release: 0.99.2 (bug/b1-e34-defects)

The scheduler row read 'quota reaped 0' — the raw key, humanised — because the new result key had no
label (found in the browser check; nothing tests result-key labels).

.dockerignore

@@ -4,6 +4,7 @@
 *.env.local
 **/__pycache__/
 **/*.pyc
+**/.coverage
 .venv/
 venv/
 # Docker patterns are anchored at the context root with NO implicit `**/` (unlike .gitignore, where

.env.example

@@ -28,6 +28,14 @@ ALLOWED_EMAILS=
 # Admin emails (subset of the above) get the admin role.
 ADMIN_EMAILS=
 
+# Logging (E22). LOG_LEVEL is the BOOTSTRAP level — the level a process starts at and the only one
+# the worker/download child read; the API's running level is the admin Configuration page's
+# `log_level` knob, which takes effect without a restart. TRACE | DEBUG | INFO | WARNING | ERROR, in
+# any letter case; any other value stops the boot with an error. LOG_FORMAT is "json" (one object per
+# line — what a log collector ingests) or "text" (a readable line for a local terminal).
+LOG_LEVEL=INFO
+LOG_FORMAT=json
+
 # Optional: origin of a separately-served frontend dev server (enables CORS). Leave empty in production.
 FRONTEND_ORIGIN=
 

.gitea/workflows/check.yml

@@ -1,71 +0,0 @@
-# Server-side mirror of the local `siftlode check` gate (E1 S3 / DX-1), so `main` is always green
-# independent of any one machine. Two parallel jobs match the two local lanes:
-#   frontend — tsc (3 projects) + eslint + prettier --check + knip + vitest
-#   backend  — ruff + the full pytest suite (pure-logic + the DB-backed lane) against a real Postgres
-#
-# Runner: the lab act_runner (LXC 115) mounts the Docker socket, so service containers work. The backend
-# runs pytest NATIVELY against a `postgres` service (not the local docker-compose.test.yml, whose
-# bind-mount is unreliable under a mounted-socket runner); fidelity is kept by installing the SAME pinned
-# requirements, Postgres 16, and ffmpeg, and by setting SIFTLODE_TEST_REQUIRE_DB=1 so a DB/ffmpeg that
-# failed to come up is a hard error, never a silent skip. node 22 matches the prod build image
-# (Dockerfile `node:22-alpine`).
-name: gate
-
-on: [push, pull_request]
-
-jobs:
-  frontend:
-    runs-on: ubuntu-latest
-    container: node:22-bookworm
-    defaults:
-      run:
-        working-directory: frontend
-    steps:
-      - uses: actions/checkout@v4
-      - run: npm ci
-      - run: npm run typecheck
-      - run: npm run lint
-      - run: npm run format:check
-      - run: npm run knip
-      - run: npm run test
-
-  backend:
-    runs-on: ubuntu-latest
-    # A combined python+node image: python 3.12 (matches the prod Dockerfile) for the tests, and node
-    # so the JS actions (checkout) run. actions/setup-python is NOT usable here — the runner can't reach
-    # GitHub's python-versions download, so it must be baked into the image.
-    container: nikolaik/python-nodejs:python3.12-nodejs20
-    services:
-      postgres:
-        image: postgres:16-alpine
-        env:
-          POSTGRES_USER: siftlode
-          POSTGRES_PASSWORD: siftlode
-          POSTGRES_DB: siftlode_test
-        options: >-
-          --health-cmd "pg_isready -U siftlode -d siftlode_test"
-          --health-interval 2s
-          --health-timeout 3s
-          --health-retries 30
-    env:
-      DATABASE_URL: postgresql+psycopg://siftlode:siftlode@postgres:5432/siftlode_test
-      SIFTLODE_TEST_REQUIRE_DB: "1"
-    steps:
-      - uses: actions/checkout@v4
-      # ffmpeg for the poster/thumbnail tests; postgresql-client for the readiness wait below.
-      - run: apt-get update && apt-get install -y --no-install-recommends ffmpeg postgresql-client
-      - name: Wait for Postgres
-        run: until pg_isready -h postgres -U siftlode -d siftlode_test; do sleep 1; done
-      - run: pip install --no-cache-dir -r backend/requirements.lock -r backend/requirements-dev.txt
-      - name: ruff
-        run: python -m ruff check --config backend/ruff.toml backend scripts
-      - name: ruff format --check
-        run: python -m ruff format --check --config backend/ruff.toml backend scripts
-      - name: pytest
-        run: python -m pytest
-        working-directory: backend
-      # Dependency CVE audit over the pinned lock. Non-blocking for now (informational) — a real CVE
-      # shows up in the run without failing the gate; tighten to a hard fail once the noise is known.
-      - name: pip-audit (advisory)
-        run: pip install --no-cache-dir pip-audit && pip-audit -r backend/requirements.lock
-        continue-on-error: true

.gitignore

@@ -9,6 +9,7 @@ __pycache__/
 venv/
 .mypy_cache/
 .pytest_cache/
+.coverage
 .ruff_cache/
 
 # Node / frontend
@@ -37,3 +38,7 @@ backups/
 Thumbs.db
 .idea/
 .vscode/
+
+# TypeScript incremental build artifacts — machine-local, absolute-path-bearing, and no gate lane
+# reads or regenerates them, so a committed one is stale from the commit that adds it.
+*.tsbuildinfo

.nvmrc

@@ -0,0 +1 @@
+24

.python-version

@@ -0,0 +1 @@
+3.12

.trivyignore.yaml

@@ -0,0 +1,15 @@
+# The CVE gate's allowlist (E32). The publish gate scans the built image and the committed tree with
+# Trivy and fails on any HIGH or CRITICAL finding that has a fix available. A finding that cannot be
+# fixed yet is let through ONLY by an entry here, and every entry carries both:
+#   statement   — why it is acceptable for now (what is exposed, what is not)
+#   expired_at  — the date it stops being accepted; after it, the gate fails again on its own
+# A contract test (frontend/src/lib/gates.contract.test.ts) refuses an entry without either, and
+# docs/gates.md lists them.
+#
+# Example:
+#   vulnerabilities:
+#     - id: CVE-2026-00000
+#       statement: in libfoo's XML path, which nothing in the image calls; no fixed Debian build yet
+#       expired_at: 2026-12-31
+vulnerabilities: []
+secrets: []

CLAUDE.md

@@ -7,17 +7,21 @@ worse than no rule — if you change how the code works, change this file in the
 
 Siftlode is a self-hosted YouTube reader (+ a local Plex browser/player and a download center). It is a
 **modular monolith**: one Postgres, one FastAPI process (uvicorn + APScheduler), one worker process
-(yt-dlp/ffmpeg), one PO-token sidecar. React 18 + react-query frontend with a router-free navigation
+(yt-dlp/ffmpeg), one PO-token sidecar. React 19 (with the React Compiler) + react-query frontend, with a
-provider.
+router-free navigation provider.
 
 ## The gate — never merge or ship red
 
-The local gate is **ruff (`check` + `format --check`) + typecheck (3 tsc projects) + eslint + prettier
+The local gate is **ruff (`check` + `format --check`) + basedpyright (backend) + the lock check + typecheck (3 tsc projects) +
---check + knip + vitest + pytest**. It runs on every `publish`, which **refuses to build on failure — there is deliberately no
+eslint + prettier --check + knip (both modes) + vitest + vite build + pytest + pip-audit + npm audit**, then a coverage report (no threshold). It runs
-`--skip` flag.** Run it and get it green **before** any merge, release, or handing work to review.
+before every `release` merges and on every `publish`, which **refuses to build on failure — there is deliberately no `--skip`
+flag**; a release also requires the Playwright suite, and the image must pass a Trivy scan before it is pushed. Run it and get it
+green **before** any merge, release, or handing work to review.
 
-- Full gate (agents with the private ops dispatcher): `siftlode check`.
+- Full gate: `bash scripts/check.sh` (`lint` for the static lanes only); the private ops dispatcher's `siftlode check` runs the same script. Single lanes: [`docs/development.md`](docs/development.md) § The gate.
-- Directly, without the dispatcher: `cd frontend && npm run typecheck && npm run lint && npm run knip && npm run test`, plus backend `ruff check backend` + `ruff format --check backend` and `pytest` (the DB-lane needs Postgres).
+- **There is no CI** (owner, 2026-10-04): this gate, run here, is the only one — green before every hand-over, and `release` refuses to merge without it.
+- **A gate with any real hit is never switched off or narrowed** — not a lint rule, a compiler flag, a lane or a release step. Fix the sites, or suppress each one inline naming its invariant (`-- why`, lint-enforced). What stays narrowed is in [`docs/gates.md`](docs/gates.md), and a contract test fails when a narrowing is missing from it. The effective ESLint config of every file and the gate/lane lines are pinned in `frontend/src/lib/gates.snapshot.json`: a new code file or any gate change fails until you regenerate it (`GATE_SNAPSHOT=update`, see gates.md) — then read that diff as a gate change. A comment saying "X is enforced by Y" is checked by running Y (E32: one such comment hid 95 functions the React Compiler skipped).
+- **The backend type lane runs in the test image, not on the host** — `requirements.lock` cannot be installed on Windows (uvloop), and a type checker without the real packages reports nonsense. It fails on NEW errors only; the pre-existing ones sit in `backend/.basedpyright/baseline.json` and burn down epic by epic. Fixed some? Regenerate with `--writebaseline`.
 - New DB-backed test? It must actually run, not silently skip — assume `SIFTLODE_TEST_REQUIRE_DB=1`.
 
 ## Workflow
@@ -45,20 +49,22 @@ how it gets built. Concretely:
   behind each option and the weight of the decision. These decision points are where the product owner
   engages with the build and learns from it, so make them understandable — a reasoned menu, not a bare one.
 - **Mockups for design/UX decisions.** For any visual or interaction choice, show mockup options to
-  choose from before building.
+  choose from before building. How the UI looks and behaves is decided in
+  [`docs/DESIGN.md`](docs/DESIGN.md) (its variants render in the dev-only `/design-lab`): build to it,
+  and change it only through its decision log.
 - **Challenge unsound requests.** If a request is technically ill-advised and you can back it with real
   arguments, push back or propose an alternative — don't comply blindly.
 - **Verify UI in the browser before calling it done.** tsc/unit/DOM tests miss visual and interaction
-  bugs; do a real visual check of any UI change.
+  bugs; do a real visual check of any UI change, and run DESIGN.md's four-step Conformance check.
 - **Be terse everywhere else.** Outside those decision points, lead with the answer or decision. During
   implementation and debugging especially: no restating the same point reworded, no filler.
 
 ## Architecture shape (don't fight it)
 
 `routes/*` (HTTP only: validation, auth-dependency, serialization) → domain packages `sync/`, `plex/`,
-`downloads/`, `youtube/` (shared by routes + scheduler + worker) → flat core (`auth`, `quota`,
+`downloads/`, `youtube/`, `issues/` (shared by routes + scheduler + worker) → flat core (`auth`, `quota`,
-`sysconfig`, `state`, `audit`, `notifications`, `userscope`, `realtime`, `ratelimit`, `progress`,
+`sysconfig`, `state`, `audit`, `notifications`, `userscope`, `realtime`, `ratelimit`, `bodylimit`, `progress`, `collation`,
-`security`) → `models/db/config`. **Dependency direction is one-way: core never imports routes.**
+`security`, `logs`, `logsink`, `diagnostics`, `debugsession`) → `models/db/config`. **Dependency direction is one-way: core never imports routes.**
 Frontend: `main.tsx` provider tree → `App.tsx` shell → `moduleRegistry` route-chunks → module
 components → the **platform layer `components/ui/`** (domain-free primitives: the popover engine and
 its menus, `Modal`/`Overlay`, `Button`/`form`, `DataTable`/`VirtualGrid`, `QueryState`, …). That
@@ -74,27 +80,39 @@ CQRS, a repository/DI framework, a router library, a generic CRUD-router factory
 
 ## Canonical patterns — when you have a choice, pick THIS
 
-| Area                    | Use                                                          | Not                                                    |
+| Area                                 | Use                                                                                                                                                                                               | Not                                                                                                                                                     |
-| ----------------------- | ------------------------------------------------------------ | ------------------------------------------------------ |
+| ------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| ORM                     | SQLAlchemy 2.0 `select()`                                    | `db.query(...)`                                        |
+| ORM                                  | SQLAlchemy 2.0 `select()`                                                                                                                                                                         | `db.query(...)`                                                                                                                                         |
-| Request body            | a Pydantic model                                             | `payload: dict`                                        |
+| Request body                         | a Pydantic model                                                                                                                                                                                  | `payload: dict`                                                                                                                                         |
-| Ownership check         | `owned_or_404` / `get_or_404`                                | ad-hoc `_own_playlist`/`_owned`/`_own_job`             |
+| Request SIZE                         | nothing — `bodylimit` caps every route (SEC-20); need more? add a ceiling to `ROUTE_MAX_BODY_BYTES`                                                                                               | `capped_body` on a route that does not read raw bytes itself                                                                                            |
-| Admin gate import       | `app.auth`                                                   | the `app.routes.admin` re-export                       |
+| Ownership check                      | `owned_or_404` / `get_or_404`                                                                                                                                                                     | ad-hoc `_own_playlist`/`_owned`/`_own_job`                                                                                                              |
-| Upstream failure status | `422` with JSON detail                                       | `502` (reserve 502 for raw byte-proxies)               |
+| Admin gate import                    | `app.auth`                                                                                                                                                                                        | the `app.routes.admin` re-export                                                                                                                        |
-| User-facing error       | structured `detail` (`{code, reason}`)                       | a bare English string under a bilingual UI             |
+| Upstream failure status              | `422` with JSON detail                                                                                                                                                                            | `502` (reserve 502 for raw byte-proxies)                                                                                                                |
-| Mutation response       | `{"ok": true, …}` (`{"deleted": id}`)                        | ad-hoc `{"unsubscribed"}`/`{"cleared": n}`/echo        |
+| User-facing error                    | structured `detail` (`{code, reason}`)                                                                                                                                                            | a bare English string under a bilingual UI                                                                                                              |
-| Transaction commit      | the handler/job-wrapper commits; helpers `flush`             | helpers that commit (only `quota` owns its session)    |
+| Mutation response                    | `{"ok": true, …}` (`{"deleted": id}`), test-held                                                                                                                                                  | a bare `{"unsubscribed"}`/`{"cleared": n}`, no `ok`                                                                                                     |
-| Admin knob storage      | `sysconfig` (values)                                         | env / `app_state` / `scheduler_settings` scatter       |
+| Transaction commit                   | the handler commits; helpers `flush` (test-held list)                                                                                                                                             | helpers that commit (only `quota` owns its session)                                                                                                     |
-| Job status word         | the `downloads/status.py` StrEnum                            | loose `"canceled"`/`"cancelled"` literals              |
+| Admin knob storage                   | `sysconfig` (values)                                                                                                                                                                              | env / `app_state` / `scheduler_settings` scatter                                                                                                        |
-| Current time            | inline `datetime.now(timezone.utc)`                          | new `now_utc`/`_now` helpers                           |
+| Job status word                      | the `downloads/status.py` StrEnum                                                                                                                                                                 | loose `"canceled"`/`"cancelled"` literals                                                                                                               |
-| FE per-account state    | `usePref`                                                    | raw `readAccount`/`writeAccount` or bespoke hooks      |
+| Current time                         | inline `datetime.now(timezone.utc)`                                                                                                                                                               | new `now_utc`/`_now` helpers                                                                                                                            |
-| FE action button        | `ui/Button` — pick a `SIZES` rung                            | a hand-rolled `<button className="px-3 py-2 …">`       |
+| FE per-account state                 | `usePref`                                                                                                                                                                                         | raw `readAccount`/`writeAccount` or bespoke hooks                                                                                                       |
-| FE keyboard shortcut    | a row in `lib/keymap.ts` + `dispatch()`                      | a private `keydown` listener with its own key literals |
+| FE button                            | `ui/Button` (a `SIZES` rung) · `IconButton` · `Pressable` (a clickable card/row) · `Chip` · `DragHandle`                                                                                          | a raw `<button>` or `FocusSafeButton` outside the control primitives (lint-enforced)                                                                    |
-| FE list/grid keys       | `NavGrid`/`keyboardNav` + `navItemAttrs` + `data-nav-open`   | a private arrow-key / roving-tabindex handler          |
+| FE keyboard shortcut                 | a row in `lib/keymap.ts` + `dispatch()`                                                                                                                                                           | a private `keydown` listener with its own key literals                                                                                                  |
-| FE view push (history)  | `rememberNavPosition()` right before `history.pushState`     | a push that forgets: Back strands focus on `<body>`    |
+| FE list/grid keys                    | `NavGrid`/`keyboardNav` + `navItemAttrs` + `data-nav-open`                                                                                                                                        | a private arrow-key / roving-tabindex handler                                                                                                           |
-| FE control              | the chip/segmented primitives (`Select` only for >6 options) | a native `<select>`                                    |
+| FE view push (history)               | `rememberNavPosition()` right before `history.pushState`                                                                                                                                          | a push that forgets: Back strands focus on `<body>`                                                                                                     |
-| FE loading/empty/error  | `QueryState` / `StateMessage`                                | ad-hoc conditionals                                    |
+| FE text field / checkbox / slider    | `ui/Input` · `Textarea` · `InnerInput` in a `fieldClass` frame · `Checkbox`/`CheckRow` · `Slider`                                                                                                 | a raw `<input>`/`<textarea>` anywhere but those primitives (lint-enforced)                                                                              |
-| FE failed write         | `meta.failTitle`, else `notifyActionError(err, title)`       | a bare `catch {}` or `notify()` with no error attached |
+| FE glass surface                     | a role: `SURFACE.panel` / `card` / `menu` / `popover` / `tooltip` (`ui/surfaces.ts`)                                                                                                              | a bare `glass`/`glass-card`/`glass-menu` beside a hand-picked radius (lint-enforced)                                                                    |
+| FE static label / button-like link   | `Badge` (flat, no edge) · `LinkButton` / `IconLink`                                                                                                                                               | a glass pill that reads as a chip; an `<a>` hand-painted like a button                                                                                  |
+| FE control                           | `SegmentedToggle`; a long or user's list: `FilterSelect`                                                                                                                                          | a native `<select>` (lint-enforced)                                                                                                                     |
+| FE loading/empty/error               | `QueryState` / `StateMessage`                                                                                                                                                                     | ad-hoc conditionals                                                                                                                                     |
+| FE failed write                      | `meta.failTitle`, else `notifyActionError(err, title)`                                                                                                                                            | a bare `catch {}` or `notify()` with no error attached                                                                                                  |
+| FE network call                      | `req`/`uploadReq`/`beacon`, or `lib/http.ts` `send` (public pages)                                                                                                                                | a bare `fetch`/`XMLHttpRequest`/`WebSocket` (lint-enforced)                                                                                             |
+| FE diagnostics                       | `lib/log.ts` (`log.info("event.code", msg, attrs)`); a filter change → `lib/filterLog.ts`, a player → `lib/mediaLog.ts`                                                                           | `console.*` (lint-enforced); a value only a higher level may carry, logged unguarded                                                                    |
+| FE state that MIRRORS a prop/query   | derive it, or hold a draft (`draft ?? server`)                                                                                                                                                    | a copy an effect re-syncs — that paints the stale value first                                                                                           |
+| FE state reset on a key change       | adjust it DURING render against a remembered key                                                                                                                                                  | `useEffect(() => setX(init), [key])`                                                                                                                    |
+| FE effect that must not re-subscribe | `useEffectEvent` (React 19.2+); a handler or timer: `useLatest` / `useStableCallback`                                                                                                             | a render-phase `ref.current = x` — the React Compiler skips the component                                                                               |
+| FE hook suppression                  | an inline `eslint-disable-next-line` NAMING the invariant                                                                                                                                         | `eslint-suppressions.json` (deliberately absent — E27 S2)                                                                                               |
+| FE popover / scroll-fade ref         | destructure it: `const { setReference } = pop.refs`, then `ref={setReference}`                                                                                                                    | `ref={pop.refs.setReference}` / `ref={fade.ref}` — the compiler then treats the whole object as a ref and skips                                         |
+| FE async handler with cleanup        | `try { … } catch { … }`, then the cleanup after it (steps that `return` early go in their own function); a conditional, loop or `?.` the `try` needs goes before it or into a module-level helper | `try/finally`, or a `?:` / `??` / loop / `?.` inside `try` — the React Compiler skips the whole component (and `react-hooks/todo` misses most of these) |
 
 ## House rules that bite (learned from real bugs)
 

Dockerfile

@@ -10,7 +10,9 @@ ARG GIT_SHA=unknown
 ARG BUILD_DATE=
 
 # Stage 1: build the frontend SPA
-FROM node:22-alpine AS frontend
+# The major is pinned to `.nvmrc` and held there by frontend/src/lib/nodeEngine.contract.test.ts:
+# this image and the dev machine the gate runs on must be ONE engine (see that test for why).
+FROM node:24-alpine AS frontend
 WORKDIR /fe
 # Deps first, keyed on the lockfile, so editing app source never re-resolves node_modules.
 COPY frontend/package.json frontend/package-lock.json ./
@@ -51,8 +53,11 @@ RUN adduser --disabled-password --gecos "" appuser \
 # Harmless on hosts with no /dev/dri (the driver just goes unused). Deno: the JavaScript runtime yt-dlp
 # uses to solve YouTube's "n" signature challenge for the web player clients (required alongside the PO
 # token, else only storyboards are offered). yt-dlp auto-detects deno on PATH.
+# `apt-get upgrade` (E32): the base image is a snapshot, and the Debian security fixes released since it
+# was cut are only in the archive — the first Trivy scan found one (pcre2) fixed upstream but not in it.
 ARG DENO_VERSION=v2.9.1
 RUN apt-get update \
+    && apt-get upgrade -y --no-install-recommends \
     && apt-get install -y --no-install-recommends ffmpeg intel-media-va-driver unzip curl ca-certificates \
     && curl -fsSL "https://github.com/denoland/deno/releases/download/${DENO_VERSION}/deno-x86_64-unknown-linux-gnu.zip" -o /tmp/deno.zip \
     && unzip -o /tmp/deno.zip -d /usr/local/bin \
@@ -74,11 +79,14 @@ RUN printf 'Types: deb\nURIs: http://deb.debian.org/debian\nSuites: bookworm\nCo
     && rm -rf /var/lib/apt/lists/*
 
 # Install from the PINNED lock (requirements.lock, generated from requirements.txt) so the image is
-# byte-reproducible and the exact versions are the ones pip-audit scanned in CI — not whatever floats
+# byte-reproducible and the exact versions are the ones the gate's pip-audit scanned — not whatever
-# up from the floor ranges on build day. Regenerate the lock when requirements.txt changes (see its header).
+# floats up from the floor ranges on build day. Regenerate the lock when requirements.txt changes (see its header).
+# pip itself is removed once it has installed the lock (E32): nothing runs it at runtime, and it carries
+# its own VENDORED copies of urllib3, msgpack and setuptools, which Trivy reports as the image's own —
+# an attack surface and a CVE stream for a tool the running app never calls.
 COPY backend/requirements.lock .
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip install --upgrade pip && pip install -r requirements.lock
+    pip install --upgrade pip && pip install -r requirements.lock && pip uninstall -y pip
 
 COPY --chown=appuser backend/ .
 # Read at runtime by config.py as the app_version (a committed single source of truth).

VERSION

@@ -1 +1 @@
-0.87.0
+0.99.2

backend/alembic.ini

@@ -1,6 +1,8 @@
 [alembic]
 script_location = alembic
 prepend_sys_path = .
+# Split the path list the way the OS does (alembic's legacy space/comma/colon split is deprecated).
+path_separator = os
 
 [loggers]
 keys = root,sqlalchemy,alembic

backend/alembic/env.py

@@ -1,17 +1,17 @@
-from logging.config import fileConfig
-
 from alembic import context
 from sqlalchemy import engine_from_config, pool
 
 from app.config import settings
 from app.db import Base
+from app.logs import configure_logging
 import app.models  # noqa: F401  (import so models register on Base.metadata)
 
 config = context.config
 config.set_main_option("sqlalchemy.url", settings.database_url)
 
-if config.config_file_name is not None:
+# Migrations run at container start (entrypoint.sh), so their output is part of the API's log
-    fileConfig(config.config_file_name)
+# stream — one format with everything else, not alembic.ini's own fileConfig shape.
+configure_logging("alembic")
 
 target_metadata = Base.metadata
 

backend/alembic/versions/0085_users_last_login_at.py

@@ -0,0 +1,26 @@
+"""add users.last_login_at — when the account last signed in (E22 S1)
+
+Revision ID: 0085_users_last_login_at
+Revises: 0084_dl_job_error_detail
+Create Date: 2026-09-16
+
+Stamped by the one sign-in path (`routes/auth._establish_session`), which the Google callback and
+password login share. No backfill is possible: the signal exists only at sign-in, so existing rows
+read NULL until each account signs in again, and the admin UI says so rather than inventing a date.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0085_users_last_login_at"
+down_revision: str | None = "0084_dl_job_error_detail"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.add_column("users", sa.Column("last_login_at", sa.DateTime(timezone=True), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("users", "last_login_at")

backend/alembic/versions/0086_log_events.py

@@ -0,0 +1,48 @@
+"""add log_events — stored log events for diagnostics (E22 S3)
+
+Revision ID: 0086_log_events
+Revises: 0085_users_last_login_at
+Create Date: 2026-09-17
+
+Written by app/logsink.py's background writer (WARNING and above, request lines), pruned by the log_gc
+job after `log_retention_days`. Sized against prod on 2026-09-16: ~2k rows on a normal day, ~14k on a
+YouTube-RSS-outage day before flood control — a few MB over the 7-day window. `user_id` carries no
+foreign key on purpose (see models.LogEvent).
+"""
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects import postgresql
+
+revision: str = "0086_log_events"
+down_revision: str | None = "0085_users_last_login_at"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "log_events",
+        sa.Column("id", sa.BigInteger(), primary_key=True),
+        sa.Column("ts", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("level", sa.String(10), nullable=False),
+        sa.Column("sev", sa.SmallInteger(), nullable=False),
+        sa.Column("source", sa.String(16), nullable=False),
+        sa.Column("logger", sa.String(128), nullable=False),
+        sa.Column("event", sa.String(64), nullable=True),
+        sa.Column("msg", sa.Text(), nullable=False),
+        sa.Column("attrs", postgresql.JSONB(), nullable=True),
+        sa.Column("exc", sa.Text(), nullable=True),
+        sa.Column("request_id", sa.String(64), nullable=True),
+        sa.Column("user_id", sa.Integer(), nullable=True),
+        sa.Column("debug_session_id", sa.String(64), nullable=True),
+        sa.Column("job_id", sa.Integer(), nullable=True),
+        sa.Column("version", sa.String(64), nullable=True),
+    )
+    op.create_index("ix_log_events_ts", "log_events", ["ts"])
+    op.create_index("ix_log_events_request_id", "log_events", ["request_id"])
+    op.create_index("ix_log_events_user_ts", "log_events", ["user_id", "ts"])
+
+
+def downgrade() -> None:
+    op.drop_table("log_events")

backend/alembic/versions/0087_users_diagnostics_level.py

@@ -0,0 +1,31 @@
+"""add users.diagnostics_level + diagnostics_until — a user's consented diagnostics level (E22 S5)
+
+Revision ID: 0087_users_diagnostics_level
+Revises: 0086_log_events
+Create Date: 2026-09-17
+
+Both NULL = the base level (Alap). A raised level is ALWAYS paired with its expiry, computed by the
+server at consent time (`app/diagnostics.py`), so no client can ask for a level that never ends.
+Columns rather than `users.preferences`: that JSON is written wholesale by the client, and a consent
+with a lifetime must not be something a client can write.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0087_users_diagnostics_level"
+down_revision: str | None = "0086_log_events"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.add_column("users", sa.Column("diagnostics_level", sa.String(8), nullable=True))
+    op.add_column(
+        "users", sa.Column("diagnostics_until", sa.DateTime(timezone=True), nullable=True)
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("users", "diagnostics_until")
+    op.drop_column("users", "diagnostics_level")

backend/alembic/versions/0088_channels_feed_fallback_at.py

@@ -0,0 +1,30 @@
+"""add channels.feed_fallback_at — when a channel's uploads were last read through the API because
+its RSS feed would not answer
+
+Revision ID: 0088_channels_feed_fallback_at
+Revises: 0087_users_diagnostics_level
+Create Date: 2026-09-19
+
+YouTube's RSS feeds go dark for hours at a time, per channel (2026-09-18/19: 393 of 396 answered 404
+for four hours; the same channel answers 404 to every retry). The RSS poll then reads the newest
+uploads through the Data API instead, at a quota cost, and at most once per `rss_fallback_hours` per
+channel: this column is that clock. NULL = never needed it.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0088_channels_feed_fallback_at"
+down_revision: str | None = "0087_users_diagnostics_level"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.add_column(
+        "channels", sa.Column("feed_fallback_at", sa.DateTime(timezone=True), nullable=True)
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("channels", "feed_fallback_at")

backend/alembic/versions/0089_debug_sessions.py

@@ -0,0 +1,81 @@
+"""add debug_sessions + debug_screenshots, and index log_events.debug_session_id (E16)
+
+Revision ID: 0089_debug_sessions
+Revises: 0088_channels_feed_fallback_at
+Create Date: 2026-09-19
+
+A debug session is one recording: Start → reproduce → Stop (app/debugsession.py). Every log line
+of that user's requests while it runs carries its id in `log_events.debug_session_id` — the column
+E22 created and left NULL — and the issue it ends in is built from exactly those rows, so the column
+gets its index now. Partial: nearly every row has no session, and an index of NULLs would be most
+of the table for nothing.
+
+At most one OPEN session per user, held by a partial unique index rather than by the start handler
+alone: two tabs pressing Start at once would otherwise both pass a "none open" check.
+
+Screenshots are rows, not files: bounded (size and count per session, app/debugsession.py), gone
+with their session by cascade, and a self-host has no file volume to configure for them.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0089_debug_sessions"
+down_revision: str | None = "0088_channels_feed_fallback_at"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.create_table(
+        "debug_sessions",
+        sa.Column("id", sa.String(16), primary_key=True),
+        sa.Column(
+            "user_id",
+            sa.Integer,
+            sa.ForeignKey("users.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column("level", sa.String(8), nullable=False),
+        sa.Column("started_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("until", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("ended_at", sa.DateTime(timezone=True), nullable=True),
+        sa.Column("ended_reason", sa.String(16), nullable=True),
+        sa.Column("issue_url", sa.String(512), nullable=True),
+    )
+    op.create_index(
+        "uq_debug_sessions_open_per_user",
+        "debug_sessions",
+        ["user_id"],
+        unique=True,
+        postgresql_where=sa.text("ended_at IS NULL"),
+    )
+    op.create_table(
+        "debug_screenshots",
+        sa.Column("id", sa.Integer, primary_key=True),
+        sa.Column(
+            "session_id",
+            sa.String(16),
+            sa.ForeignKey("debug_sessions.id", ondelete="CASCADE"),
+            nullable=False,
+            index=True,
+        ),
+        sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+        sa.Column("mime", sa.String(16), nullable=False),
+        sa.Column("size", sa.Integer, nullable=False),
+        sa.Column("data", sa.LargeBinary, nullable=False),
+    )
+    op.create_index(
+        "ix_log_events_debug_session",
+        "log_events",
+        ["debug_session_id"],
+        postgresql_where=sa.text("debug_session_id IS NOT NULL"),
+    )
+
+
+def downgrade() -> None:
+    op.drop_index("ix_log_events_debug_session", table_name="log_events")
+    op.drop_table("debug_screenshots")
+    op.drop_index("uq_debug_sessions_open_per_user", table_name="debug_sessions")
+    op.drop_table("debug_sessions")

backend/alembic/versions/0090_debug_session_draft.py

@@ -0,0 +1,27 @@
+"""add debug_sessions.draft — the issue being written from a session (E16 S3)
+
+Revision ID: 0090_debug_session_draft
+Revises: 0089_debug_sessions
+Create Date: 2026-09-19
+
+The composer saves as it goes: the title, the rendered markdown body that an export or an issue
+carries, the log rows the user took out, and the form's own fields so reopening it later (Settings →
+Diagnosztika lists past sessions) shows what was written. NULL until the composer first saves.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+from sqlalchemy.dialects.postgresql import JSONB
+
+revision: str = "0090_debug_session_draft"
+down_revision: str | None = "0089_debug_sessions"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.add_column("debug_sessions", sa.Column("draft", JSONB, nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("debug_sessions", "draft")

backend/alembic/versions/0091_plex_tagline_hu.py

@@ -0,0 +1,27 @@
+"""add tagline_hu to plex_items + plex_shows — the HU overlay's tagline (E30 S1)
+
+Revision ID: 0091_plex_tagline_hu
+Revises: 0090_debug_session_draft
+Create Date: 2026-09-29
+
+The Hungarian overlay section lists a tagline beside the title and summary; the info page showed the
+English one to a Hungarian reader. NULL until the next overlay sync fills it (serving falls back to EN).
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0091_plex_tagline_hu"
+down_revision: str | None = "0090_debug_session_draft"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.add_column("plex_items", sa.Column("tagline_hu", sa.Text(), nullable=True))
+    op.add_column("plex_shows", sa.Column("tagline_hu", sa.Text(), nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("plex_shows", "tagline_hu")
+    op.drop_column("plex_items", "tagline_hu")

backend/alembic/versions/0092_plex_person_biography_hu.py

@@ -0,0 +1,31 @@
+"""add biography_hu to plex_people — TMDB's Hungarian biography (E30 S4)
+
+Revision ID: 0092_plex_person_biography_hu
+Revises: 0091_plex_tagline_hu
+Create Date: 2026-09-29
+
+The person page showed a Hungarian reader the English biography although TMDB has a Hungarian one for
+most people. The person fetch now appends `translations` and keeps the Hungarian biography beside the
+English one. Already-enriched people are un-stamped, so their next page open refetches them once (the
+same lazy path as a first open) and fills it; until then the English biography stands in.
+"""
+
+import sqlalchemy as sa
+from alembic import op
+
+revision: str = "0092_plex_person_biography_hu"
+down_revision: str | None = "0091_plex_tagline_hu"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.add_column("plex_people", sa.Column("biography_hu", sa.Text(), nullable=True))
+    op.execute(
+        "UPDATE plex_people SET tmdb_enriched_at = NULL "
+        "WHERE tmdb_person_id IS NOT NULL AND tmdb_enriched_at IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    op.drop_column("plex_people", "biography_hu")

backend/alembic/versions/0093_plex_person_known_for_votes.py

@@ -0,0 +1,29 @@
+"""re-fetch linked people — "Known for" ranked by votes, without guest visits (E30 S6)
+
+Revision ID: 0093_plex_person_known_for_votes
+Revises: 0092_plex_person_biography_hu
+Create Date: 2026-09-30
+
+A person's "Known for" pool was their credits ranked by TMDB popularity, so talk-show visits as
+themselves crowded out the films (Tom Cruise: 14 of 15 were talk shows, none in the library). The pool is
+now ranked by vote count, without guest visits, with directed works, and deeper. Already-enriched people
+are un-stamped so the people_enrich job fetches them again; their biography stays on the row meanwhile.
+"""
+
+from alembic import op
+
+revision: str = "0093_plex_person_known_for_votes"
+down_revision: str | None = "0092_plex_person_biography_hu"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    op.execute(
+        "UPDATE plex_people SET tmdb_enriched_at = NULL "
+        "WHERE tmdb_person_id IS NOT NULL AND tmdb_enriched_at IS NOT NULL"
+    )
+
+
+def downgrade() -> None:
+    pass  # the watermark only says "fetched"; a refetch under the old ranking is not worth undoing to

backend/alembic/versions/0094_plex_genres_hu_everywhere.py

@@ -0,0 +1,45 @@
+"""Hungarian genres on every Plex row, not only on overlaid ones (E30 review)
+
+Revision ID: 0094_plex_genres_hu_everywhere
+Revises: 0093_plex_person_known_for_votes
+Create Date: 2026-09-30
+
+The HU genre facets and filter read `genres_hu`, which only the overlay pass wrote — so a library with
+no HU twin, or a guid-less row, offered English genres a Hungarian filter could never match. The sync
+now writes `genres_hu` for every row (plex/sync.py `_apply_facet_fields`); this fills the rows it has
+not reached yet, from their own English list through the same map.
+"""
+
+import json
+
+import sqlalchemy as sa
+from alembic import op
+
+from app.plex.genres import hu_genres
+
+revision: str = "0094_plex_genres_hu_everywhere"
+down_revision: str | None = "0093_plex_person_known_for_votes"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    for table in ("plex_items", "plex_shows"):
+        rows = conn.execute(
+            sa.text(
+                f"SELECT id, genres FROM {table} "
+                "WHERE genres_hu IS NULL AND jsonb_typeof(genres) = 'array'"
+            )
+        ).fetchall()
+        updates = [
+            {"id": rid, "g": json.dumps(hu)} for rid, genres in rows if (hu := hu_genres(genres))
+        ]
+        if updates:
+            conn.execute(
+                sa.text(f"UPDATE {table} SET genres_hu = CAST(:g AS jsonb) WHERE id = :id"), updates
+            )
+
+
+def downgrade() -> None:
+    pass  # derived from `genres`; the sync writes the same values again

backend/alembic/versions/0095_videos_channel_cover_index.py

@@ -0,0 +1,83 @@
+"""videos: the per-channel index covers what the channel manager and the feed sidebar read (E31 S4d)
+
+The channel manager (`GET /api/channels`) sums each subscribed channel's videos: count, last upload, total
+duration. The feed sidebar (`/feed/summary`, `/feed/years`) counts the reader's videos per channel and per
+year under the default filters: available, not via explore, not a short, not live. Every one of them found
+its rows through `ix_videos_channel_pub_desc` (channel_id, published_at DESC NULLS LAST, id) and then read
+each video's heap page for the rest: ~270k rows at ~2.5 KB each, 681 MB on prod. That stays in cache only
+while nothing else needs the memory. Measured on prod (2026-10-02): `/api/channels` 51–61 ms warm, but
+2.25 s and 1.96 s on a first open in the browser and 6.4 s under EXPLAIN with 30k blocks read from disk;
+`/feed/summary` and `/feed/years` ~140 ms warm, ~800 ms cold.
+
+The same key with those columns INCLUDEd makes them Index Only Scans. On a prod-copy localdev, after a
+VACUUM, the manager's sums read 0 heap pages and the summary reads the 25 MB index instead of the heap.
+The visibility map is what lets them skip the heap; prod's autovacuum kept it complete (relallvisible ==
+relpages). The key columns are unchanged, so every reader of the old index (the per-channel feed, the FK
+cascade, the sync upsert, channel detail) rides this one. The old index is dropped.
+tests/test_channel_list_index_only.py holds the queries and this column list together.
+
+CONCURRENTLY (autocommit_block) builds it without blocking a live prod. IF (NOT) EXISTS keeps the drift
+test's downgrade → re-upgrade round trip idempotent. A re-run after a build that failed part-way rebuilds
+the INVALID leftover, and the old index is dropped only once the new one is valid: a re-run that kept the
+leftover and dropped the old one left `videos` with no usable channel_id index (E31 S5 review).
+"""
+
+from alembic import op
+from sqlalchemy import text
+
+revision: str = "0095_videos_channel_cover_index"
+down_revision: str | None = "0094_plex_genres_hu_everywhere"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+
+def _valid(name: str) -> bool | None:
+    """The index's `indisvalid`, or None when there is no index of that name."""
+    row = (
+        op.get_bind()
+        .execute(
+            text(
+                "SELECT i.indisvalid FROM pg_index i JOIN pg_class c ON c.oid = i.indexrelid "
+                "WHERE c.relname = :name"
+            ),
+            {"name": name},
+        )
+        .first()
+    )
+    return None if row is None else bool(row[0])
+
+
+def _build(name: str, ddl: str) -> None:
+    """CREATE INDEX CONCURRENTLY `name` (`ddl` is what follows the name), re-runnable. A build that
+    failed part-way leaves an INVALID index under the name, and `IF NOT EXISTS` would keep it: the
+    planner never uses it, and nothing says so (E31 S5 review). Such a leftover is dropped and built
+    again."""
+    if _valid(name) is False:
+        op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")
+    op.execute(f"CREATE INDEX CONCURRENTLY IF NOT EXISTS {name} {ddl}")
+
+
+def _swap(build: str, ddl: str, drop: str) -> None:
+    _build(build, ddl)
+    if not _valid(build):
+        raise RuntimeError(f"{build} did not build valid; {drop} is kept")
+    op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {drop}")
+
+
+def upgrade() -> None:
+    with op.get_context().autocommit_block():
+        _swap(
+            "ix_videos_channel_pub_cover",
+            "ON videos (channel_id, published_at DESC NULLS LAST, id) "
+            "INCLUDE (duration_seconds, unavailable_since, via_explore, is_short, live_status)",
+            "ix_videos_channel_pub_desc",
+        )
+
+
+def downgrade() -> None:
+    with op.get_context().autocommit_block():
+        _swap(
+            "ix_videos_channel_pub_desc",
+            "ON videos (channel_id, published_at DESC NULLS LAST, id ASC)",
+            "ix_videos_channel_pub_cover",
+        )

backend/alembic/versions/0096_videos_sort_indexes.py

@@ -0,0 +1,86 @@
+"""videos: an index per feed sort, so a sorted page stops after its rows (E31 S4e)
+
+The feed's views, duration and title sorts had no index. Every page (the first and each next cursor)
+read the whole videos heap, filtered it and kept a top-N: on prod ~350 ms a page for views/duration,
+550–780 ms for title (the ICU comparison on every row), 2.3 s for page 2 when the copies of one request
+queued. With an index in the exact ORDER BY (`feedquery.order_by`: the key, NULLS LAST, then id) the
+planner walks it and stops after the page. On a prod-copy localdev every plain sort fell from 180–310 ms
+to ~1 ms, deep pages (20) to 3–5 ms, and narrow filters (a tag, a channel, a year, nothing found) stayed
+at 0.3–25 ms.
+
+One index per direction: a backward scan of `DESC NULLS LAST` gives `ASC NULLS FIRST`, not the
+`ASC NULLS LAST` the asc sorts ask for. The title pair is per reader language (`app.collation.ICU`):
+an index only serves an ORDER BY in its own collation.
+
+Cost: ~130 MB (localdev: 10 + 10 + 8 + 8 + 4 × 23). Writes stay cheap: view_count and duration are
+written by the enrich, which rewrites indexed columns already (the revalidation job only touches
+last_checked_at), so no HOT update is lost. A search combined with one of these sorts can now walk the
+index too: a term matching ~2 % of the videos took 130–200 ms instead of 7 (still under the plain sort it
+replaces); rarer and broader terms stay fast. Fencing the index off under a search made broad terms
+(`m`, `the`) take 2 s, so it is left to the planner.
+
+The title indexes are the schema's first ones on an ICU collation: an ICU upgrade in the Postgres image
+reorders them, and `app.collation.heal_drift` rebuilds them at startup when it sees one.
+tests/test_feed_sort_indexes.py holds each sort and its index together.
+
+CONCURRENTLY (autocommit_block) builds them without blocking a live prod; IF (NOT) EXISTS keeps the drift
+test's downgrade → re-upgrade round trip idempotent.
+"""
+
+from alembic import op
+from sqlalchemy import text
+
+revision: str = "0096_videos_sort_indexes"
+down_revision: str | None = "0095_videos_channel_cover_index"
+branch_labels: str | None = None
+depends_on: str | None = None
+
+# name → the indexed ORDER BY (feedquery.sort_keys + order_by: key NULLS LAST, then id).
+INDEXES = {
+    "ix_videos_views_desc": "view_count DESC NULLS LAST, id",
+    "ix_videos_views_asc": "view_count ASC NULLS LAST, id",
+    "ix_videos_duration_desc": "duration_seconds DESC NULLS LAST, id",
+    "ix_videos_duration_asc": "duration_seconds ASC NULLS LAST, id",
+    "ix_videos_title_en": '(lower(title) COLLATE "en-x-icu") ASC NULLS LAST, id',
+    "ix_videos_title_en_desc": '(lower(title) COLLATE "en-x-icu") DESC NULLS LAST, id',
+    "ix_videos_title_hu": '(lower(title) COLLATE "hu-x-icu") ASC NULLS LAST, id',
+    "ix_videos_title_hu_desc": '(lower(title) COLLATE "hu-x-icu") DESC NULLS LAST, id',
+}
+
+
+def _valid(name: str) -> bool | None:
+    """The index's `indisvalid`, or None when there is no index of that name."""
+    row = (
+        op.get_bind()
+        .execute(
+            text(
+                "SELECT i.indisvalid FROM pg_index i JOIN pg_class c ON c.oid = i.indexrelid "
+                "WHERE c.relname = :name"
+            ),
+            {"name": name},
+        )
+        .first()
+    )
+    return None if row is None else bool(row[0])
+
+
+def _build(name: str, ddl: str) -> None:
+    """CREATE INDEX CONCURRENTLY `name` (`ddl` is what follows the name), re-runnable. A build that
+    failed part-way leaves an INVALID index under the name, and `IF NOT EXISTS` would keep it: the
+    planner never uses it, and nothing says so (E31 S5 review). Such a leftover is dropped and built
+    again."""
+    if _valid(name) is False:
+        op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")
+    op.execute(f"CREATE INDEX CONCURRENTLY IF NOT EXISTS {name} {ddl}")
+
+
+def upgrade() -> None:
+    with op.get_context().autocommit_block():
+        for name, columns in INDEXES.items():
+            _build(name, f"ON videos ({columns})")
+
+
+def downgrade() -> None:
+    with op.get_context().autocommit_block():
+        for name in INDEXES:
+            op.execute(f"DROP INDEX CONCURRENTLY IF EXISTS {name}")

backend/app/README.md

@@ -13,16 +13,19 @@ Three layers, by *who* runs the code — not by feature:
    response models, status-code mapping, serialization. It owns logic that runs *only* on a request.
    Business logic that a request happens to trigger does **not** live here if anything else runs it too.
 
-2. **`plex/`, `downloads/`, `sync/` — domain logic shared by more than one caller.** Anything the
+2. **`plex/`, `downloads/`, `sync/`, `issues/` — domain logic shared by more than one caller.** Anything the
    **scheduler or the worker** runs (not just a request) lives in one of these domain packages, so the
    route, the worker, and the scheduler share one definition. Examples from this codebase:
    `downloads/service.py` owns enqueue + the job state transitions (routes call it); `plex/sync.py` is
-   run by both the admin route and the scheduler; `downloads/pipeline.py` is the worker's child pipeline.
+   run by both the admin route and the scheduler; `downloads/pipeline.py` is the worker's child pipeline;
+   `issues/bundle.py` is the one definition of a bug report's files, for the zip export and the tracker.
 
 3. **Flat `app/*.py` core modules — cross-cutting primitives.** `auth`, `sysconfig`, `quota`, `audit`,
-   `notifications`, `userscope`, `realtime`, `ratelimit`, `progress`, … are single-file primitives many
+   `notifications`, `userscope`, `realtime`, `ratelimit`, `progress`, `logs` (the one logging
-   layers import. They depend on nothing above them (a core module must never import a `routes/…`
+   configuration), `logsink` (its `log_events` table), `diagnostics` (a user's consented log
-   module).
+   level), `debugsession` (a recording that tags its user's lines, E16) and `collation` (A→Z in the
+   reader's alphabet, E30), … are single-file primitives many layers import. They depend on nothing above them (a core
+   module must never import a `routes/…` module).
 
 **Extraction trigger: the SECOND consumer.** Keep new logic inline in its first caller. The moment a
 *second* caller needs it (a worker as well as a route; a scheduler as well as a route), extract it down
@@ -69,7 +72,9 @@ mutation it records, so an audit entry can never exist without its change (or vi
 ## The gate
 
 Everything below must be green before a change is done (run via `siftlode check`):
-`ruff` (backend) · `tsc` ×3 + `eslint` + `prettier` + `knip` (frontend) · `vitest` · `pytest`. The
+`ruff` + `basedpyright` (backend) · the lock check + `tsc` ×3 + `eslint` + `prettier` + `knip` +
-pytest lane runs in `backend/docker-compose.test.yml` — an ephemeral Postgres + the backend tree
+`vite build` (frontend) · `vitest` · `pytest`. The pytest lane runs in `backend/docker-compose.test.yml`
-bind-mounted — so the DB-backed integration tests (the `db` fixture) actually run; `SIFTLODE_TEST_
+— an ephemeral Postgres + the backend tree bind-mounted — so the DB-backed integration tests (the `db`
-REQUIRE_DB=1` there turns a would-be silent skip into a hard failure. See `docs/development.md`.
+fixture) actually run; a missing database is a hard failure, not a silent skip, unless a hand run asks
+for pure-logic only with `SIFTLODE_TEST_REQUIRE_DB=0`. See `docs/development.md`, and `docs/gates.md`
+for what each gate deliberately leaves out.

backend/app/audit.py

@@ -58,6 +58,9 @@ class AuditAction:
     DOWNLOAD_CREDENTIAL_DELETE = "download_credential_delete"
     # The log itself was cleared (a single row survives the clear as tamper-evidence)
     AUDIT_CLEAR = "audit_clear"
+    # E16: a user sent a debug session's issue to the tracker — the one act that takes their log
+    # off this instance, so it is recorded (target = the session, summary = the issue's URL).
+    ISSUE_CREATE = "issue_create"
 
 
 def record(

backend/app/auth.py

@@ -1,19 +1,20 @@
-import logging
-
 import httpx
 from fastapi import BackgroundTasks, Depends, HTTPException, Request
 from sqlalchemy import delete, func, select
 from sqlalchemy.orm import Session
 from starlette.requests import HTTPConnection
 
+from app import debugsession
 from app import email as email_mod
+from app import logs
 from app import mailtext
 from app import sysconfig
 from app.config import settings
 from app.db import get_db
-from app.models import Invite, User
+from app.models import Invite, LogEvent, User
 from app.security import decrypt_optional
 from app.utils import canonical_email
+from app.refusal import refused
 
 PASSWORD_MIN_LEN = 10
 
@@ -37,7 +38,7 @@ READ_SCOPES = f"{BASE_SCOPES} {READ_SCOPE}"
 WRITE_SCOPES = f"{BASE_SCOPES} {READ_SCOPE} {WRITE_SCOPE}"
 _YOUTUBE_SCOPES = frozenset({READ_SCOPE, WRITE_SCOPE})
 
-log = logging.getLogger("siftlode.auth")
+log = logs.get_logger(__name__)
 
 
 def google_enabled(db: Session) -> bool:
@@ -220,7 +221,7 @@ def revoke_google_token(token: str | None) -> None:
             timeout=10,
         )
         log.info("Google token revoke: status=%s", resp.status_code)
-    except Exception:
+    except Exception:  # noqa: BLE001 — a revoke we cannot make must not fail the sign-out
         log.warning("Google token revoke failed", exc_info=True)
 
 
@@ -245,6 +246,12 @@ def purge_user(db: Session, user: User, background: BackgroundTasks) -> None:
     )
     # Raw delete so Postgres applies ON DELETE CASCADE / SET NULL on every dependent table.
     db.execute(delete(User).where(User.id == user_id))
+    # Stored log events name the account by id with no foreign key to cascade from (models.LogEvent),
+    # so they go explicitly — they are request history, i.e. personal data. Forgotten FIRST: this very
+    # request's access line, a queued row, a later background task would otherwise put rows back
+    # under the id after this DELETE (logs.forget_user).
+    logs.forget_user(user_id)
+    db.execute(delete(LogEvent).where(LogEvent.user_id == user_id))
     # Every canonically-equal invite goes, not just the literal spelling: is_allowed matches on the
     # canonical form, so leaving a sibling row behind would keep granting access to a mailbox whose
     # account was just erased — and prod really does carry approved pairs for one mailbox.
@@ -297,6 +304,13 @@ def resolved_user_id(conn: HTTPConnection, *, strict: bool = True) -> tuple[int
         try:
             hid = int(hdr)
         except ValueError:
+            # Under `strict` this becomes a 401 and the tab authenticates as nobody, which reads
+            # in the browser exactly like a session that expired. The value itself stays out —
+            # it came from a request, and `wallet` says which accounts were actually on offer.
+            log.debug(
+                "the active-account header could not be read as an id",
+                extra=logs.event("auth.account_header.invalid", accounts=len(wallet)),
+            )
             hid = None
         if hid is not None and hid in wallet:
             return hid, hid == default_id
@@ -328,19 +342,19 @@ def current_user(request: Request, db: Session = Depends(get_db)) -> User:
 
 def _session_user(request: Request, db: Session, user_id: int | None, is_default: bool) -> User:
     if not user_id:
-        raise HTTPException(status_code=401, detail="Not authenticated")
+        raise refused(401, "not_authenticated")
     user = db.get(User, user_id)
     if user is None or not user.is_active or user.is_suspended:
         # Suspended/deactivated mid-session → end it so the block takes effect at once.
         _end_account_session(request, user_id, is_default)
-        raise HTTPException(status_code=401, detail="Not authenticated")
+        raise refused(401, "not_authenticated")
     # SA4: reject a cookie whose recorded epoch is behind the account's current one (revoked by a
     # password reset/change or "log out everywhere"). A cookie with no recorded epoch (pre-SA4, or an
     # account signed in before this shipped) is treated as epoch 0 — so the FIRST bump revokes it too.
     epochs = request.session.get("epochs") or {}
     if int(epochs.get(str(user_id), 0)) != (user.session_epoch or 0):
         _end_account_session(request, user_id, is_default)
-        raise HTTPException(status_code=401, detail="Not authenticated")
+        raise refused(401, "not_authenticated")
     # Always keep the session default in the switchable list — covers sessions created before
     # multi-session existed (their account_ids was never seeded), so the switcher isn't empty.
     default_id = request.session.get("user_id")
@@ -348,6 +362,17 @@ def _session_user(request: Request, db: Session, user_id: int | None, is_default
     if default_id and default_id not in accounts:
         accounts.append(default_id)
         request.session["account_ids"] = accounts[-MAX_SESSION_ACCOUNTS:]
+    # Attribute this request's log lines to the account (E22). Here, not in the request middleware:
+    # this is the one place a session becomes an account, and the middleware runs outside both the
+    # session middleware and this task's context. `request.state` is shared across middleware layers
+    # (it lives on the ASGI scope), so the access-log line can read it back. The user's consented
+    # diagnostics level rides along: their requests are logged — and stored — in that detail (S5).
+    # So does a debug session the request names (E16): only the caller's own open one counts, and a
+    # header naming anything else is ignored rather than refused — a stale tab must keep working.
+    session = debugsession.from_header(db, user, request.headers.get(debugsession.HEADER))
+    level, until = debugsession.level_for(user, session)
+    logs.set_user(user.id, level, until, session.id if session else None)
+    request.state.user_id = user.id
     return user
 
 
@@ -361,12 +386,20 @@ def optional_current_user(request: Request, db: Session = Depends(get_db)) -> Us
     user_id, is_default = resolved_user_id(request, strict=False)
     try:
         return _session_user(request, db, user_id, is_default)
+    # silent-ok: this is the anonymous case as much as the refused one — every landing page hit
+    # comes through here. The second attempt below is the handling, and it reports.
     except HTTPException:
         if is_default:
             return None
     try:
         return _session_user(request, db, *resolved_user_id(request, strict=False))
     except HTTPException:
+        # Both the pinned account AND the default one were refused, so the tab is signed out
+        # without being told why — and it got here from a session that named a real account.
+        log.info(
+            "a session was resolved to nobody: neither the named nor the default account held",
+            extra=logs.event("auth.session.probe_rejected", user_id=user_id),
+        )
         return None
 
 
@@ -390,10 +423,7 @@ def require_write_scope(user: User = Depends(require_human)) -> User:
     checks that had drifted into three different 403 messages across channels/playlists. Layers on
     `require_human` because writing to YouTube needs a real, non-demo identity."""
     if not has_write_scope(user):
-        raise HTTPException(
+        raise refused(403, "youtube_editing_off")
-            status_code=403,
-            detail="Enable YouTube editing in Settings first.",
-        )
     return user
 
 
@@ -401,7 +431,7 @@ def admin_user(user: User = Depends(current_user)) -> User:
     """Dependency: require the signed-in user to be an admin. Single source of truth for the
     admin gate — every admin route/action depends on this rather than re-checking the role."""
     if user.role != "admin":
-        raise HTTPException(status_code=403, detail="Admin only")
+        raise refused(403, "admin_only")
     return user
 
 

backend/app/bodylimit.py

@@ -0,0 +1,231 @@
+"""The global request-body cap (SEC-20).
+
+**The hole.** FastAPI reads a declared body IN FULL before auth or any other dependency runs — the
+body parameter has to be parsed to be passed in. So `POST /auth/password-login` drains a 32 MiB
+upload from an anonymous caller and only then answers 401. 70 of the 72 body-declaring routes behave
+the same way: one unauthenticated client can spend the process's memory and time at will.
+
+**Why a middleware and not `capped_body` on every route.** `routes/_common.py:capped_body` is the
+right tool where a route genuinely takes bulk bytes, and the two routes that use it use it
+correctly. But putting it on all 70 others would (a) fight the canonical "Request body → a Pydantic
+model" row, since capping means taking the body as raw bytes and parsing it by hand, and (b) leave
+the NEXT route added uncovered, because nothing would remind its author. A default that has to be
+opted into is not a default. This module inverts it: the cap applies to everything, and the few
+routes that need more opt OUT, in one list, next to the code that reads it.
+
+**Why raw ASGI and not `@app.middleware("http")`.** Starlette's BaseHTTPMiddleware cannot wrap
+`receive`, so it can only read `Content-Length` — which a client may omit entirely (chunked) or
+simply understate. The bytes have to be counted as they arrive, which is the same rule `capped_body`
+already applies per-route, and counting needs the receive channel.
+
+**Why the refusal is byte-identical to `capped_body`'s.** Two shapes for one fact is how the
+`cancelled`/`canceled` bug happened. The 413 body here is built through the same `JSONResponse`
+FastAPI renders `HTTPException(413, {"code": "too_large"})` with — not a hand-rolled `json.dumps`,
+whose default separators differ by two bytes — and `tests/test_body_limit.py` compares the RAW BYTES
+of the two against each other rather than against a literal, so they cannot drift apart. The same
+reasoning decides where the middleware is registered (`main.py`): just inside CORS, so this 413
+carries the same headers `capped_body`'s does.
+
+**No admin knob, deliberately.** The ceiling is a constant, not a `sysconfig` value: an attacker-
+facing limit that an admin can raise from the UI is a limit that can be raised by whoever gets the
+admin session, and nobody has asked to tune it. If a self-hoster ever needs a different ceiling it
+belongs in `config.py` (env, set before the process starts), not in the database.
+"""
+
+import re
+
+from starlette.responses import JSONResponse
+from starlette.types import ASGIApp, Message, Receive, Scope, Send
+
+from app import debugsession
+from app.config import settings
+
+# 1 MiB. The largest body any route on the default ceiling legitimately takes is 512 KiB (the client
+# log sink, `routes/diagnostics.py:BODY_MAX_BYTES`), so this leaves 2x headroom — a ceiling that
+# refuses a legitimate request would be a worse bug than the one this module fixes. That ratio is
+# the load-bearing fact, so `test_body_limit.py` holds it rather than the figure above.
+DEFAULT_MAX_BODY_BYTES = 1024 * 1024
+
+# The routes that need MORE than the default, as (method, path TEMPLATE) → their own ceiling. The
+# template, not a concrete path, so the list reads like the route decorators it mirrors and
+# `test_body_limit.py` can assert that each one still resolves to a real route: a renamed route
+# fails the gate instead of silently losing its allowance in prod.
+#
+# These are RAISED CEILINGS, NOT EXEMPTIONS, and the difference is the whole point. An exemption
+# would mean "this route is counted by something else" — which is only true when the route's own
+# limit applies to the bytes as they ARRIVE. That, in turn, is only true of a route with NO DECLARED
+# BODY PARAMETER: FastAPI parses and spools a declared body BEFORE the first dependency runs, so a
+# route that declares one has already stored the bytes by the time its own cap — or its auth — gets
+# a say.
+#
+# Both listed routes now satisfy that. The screenshot route always did (`capped_body` streams it,
+# and the owner check resolves first). The upload route did NOT until P1: it declared
+# `file: UploadFile = File(...)`, so an anonymous caller could make the process write the whole
+# per-file allowance to temp before `require_human` refused them — this middleware bounded that
+# number but could not stop the writing. It now reads its own body from a dependency declared after
+# the auth one, so its in-handler cap measures bytes as they arrive, like everything else here.
+#
+# That rule is no longer only written here: `test_body_limit.py` parses the routes and fails the
+# gate if any route in this dict declares a body parameter. A comment could not have stopped the
+# defect it describes; the test can.
+#
+# A route belongs here ONLY IF ITS OWN CAP IS LARGER THAN THE DEFAULT. That is why
+# `POST /api/diagnostics/events` is absent: its `capped_body` limit is 512 KiB, already below the
+# default, so listing it could only widen what it accepts.
+ROUTE_MAX_BODY_BYTES: dict[tuple[str, str], int] = {
+    # Ingesting a local file. The body IS the file since P1 — no multipart envelope to allow for —
+    # so this ceiling and the route's own cap count exactly the same bytes. `sysconfig`'s
+    # `download_upload_max_bytes` is bounded by this same constant (see sysconfig.SPECS), so an
+    # admin can only ever LOWER the route's cap below the edge, never raise it past one.
+    #
+    # WHICH LAYER ANSWERS, measured rather than assumed (deep-review round 1):
+    #   * knob LOWERED (an override below the default) — the route's cap is the smaller number, so
+    #     `capped_stream_to` trips first and answers 422 `upload_too_large` carrying the admin's own
+    #     figure, after draining the rest of the body so the client can read it.
+    #   * knob at the DEFAULT — the two numbers are EQUAL, and the edge always wins: a declared
+    #     Content-Length over the cap is refused here before a byte is read, and an undeclared body
+    #     trips this counter on the same chunk the route would have tripped on. The route's 422 is
+    #     unreachable in that configuration, and no amount of headroom would change it — the
+    #     declared-length check fires first for any body big enough to matter, and the route's drain
+    #     pushes the count past the edge for the rest.
+    #
+    # So this 413 is a refusal a real user meets on a default install, not a theoretical edge, and
+    # it has to be as readable as the route's own: the client localizes `too_large` (`errors.tooLarge`)
+    # exactly as it does `upload_too_large`, and the upload dialog refuses an over-cap file before
+    # sending it at all, off the `upload_max_bytes` that `GET /api/downloads/usage` reports.
+    ("POST", "/api/downloads/upload"): settings.download_upload_max_bytes,
+    ("POST", "/api/debug-sessions/{sid}/screenshots"): debugsession.SCREENSHOT_MAX_BYTES,
+}
+
+# The refusal, byte-for-byte what `capped_body`'s `HTTPException(413, {"code": "too_large"})` renders
+# as — built through the same JSONResponse so the separators match too, and asserted against a real
+# route's raw bytes in `test_body_limit.py` rather than against a literal here.
+TOO_LARGE_BODY = JSONResponse({"detail": {"code": "too_large"}}, status_code=413).body
+
+_PATH_PARAM = re.compile(r"\{[^/{}]+\}")
+
+
+def template_to_regex(template: str) -> re.Pattern[str]:
+    """`/api/debug-sessions/{sid}/screenshots` → a regex matching one concrete path.
+
+    A path parameter matches `[^/]+` — one segment, never across a `/`, so `{sid}` cannot swallow a
+    deeper path and hand an exemption to a route that was never granted one.
+    """
+    literals = [re.escape(part) for part in _PATH_PARAM.split(template)]
+    return re.compile("^" + "[^/]+".join(literals) + "$")
+
+
+def _declared_length(scope: Scope) -> int | None:
+    """The `Content-Length` header as an int, or None when absent or unparseable (a chunked body has
+    none, and a malformed one is not a number we may trust — both fall through to counting)."""
+    for name, value in scope.get("headers", ()):
+        if name == b"content-length":
+            try:
+                return int(value)
+            except ValueError:
+                # silent-ok: a malformed Content-Length is the caller's doing, not a fault of ours,
+                # and it is about to be answered anyway — returning None drops the request into the
+                # counting path, which refuses on the real bytes. Logging it would hand an anonymous
+                # client a free line in our log for every request it cares to malform.
+                return None
+    return None
+
+
+async def _send_too_large(send: Send) -> None:
+    await send(
+        {
+            "type": "http.response.start",
+            "status": 413,
+            "headers": [
+                (b"content-type", b"application/json"),
+                (b"content-length", str(len(TOO_LARGE_BODY)).encode()),
+            ],
+        }
+    )
+    await send({"type": "http.response.body", "body": TOO_LARGE_BODY})
+
+
+class BodyLimitMiddleware:
+    """Refuses any request body over its route's ceiling, before the application sees it.
+
+    Registered just inside CORS in `main.py` — so the 413 is decorated like every other response
+    and a browser can actually read it — but outside the session, the setup gate and every auth
+    dependency, none of which run for a request that is already too large.
+
+    Every request gets a finite ceiling: `ROUTE_MAX_BODY_BYTES` raises it for the two routes that
+    genuinely take bulk bytes, and there is no path that skips counting altogether.
+    """
+
+    def __init__(
+        self,
+        app: ASGIApp,
+        *,
+        limit: int = DEFAULT_MAX_BODY_BYTES,
+        routes: dict[tuple[str, str], int] | None = None,
+    ) -> None:
+        self.app = app
+        self.limit = limit
+        self._routes = tuple(
+            (method, template_to_regex(path), cap)
+            for (method, path), cap in (ROUTE_MAX_BODY_BYTES if routes is None else routes).items()
+        )
+
+    def limit_for(self, method: str, path: str) -> int:
+        """This request's ceiling: the route's own if it has one, else the global default. Every
+        request gets a finite number — there is no path through here that is simply not counted."""
+        for m, rx, cap in self._routes:
+            if m == method and rx.match(path):
+                return cap
+        return self.limit
+
+    async def __call__(self, scope: Scope, receive: Receive, send: Send) -> None:
+        if scope["type"] != "http":
+            await self.app(scope, receive, send)
+            return
+        limit = self.limit_for(scope.get("method", ""), scope.get("path", ""))
+
+        declared = _declared_length(scope)
+        if declared is not None and declared > limit:
+            # Refused on the declaration alone: nothing downstream is called, and not one byte of
+            # the body is read.
+            await _send_too_large(send)
+            return
+
+        seen = 0
+        refused = False
+        started = False
+
+        async def counting_receive() -> Message:
+            nonlocal seen, refused
+            message = await receive()
+            if message["type"] != "http.request":
+                return message
+            seen += len(message.get("body", b""))
+            if seen > limit:
+                # The app is blocked on this receive, so it has not answered yet. Telling it the
+                # client went away is the only in-band way to stop it reading: `request.stream()`
+                # turns an `http.disconnect` into ClientDisconnect, which unwinds the handler. Our
+                # own 413 goes out below, once the app has finished unwinding.
+                refused = True
+                return {"type": "http.disconnect"}
+            return message
+
+        async def guarded_send(message: Message) -> None:
+            nonlocal started
+            if refused and not started:
+                return  # discard the app's answer to the disconnect; the 413 below is the response
+            started = True
+            await send(message)
+
+        try:
+            await self.app(scope, counting_receive, guarded_send)
+        except Exception:  # only swallowed when WE caused it, see below
+            # ClientDisconnect (or whatever a handler raises on a truncated read) is the expected
+            # consequence of the disconnect we just faked, and the caller must get the 413 rather
+            # than a 500. Anything raised when we did NOT refuse is a real failure and is re-raised
+            # untouched.
+            if not refused:
+                raise
+
+        if refused and not started:
+            await _send_too_large(send)

backend/app/collation.py

@@ -0,0 +1,83 @@
+"""Alphabetical order in the reader's language (E30 S3).
+
+Why not the database default: the Postgres image is Alpine, and musl's `en_US.utf8` collation is BYTE
+order — every accented initial sorted after "Z" ("Zorró < Életrajzi < Ötödik"), in every A→Z list the
+server orders. Postgres ships the ICU collations on the same image (dev, test, prod alike), so a text
+sort names one: Hungarian puts "Ö" after "O" and "Cs" after "Cu", English folds the accents in.
+
+`collated` wraps an ORDER BY expression — and because a keyset cursor compares with the SAME expression
+(`feedquery.sort_keys`), the cursor pages in the same order it sorted in. `sorted_words` orders a Python
+list the same way through the database, so a list built in Python matches one built in SQL exactly.
+
+The feed's title sorts ride indexes built in these collations (migration 0096), and an index keeps the
+order of the ICU that built it. `postgres:16-alpine` is a floating tag: an image update can bring a newer
+ICU, after which those indexes disagree with the comparisons a query makes (a keyset page skips or
+repeats a title). `heal_drift` rebuilds them at startup once Postgres reports the version moved."""
+
+from typing import Any
+
+from sqlalchemy import Engine, SQLColumnExpression, column, func, select, text, values
+from sqlalchemy.orm import Session
+from sqlalchemy.sql.elements import ColumnElement
+from sqlalchemy.types import String
+
+from app import logs
+from app.mailtext import normalize
+
+log = logs.get_logger(__name__)
+
+ICU: dict[str, str] = {"en": "en-x-icu", "hu": "hu-x-icu"}
+
+
+def collation(lang: str | None) -> str:
+    """The ICU collation for a UI language; anything unknown sorts as English."""
+    return ICU[normalize(lang)]
+
+
+def collated(expr: SQLColumnExpression[Any], lang: str | None) -> ColumnElement[Any]:
+    """`expr` compared in the reader's alphabet. Case-folded, so "apple" and "Apple" sit together."""
+    return func.lower(expr).collate(collation(lang))
+
+
+def sorted_words(db: Session, words: list[str], lang: str | None) -> list[str]:
+    """`words` in the reader's alphabetical order, by the same ICU collation as `collated`."""
+    if not words:
+        return []
+    w = values(column("w", String), name="words").data([(x,) for x in words])
+    return list(db.scalars(select(w.c.w).order_by(collated(w.c.w, lang), w.c.w)))
+
+
+def heal_drift(engine: Engine) -> list[str]:
+    """Rebuild every index on an ICU collation whose library version moved, then record the new version.
+    Returns the rebuilt indexes. CONCURRENTLY: the feed keeps reading while it runs."""
+    with engine.connect().execution_options(isolation_level="AUTOCOMMIT") as conn:
+        drifted = conn.execute(
+            text(
+                "SELECT oid, collname FROM pg_collation WHERE collname = ANY(:names) "
+                "AND collversion IS DISTINCT FROM pg_collation_actual_version(oid)"
+            ),
+            {"names": list(ICU.values())},
+        ).all()
+        if not drifted:
+            return []
+        indexes = list(
+            conn.execute(
+                text(
+                    "SELECT DISTINCT d.objid::regclass::text FROM pg_depend d "
+                    "JOIN pg_class c ON c.oid = d.objid AND c.relkind = 'i' "
+                    "WHERE d.classid = 'pg_class'::regclass "
+                    "AND d.refclassid = 'pg_collation'::regclass AND d.refobjid = ANY(:oids)"
+                ),
+                {"oids": [row.oid for row in drifted]},
+            ).scalars()
+        )
+        log.warning(
+            "ICU collation version changed (%s): rebuilding %d indexes",
+            ", ".join(row.collname for row in drifted),
+            len(indexes),
+        )
+        for name in indexes:
+            conn.execute(text(f"REINDEX INDEX CONCURRENTLY {name}"))
+        for row in drifted:
+            conn.execute(text(f'ALTER COLLATION "{row.collname}" REFRESH VERSION'))
+        return indexes

backend/app/config.py

@@ -1,6 +1,7 @@
 from pathlib import Path
+from typing import Literal, get_args, get_origin
 
-from pydantic import model_validator
+from pydantic import Field, ValidationInfo, field_validator, model_validator
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 # Shipped placeholder; production must override it. Used by the startup guard below.
@@ -15,11 +16,40 @@ def _version_from_file() -> str:
             v = p.read_text(encoding="utf-8").strip()
             if v:
                 return v
+        # silent-ok: this runs at import, and `app.logs` imports THIS module — a log call here
+        # would be a circular import. The loop tries the other location, then falls back to "dev".
         except OSError:
             continue
     return "dev"
 
 
+# Closed value sets. A setting typed with one of these is a CHOICE: pydantic rejects anything else at
+# boot, and sysconfig derives the admin picker's options from the same Literal — so the list lives in
+# exactly one place and a consumer never has to guess what a typo should fall back to.
+SearchSource = Literal["scrape", "api"]
+DownloadLayout = Literal["plex", "flat"]
+LogLevel = Literal["TRACE", "DEBUG", "INFO", "WARNING", "ERROR"]
+LogFormat = Literal["json", "text"]
+TranscodeHw = Literal["auto", "on", "off"]
+# libx264's presets, fastest first (`placebo` left out on purpose: no measurable gain for its cost).
+TranscodePreset = Literal[
+    "ultrafast", "superfast", "veryfast", "faster", "fast", "medium", "slow", "slower", "veryslow"
+]
+
+
+def choices_of(annotation: object) -> tuple[str, ...]:
+    """The allowed values of a choice-typed setting, or () for any other type."""
+    return tuple(get_args(annotation)) if get_origin(annotation) is Literal else ()
+
+
+def canonical_choice(options: tuple[str, ...], raw: object) -> str | None:
+    """The option `raw` names, in its canonical spelling (any letter case, surrounding space
+    ignored), or None. The ONE matching rule for a choice value from the env and from the admin
+    API, so a spelling one accepts the other cannot refuse."""
+    folded = str(raw).strip().casefold()
+    return next((option for option in options if option.casefold() == folded), None)
+
+
 class Settings(BaseSettings):
     model_config = SettingsConfigDict(env_file=".env", extra="ignore")
 
@@ -85,10 +115,13 @@ class Settings(BaseSettings):
     search_daily_limit_per_user: int = 70
     # Source for live YouTube search: "scrape" (InnerTube, no API quota) or "api" (search.list,
     # 100 units/page). Scrape is the default; flip to "api" if scraping is blocked/breaks.
-    search_source: str = "scrape"
+    search_source: SearchSource = "scrape"
     # Recent-first backfill: how far back to fetch on the first pass per channel.
     backfill_recent_max_videos: int = 100
     backfill_recent_max_days: int = 365
+    # A channel whose RSS feed has not answered for this many hours gets its newest uploads read
+    # through the Data API instead (1 unit), at most once per this many hours. 0 = never.
+    rss_fallback_hours: int = 2
 
     # Shorts are confirmed by probing youtube.com/shorts/<id>. Only videos at or below
     # this duration (and not livestreams) are probed; longer videos are never Shorts.
@@ -155,6 +188,9 @@ class Settings(BaseSettings):
     # per-client responses are flaky, and android_vr can get bot-flagged under heavy load) the
     # worker retries with the FALLBACK set — the POT-capable web clients (web needs the token +
     # Deno) plus android for audio. Both admin-tunable when YouTube shifts what works.
+    # A SIGNED-IN download (the user added YouTube cookies) keeps only the cookie-capable clients
+    # of these two sets — yt-dlp drops the rest, so android_vr / android / ios never run there.
+    # See adapters.YouTubeAdapter._COOKIE_CAPABLE.
     download_player_clients: str = "android_vr"
     download_player_clients_fallback: str = "web_safari,web,android"
 
@@ -211,7 +247,11 @@ class Settings(BaseSettings):
     # Per-file cap for local uploads (E7). Unlike a download (whose size is unknown until the worker
     # finishes), an upload's size is known at ingest, so this bounds a single file directly; the
     # user's overall byte footprint is still governed by download_default_max_bytes.
-    download_upload_max_bytes: int = 524_288_000  # 500 MiB
+    # ge=1: this number is also the EDGE ceiling (`bodylimit.ROUTE_MAX_BODY_BYTES`), where 0 is a
+    # real limit of zero rather than the route's "no per-file cap" sentinel — an instance set to 0
+    # would refuse every upload, including the admin page's own way back. To lift the cap in
+    # practice, set a large number.
+    download_upload_max_bytes: int = Field(default=524_288_000, ge=1)  # 500 MiB
     # Retention: an asset expires this many days after its last access; the GC job then
     # deletes it (once no job references it) after an additional grace window.
     download_retention_days: int = 30
@@ -220,11 +260,31 @@ class Settings(BaseSettings):
     # Admin audit log: how long to keep rows (days; 0 = keep forever) + how often the GC runs.
     audit_retention_days: int = 180
     audit_gc_minutes: int = 1440
+    # YouTube quota history (quota_events): days kept (0 = forever). A year, so the admin usage view's
+    # longest window (365 days) stays whole; the same daily cleanup job prunes it.
+    quota_retention_days: int = 365
     # On-disk layout: "plex" (Channel/Season YYYY/… [id].ext + .nfo/poster) or "flat".
-    download_layout: str = "plex"
+    download_layout: DownloadLayout = "plex"
     # Default SponsorBlock (skip sponsor segments) for new custom profiles.
     sponsorblock_default: bool = False
 
+    # --- Diagnostics / logging (E22) ---
+    # BOOTSTRAP level only: the running level is the `log_level` admin knob (sysconfig), applied at
+    # startup and on change without a restart. This is what a process logs at before it can reach
+    # the database — and the only knob the worker/runner children read, since they are short-lived.
+    # TRACE (below DEBUG) is the "Teljes" diagnostics level; see app/logs.py.
+    log_level: LogLevel = "INFO"
+    # "json" (one object per line — what Loki and the log_events sink ingest) or "text" (a readable
+    # line for a local terminal). Redaction applies to both.
+    log_format: LogFormat = "json"
+    # Also write events to the `log_events` table (app/logsink.py): WARNING and above, the request
+    # lines, and — once they exist — client batches. Off only where no table is wanted (the tests).
+    log_db_sink: bool = True
+    # Days a `log_events` row is kept (the privacy policy promises this), and how often the log_gc job
+    # enforces it — hourly, so a row never outlives the promise by more than an hour.
+    log_retention_days: int = 7
+    log_gc_minutes: int = 60
+
     # --- Plex integration (optional) ---
     # The Plex module plays the LOCAL physical media file directly (our own ffmpeg/direct-serve);
     # Plex is used only for metadata + optional watch-state sync. Requires the Plex server to be
@@ -262,11 +322,15 @@ class Settings(BaseSettings):
     # re-encode is CPU-heavy (the VAAPI hardware path is far cheaper and could sustain more); direct-serve
     # and remux (stream-copy) are not capped by this.
     plex_max_transcodes: int = 4
+    # Max concurrent playback sessions of ANY kind — remux (stream-copy) included. Cheap on the CPU but
+    # each one is a whole segmented copy of the film on the scratch volume, written faster than
+    # realtime. One account holds at most two regardless (plex/stream.py `_MAX_SESSIONS_PER_USER`).
+    plex_max_sessions: int = 8
     # --- Transcode (playable="transcode") re-encode tuning (E6 S1) ---
     # Hardware-encoder policy: "auto" = VAAPI (h264_vaapi) when the iGPU render node is present AND the
     # source codec is hardware-decodable, else libx264 software; "on" = force VAAPI wherever the render
     # node exists (skip the decodability gate); "off" = always software. See plex/stream.py `_select_encode`.
-    plex_transcode_hw: str = "auto"
+    plex_transcode_hw: TranscodeHw = "auto"
     # Software-fallback libx264 tuning (IGNORED on the VAAPI path). `preset` trades CPU for size; `crf`
     # trades quality for size (lower = better + bigger). `max_height` caps the SOFTWARE transcode's output
     # height (0 = keep source) so a CPU re-encode of a tall source stays affordable — the GPU path always
@@ -276,7 +340,7 @@ class Settings(BaseSettings):
     # call, not a fix. Only affects files that go through remux/transcode — a direct-played file is
     # served untouched, but those are browser-native stereo anyway, not the 5.1 mixes this is for.
     plex_audio_loudnorm: bool = False
-    plex_transcode_preset: str = "veryfast"
+    plex_transcode_preset: TranscodePreset = "veryfast"
     plex_transcode_crf: int = 20
     plex_transcode_max_height: int = 1080
     # Where on-the-fly HLS remux segments are written (transient, reaped). Empty → a `.plex-hls`
@@ -294,10 +358,22 @@ class Settings(BaseSettings):
     tmdb_api_base: str = "https://api.themoviedb.org/3"
     # Metadata language for TMDB fetches (keyword/overview locale). Stored EN by default.
     tmdb_language: str = "en-US"
+
+    # --- Bug reports (E16): where a debug session's issue goes. All three empty = no tracker; the
+    # composer still exports a zip. The repository must be PRIVATE — an issue carries the reporter's
+    # log — and app/issues/gitea.py refuses a public one. Runtime knobs (sysconfig, admin UI). ---
+    issues_gitea_url: str = ""
+    issues_gitea_repo: str = ""
+    issues_gitea_token: str = ""
     # How often the TMDB enrichment job fetches a batch of not-yet-enriched Plex titles, in minutes.
     tmdb_enrich_interval_min: int = 120
     # Titles fetched per TMDB enrichment run (TMDB is fast + generously rate-limited).
     tmdb_enrich_batch: int = 100
+    # How often the people job (E30 S6) fetches a batch of person bios from TMDB and portraits through Plex,
+    # in minutes, and how many of each per run: ~18k linked people at 300 per 10 min is about ten hours
+    # once, then only the newcomers.
+    people_enrich_interval_min: int = 10
+    people_enrich_batch: int = 300
 
     @property
     def allowed_email_set(self) -> set[str]:
@@ -325,6 +401,17 @@ class Settings(BaseSettings):
         https OAUTH_REDIRECT_URL as the signal for 'real deployment' vs local dev."""
         return self.oauth_redirect_url.lower().startswith("https://")
 
+    @field_validator("*", mode="before")
+    @classmethod
+    def _canonical_choice(cls, value: object, info: ValidationInfo) -> object:
+        """A choice setting accepts its value in any letter case (`LOG_LEVEL=debug`) and stores the
+        canonical spelling; an unknown value still fails the boot with pydantic's own message."""
+        field = cls.model_fields.get(info.field_name or "")
+        options = choices_of(field.annotation) if field else ()
+        if options and isinstance(value, str):
+            return canonical_choice(options, value) or value
+        return value
+
     @model_validator(mode="after")
     def _enforce_production_secrets(self) -> "Settings":
         """Fail fast rather than boot a public instance with the shipped placeholder

backend/app/db.py

@@ -1,6 +1,6 @@
 from collections.abc import Iterator
 
-from sqlalchemy import create_engine
+from sqlalchemy import create_engine, text
 from sqlalchemy.orm import DeclarativeBase, Session, sessionmaker
 
 from app.config import settings
@@ -12,7 +12,9 @@ class Base(DeclarativeBase):
 
 # Pool sized above the sync-endpoint threadpool (~40) so a burst of short DB sessions can't starve on
 # checkout. The image proxies deliberately DON'T hold a connection during their slow external fetch
-# (see routes/plex/), so this headroom is only for the many brief sessions, not long ones.
+# (see routes/plex/), so this headroom is only for the many brief sessions, not long ones. The one
+# long hold — a playlist sync's advisory claim, for a whole push — opens its own unpooled connection
+# (sync/playlists.py `_claim_engine`) instead of drawing a third one from here.
 engine = create_engine(
     settings.database_url,
     pool_pre_ping=True,
@@ -20,6 +22,10 @@ engine = create_engine(
     max_overflow=30,
     pool_timeout=30,
     future=True,
+    # A database error renders its bound parameters into its message — a title, an uploader, a
+    # file path, a URL — and that message is logged and stored as a job's error text. The SQL
+    # stays; the values are data, and the ids an operator needs are on the line already.
+    hide_parameters=True,
 )
 SessionLocal = sessionmaker(bind=engine, autoflush=False, expire_on_commit=False)
 
@@ -30,3 +36,50 @@ def get_db() -> Iterator[Session]:
         yield db
     finally:
         db.close()
+
+
+def release_read_connection(db: Session) -> None:
+    """End a request's READ transaction so its pooled connection goes back BEFORE a long await.
+
+    SQLAlchemy 2.0 autobegins on the first SELECT, so a request that has merely authenticated is
+    already holding one of the connections above — `idle in transaction` to Postgres, and nothing
+    times it out. A route that then awaits something slow (a request body still arriving, an
+    external fetch) pins it for the whole wait, which is precisely the invariant the pool is sized
+    on in the comment above. Measured: the next checkout after this call reuses the very backend
+    that was released, and a 1.8 s upload goes from one sustained `idle in transaction` to none.
+
+    Every route that reads its own body from a dependency needs this, and `test_body_limit.py`
+    holds that — the three that do are the ones with a raised (or self-imposed) body ceiling, and
+    they got here one at a time, which is how the upload route ended up the only one fixed.
+
+    The caller must have NOTHING PENDING, and that is checked rather than described. The session is
+    shared with every dependency that ran before, so a rollback here would silently discard a write
+    one of them staged — `service.enqueue` already learned that lesson the expensive way and uses a
+    SAVEPOINT for it. CLAUDE.md: a rule stated in a comment must also be enforced in code.
+
+    TWO checks, because the obvious one is blind to the shapes this codebase actually writes in.
+    `db.new`/`deleted`/`dirty` hold only UNFLUSHED unit-of-work: the moment a helper calls
+    `db.flush()` — CLAUDE.md's canonical split is "the handler commits; helpers flush" — the object
+    is persistent with no unflushed history and all three sets are empty again, and Core DML
+    (`db.execute(update(...))`) never enters them at all. Measured: after `add`+`flush` the sets are
+    empty, the old check passed, and the rollback lost the row. So the transaction is asked
+    directly. Postgres assigns a transaction id on the first WRITE and never for a read, so
+    `pg_current_xact_id_if_assigned()` is NULL for exactly the state this function is safe in — and
+    it sees a raw `text("UPDATE …")` too, which no session-level bookkeeping would.
+    """
+    if db.new or db.deleted or db.dirty:
+        raise RuntimeError(
+            "release_read_connection() with pending work: something ahead of this dependency "
+            "staged a write, which ending the transaction would discard. Commit it first, or do "
+            "it after the await."
+        )
+    if (
+        db.in_transaction()
+        and db.execute(text("SELECT pg_current_xact_id_if_assigned()")).scalar() is not None
+    ):
+        raise RuntimeError(
+            "release_read_connection() on a transaction that has already WRITTEN: something ahead "
+            "of this dependency flushed a row or ran an UPDATE/INSERT/DELETE, which ending the "
+            "transaction would discard. Commit it first, or do it after the await."
+        )
+    db.rollback()

backend/app/debugsession.py

@@ -0,0 +1,485 @@
+"""Debug sessions (E16): Start → reproduce → Stop, and the evidence a bug report is built from.
+
+While a session is open its user's browser names it in `X-Debug-Session` on every request. The auth
+dependency accepts the header only for the caller's OWN open session (anything else is ignored, never
+an error — a stale tab must not break the app) and then logs that request at the session's level with
+every line tagged by its id; the browser tags its own events the same way. The issue is later built
+from exactly the rows carrying the tag, so what the review list shows is what gets sent.
+
+Starting a session IS the consent for its level, for its duration: the start dialog carries the
+consent text. It needs no change to the user's own diagnostics level, and none is made — when the
+session ends the level in force is simply the user's own again.
+
+A session is capped at `MAX_MINUTES`, set by the server at Start. An open one past its cap is over
+whether or not anything has closed it yet: it reads as over at once, a new Start closes it, and the
+hourly log job sweeps the rest. Sessions go with the logs they point at, after `log_retention_days`.
+
+Core layer: imports `models`, `diagnostics`.
+"""
+
+import secrets
+from datetime import datetime, timedelta, timezone
+from enum import StrEnum
+from typing import Literal, get_args
+from uuid import uuid4
+
+from sqlalchemy import and_, delete, func, not_, select, update
+from sqlalchemy.orm import Session, defer
+
+from app import diagnostics
+from app.models import DebugScreenshot, DebugSession, LogEvent, User
+
+HEADER = "x-debug-session"
+MAX_MINUTES = diagnostics.SESSION_MINUTES
+SCREENSHOT_MAX_BYTES = 10 * 1024 * 1024
+SCREENSHOTS_MAX = 8
+# What one user may keep in screenshots across ALL their sessions (until retention takes them): the
+# per-session cap alone would let Start → 8 images → Stop repeat without end.
+SCREENSHOT_USER_MAX_BYTES = 100 * 1024 * 1024
+# The screenshot types kept, and the extension each is exported and sent under — one table, so an
+# accepted type always has a name.
+IMAGE_TYPES = {"image/png": "png", "image/jpeg": "jpg", "image/webp": "webp"}
+# What a screenshot is called in the zip and on the issue. `issues/bundle.py` builds the names from
+# this and `IMAGE_TYPES`; this module and `issues/gitea.py` read them back to tell a screenshot's
+# link from a log's. One definition, because three copies of a name is how the next rename breaks
+# `images_pending` into silence.
+SCREENSHOT_PREFIX = "screenshot-"
+# Rows an issue can carry. A 60-minute session at Teljes is a few thousand rows; past this the
+# bundle says how many it left out rather than growing without bound.
+BUNDLE_MAX = 5_000
+
+SessionLevel = Literal["DEBUG", "TRACE"]
+SESSION_LEVELS: tuple[str, ...] = get_args(SessionLevel)
+# Detail order of the diagnostics levels: a session raises the level in force, never lowers it.
+_DETAIL = {"INFO": 0, "DEBUG": 1, "TRACE": 2}
+_ID_LEN = 16
+
+
+class EndReason(StrEnum):
+    STOPPED = "stopped"
+    EXPIRED = "expired"
+
+
+def is_live(session: DebugSession, now: datetime | None = None) -> bool:
+    """Open and within its cap — THE liveness rule: an open session past its cap is over."""
+    return session.ended_at is None and session.until > (now or datetime.now(timezone.utc))
+
+
+def open_session(db: Session, user: User) -> DebugSession | None:
+    """The user's live session, or None. Read-only: one past its cap is over and reads as None;
+    `start` and the hourly sweep are what close it."""
+    session = _unclosed(db, user)
+    if session is None or not is_live(session, datetime.now(timezone.utc)):
+        return None
+    return session
+
+
+def _unclosed(db: Session, user: User) -> DebugSession | None:
+    return db.scalar(
+        select(DebugSession).where(DebugSession.user_id == user.id, DebugSession.ended_at.is_(None))
+    )
+
+
+def start(db: Session, user: User, level: SessionLevel) -> DebugSession:
+    """Open a session, or return the live one — a second tab pressing Start joins the recording
+    rather than ending it. One past its cap is closed first: the one-open-per-user index would
+    refuse the new row beside it. Flushes; the caller commits. Two concurrent Starts can still both
+    get here: the index refuses the second, and the caller joins the one that won."""
+    now = datetime.now(timezone.utc)
+    current = _unclosed(db, user)
+    if current is not None:
+        if is_live(current, now):
+            return current
+        end(current, EndReason.EXPIRED)
+        db.flush()
+    session = DebugSession(
+        id=secrets.token_hex(_ID_LEN // 2),
+        user_id=user.id,
+        level=level,
+        started_at=now,
+        until=now + timedelta(minutes=MAX_MINUTES),
+    )
+    db.add(session)
+    db.flush()
+    return session
+
+
+def stop(session: DebugSession) -> None:
+    """The user's Stop. Past its cap the session had already ended on its own, and a late Stop must
+    not rename that."""
+    end(session, EndReason.STOPPED if is_live(session) else EndReason.EXPIRED)
+
+
+def end(session: DebugSession, reason: EndReason) -> None:
+    """Close a session. An expired one is dated at its cap, not at whenever it was noticed."""
+    if session.ended_at is not None:
+        return
+    session.ended_at = (
+        session.until
+        if reason is EndReason.EXPIRED
+        else min(datetime.now(timezone.utc), session.until)
+    )
+    session.ended_reason = reason
+
+
+def from_header(db: Session, user: User, value: str | None) -> DebugSession | None:
+    """The session a request names, if it is the caller's own and still live; else None."""
+    if not value or len(value) != _ID_LEN or not all(c in "0123456789abcdef" for c in value):
+        return None
+    session = db.get(DebugSession, value)
+    if (
+        session is None
+        or session.user_id != user.id
+        or not is_live(session, datetime.now(timezone.utc))
+    ):
+        return None
+    return session
+
+
+def level_for(user: User, session: DebugSession | None) -> tuple[str, datetime | None]:
+    """The diagnostics level (and its expiry) a request is logged at: the user's own, raised to the
+    session's while one is named — whichever records more.
+
+    A named session also caps the expiry at its own: this pair is what the process tracks under the
+    session (app/logs.py), and an expiry outliving the recording would hold the root logger down
+    after it ended."""
+    own = diagnostics.effective_level(user)
+    until = user.diagnostics_until if own != diagnostics.BASE else None
+    if session is None:
+        return own, until
+    if _DETAIL[session.level] <= _DETAIL[own]:
+        return own, min(until, session.until) if until else session.until
+    return session.level, session.until
+
+
+def lock(db: Session, session: DebugSession) -> None:
+    """Re-read the session under a row lock, held until the caller commits. Everything that changes
+    what an issue is made of — the draft, the screenshots, the send itself — takes it first: a send
+    in flight sees none of them change under it, and each of them sees a send that committed."""
+    db.refresh(session, with_for_update=True)
+
+
+def save_draft(session: DebugSession, draft: dict) -> None:
+    """Keep the composer's draft (the caller validated, locked and commits). Shape: `title`, `body`
+    (the rendered markdown an export or an issue carries), `excluded` (log row ids taken out of the
+    evidence) and `form` (the composer's own fields, to reopen it as it was). A delivery's progress
+    is not the composer's to replace, and is kept."""
+    progress = (session.draft or {}).get("delivery")
+    session.draft = {**draft, "delivery": progress} if progress else draft
+
+
+# --- a delivery's progress -----------------------------------------------------------------
+# Written by app/issues/gitea.py into the draft; read HERE, so the send's decision and what the
+# composer shows are one reading of it.
+def delivery(session: DebugSession) -> dict:
+    return dict((session.draft or {}).get("delivery") or {})
+
+
+def not_uploaded(session: DebugSession) -> list[str]:
+    """The files a send that stopped half-way has still to upload — by name, as the issue carries
+    them. What the composer names to the user."""
+    return list(delivery(session).get("pending") or [])
+
+
+def images_pending(session: DebugSession) -> bool:
+    """Whether the issue's body still has to be edited to show the screenshots it already carries
+    (every upload went, the edit that shows them did not)."""
+    state = delivery(session)
+    shown = state.get("images_shown")
+    return not shown and any(n.startswith(SCREENSHOT_PREFIX) for n in state.get("links") or {})
+
+
+def unfinished(session: DebugSession) -> bool:
+    """Whether a send owes the issue anything — the retry's question."""
+    return bool(not_uploaded(session)) or images_pending(session)
+
+
+# A send claims the session while it talks to the tracker, instead of holding the row's lock across
+# those calls: a Stop, an autosave or a screenshot would queue behind a lock for the whole send.
+# They see the claim and are refused at once instead. A claim not renewed for this long was left by
+# a request that died mid-send (the process restarted), and the next send takes it over — which is
+# why a live send RENEWS it every time it reports progress (`set_delivery`). A claim is a lease, and
+# a send slower than the lease must not lose it while it is still running.
+SEND_CLAIM_SECONDS = 600
+
+
+def _claim(session: DebugSession) -> dict:
+    claim = delivery(session).get("sending")
+    return claim if isinstance(claim, dict) else {}
+
+
+def _stamp(token: str) -> dict:
+    return {"at": datetime.now(timezone.utc).isoformat(), "by": token}
+
+
+def sending(session: DebugSession) -> bool:
+    """Whether a send is under way right now (a live claim)."""
+    at = _claim(session).get("at")
+    if not at:
+        return False
+    return (datetime.now(timezone.utc) - datetime.fromisoformat(at)).total_seconds() < (
+        SEND_CLAIM_SECONDS
+    )
+
+
+def claim_send(session: DebugSession) -> str:
+    """Mark the session as being sent, and return the token that OWNS this claim. The caller holds
+    the lock and commits, which is what makes the claim visible to everyone else."""
+    token = uuid4().hex
+    _write_delivery(session, {**delivery(session), "sending": _stamp(token)})
+    return token
+
+
+def release_send(session: DebugSession, token: str) -> None:
+    """Give back our OWN claim, and only ours: one that aged out and was taken over belongs to the
+    send holding it now, and handing that one back would let a third send in beside it."""
+    if _claim(session).get("by") != token:
+        return
+    _write_delivery(session, {k: v for k, v in delivery(session).items() if k != "sending"})
+
+
+class SendClaimLost(Exception):
+    """This send's claim is no longer on the row: it aged out and another send took it over. The
+    send must stop — what it uploads from here it can no longer record, so a retry would send it
+    again — and it must not write, because the row now holds the other send's progress."""
+
+
+def claim_owner(db: Session, session: DebugSession) -> str | None:
+    """Who holds the row's claim RIGHT NOW, read from the ROW and not from the object.
+
+    `expire_on_commit` is off (app/db.py), so a checkpoint does not re-read: a send that stalled
+    past its lease still carries its own dead token in memory long after another send took over.
+    A targeted scalar, not `lock`/`db.refresh` — a refresh here would discard the `issue_url` that
+    `gitea.deliver` assigns one line before its first progress report, and that link is the one
+    thing the send must not lose."""
+    return db.scalar(
+        select(DebugSession.draft["delivery"]["sending"]["by"].as_string()).where(
+            DebugSession.id == session.id
+        )
+    )
+
+
+def set_delivery(db: Session, session: DebugSession, state: dict, token: str) -> None:
+    """Write a delivery's progress (app/issues/gitea.py keeps it; this module reads it), and renew
+    the claim while doing it — a send still reporting progress is still alive, so one slower than
+    the lease cannot expire while it is still running.
+
+    The claim is not the caller's to carry: `gitea.deliver` threads ONE snapshot of this dict
+    through a whole send, so a progress report that carried the claim wrote a stale one back at
+    every step. It is stamped from `token` here instead, and only after the ROW confirms the claim
+    is still ours. Without that check the write is a full-column overwrite with no WHERE: a send
+    that lost its lease came back and stomped the take-over send's committed links, then put its
+    own dead token back on an idle session (deep-review round 4). `release_send` has always made
+    the same check; this is the other half of it."""
+    if claim_owner(db, session) != token:
+        raise SendClaimLost(session.id)
+    state = {k: v for k, v in state.items() if k != "sending"}
+    state["sending"] = _stamp(token)
+    _write_delivery(session, state)
+
+
+def _write_delivery(session: DebugSession, state: dict) -> None:
+    # A new dict, so the JSONB column sees the change.
+    session.draft = {**(session.draft or {}), "delivery": state}
+
+
+def serialize(session: DebugSession | None, screenshots: int = 0) -> dict | None:
+    if session is None:
+        return None
+    return {
+        "id": session.id,
+        "level": session.level,
+        "started_at": session.started_at.isoformat(),
+        "until": session.until.isoformat(),
+        "ended_at": session.ended_at.isoformat() if session.ended_at else None,
+        "ended_reason": session.ended_reason,
+        "issue_url": session.issue_url,
+        # A send that stopped half-way: the files it has still to upload, and whether the issue's
+        # body has still to be edited to show the screenshots it already carries. Both are true of a
+        # send that is RUNNING too — every step commits its progress — so `sending` goes with them:
+        # without it a reload or a second tab renders a live send as a failed one, and offers a
+        # retry that is certain to be refused.
+        "pending_files": not_uploaded(session),
+        "images_pending": images_pending(session),
+        "sending": sending(session),
+        "screenshots": screenshots,
+        "draft": session.draft,
+    }
+
+
+def view(db: Session, session: DebugSession) -> dict:
+    """The session as the API serves it — one shape wherever it is served."""
+    return serialize(session, screenshot_count(db, session)) or {}
+
+
+def screenshot_count(db: Session, session: DebugSession) -> int:
+    return (
+        db.scalar(
+            select(func.count())
+            .select_from(DebugScreenshot)
+            .where(DebugScreenshot.session_id == session.id)
+        )
+        or 0
+    )
+
+
+# --- screenshots ---------------------------------------------------------------------------
+# By content, not by the declared type: what is stored is served back to a browser as an image.
+_SIGNATURES = (
+    (b"\x89PNG\r\n\x1a\n", "image/png"),
+    (b"\xff\xd8\xff", "image/jpeg"),
+)
+
+
+def sniff(data: bytes) -> str | None:
+    """The image type `data` really is — one of IMAGE_TYPES — or None."""
+    for magic, mime in _SIGNATURES:
+        if data.startswith(magic):
+            return mime
+    if len(data) >= 12 and data[:4] == b"RIFF" and data[8:12] == b"WEBP":
+        return "image/webp"
+    return None
+
+
+def lock_user(db: Session, user_id: int) -> None:
+    """Hold this user's own lock for the rest of the transaction. A session's row lock bounds what
+    THAT session may keep; the byte cap below is the user's across all of them, and a lock on one
+    session cannot serialize a check that reads them all — two uploads to two sessions both read the
+    total before either commits, and both pass."""
+    db.execute(select(func.pg_advisory_xact_lock(user_id)))
+
+
+def stored_screenshot_bytes(db: Session, user_id: int) -> int:
+    return (
+        db.scalar(
+            select(func.coalesce(func.sum(DebugScreenshot.size), 0))
+            .join(DebugSession, DebugSession.id == DebugScreenshot.session_id)
+            .where(DebugSession.user_id == user_id)
+        )
+        or 0
+    )
+
+
+class ScreenshotRefused(Exception):
+    """A screenshot that cannot be kept; `code` is the refusal's machine word."""
+
+    def __init__(self, code: str) -> None:
+        super().__init__(code)
+        self.code = code
+
+
+def add_screenshot(db: Session, session: DebugSession, data: bytes) -> DebugScreenshot:
+    """Keep one screenshot. Locks the user and the session, and flushes; the caller commits.
+    Refused while a send is under way or once the issue was sent (it can no longer carry it), past
+    the per-session count or the user's storage, or when it is not an image we can show."""
+    if not data:
+        raise ScreenshotRefused("empty")
+    mime = sniff(data)
+    if mime is None:
+        raise ScreenshotRefused("not_an_image")
+    # Under both locks: two uploads at once cannot both pass the count (the session's), nor both
+    # pass the user's byte cap from two different sessions (the user's). Always in this order.
+    lock_user(db, session.user_id)
+    lock(db, session)
+    if sending(session):
+        raise ScreenshotRefused("sending")
+    if session.issue_url is not None:
+        raise ScreenshotRefused("already_sent")
+    if screenshot_count(db, session) >= SCREENSHOTS_MAX:
+        raise ScreenshotRefused("too_many")
+    if stored_screenshot_bytes(db, session.user_id) + len(data) > SCREENSHOT_USER_MAX_BYTES:
+        raise ScreenshotRefused("storage_full")
+    shot = DebugScreenshot(
+        session_id=session.id,
+        created_at=datetime.now(timezone.utc),
+        mime=mime,
+        size=len(data),
+        data=data,
+    )
+    db.add(shot)
+    db.flush()
+    return shot
+
+
+def remove_screenshot(db: Session, session: DebugSession, shot: DebugScreenshot) -> None:
+    """Delete one screenshot, under the lock. Refused while a send is under way or unfinished: a
+    delivery names the screenshots by their order, and removing one would renumber what it still
+    has to send. Once the issue has everything, this only drops OUR copy — the issue keeps its
+    own. The caller commits."""
+    lock(db, session)
+    if sending(session):
+        raise ScreenshotRefused("sending")
+    if unfinished(session):
+        raise ScreenshotRefused("already_sent")
+    db.delete(shot)
+    db.flush()
+
+
+def screenshots(db: Session, session: DebugSession) -> list[DebugScreenshot]:
+    """The session's screenshots, oldest first, WITHOUT their bytes loaded."""
+    return list(
+        db.scalars(
+            select(DebugScreenshot)
+            .where(DebugScreenshot.session_id == session.id)
+            .options(defer(DebugScreenshot.data))
+            .order_by(DebugScreenshot.id)
+        )
+    )
+
+
+# --- the evidence --------------------------------------------------------------------------
+# The browser's own log upload (`POST` here, every few seconds while recording) is how the evidence
+# arrives, not part of it: its server lines would fill the bundle with the recording's plumbing. The
+# browser leaves its copies out for the same reason (frontend lib/diagnostics `isOwnRequest`).
+INGEST_PATH = "/api/diagnostics/events"
+
+
+def _evidence(session: DebugSession):
+    upload_line = and_(
+        LogEvent.event == "http.request", LogEvent.attrs["path"].astext == INGEST_PATH
+    )
+    # `IS TRUE`, not the bare condition: for a row with no event or no attrs it is NULL, and
+    # `NOT NULL` would drop exactly the rows that are evidence.
+    return (LogEvent.debug_session_id == session.id, not_(upload_line.is_(True)))
+
+
+def bundle(
+    db: Session, session: DebugSession, *, drop_excluded: bool = False
+) -> tuple[list[LogEvent], int]:
+    """The session's log rows, browser and server interleaved in time, capped at BUNDLE_MAX — and
+    how many rows it has in all, so a cut bundle can say so. `drop_excluded` leaves out the rows
+    the user took out in the composer: what an export or an issue carries."""
+    where = list(_evidence(session))
+    excluded = (session.draft or {}).get("excluded") or [] if drop_excluded else []
+    if excluded:
+        where.append(LogEvent.id.not_in(excluded))
+    rows = list(
+        db.scalars(
+            select(LogEvent).where(*where).order_by(LogEvent.ts, LogEvent.id).limit(BUNDLE_MAX)
+        )
+    )
+    total = len(rows)
+    if total == BUNDLE_MAX:
+        total = db.scalar(select(func.count()).select_from(LogEvent).where(*where)) or total
+    return rows, total
+
+
+# --- housekeeping (the hourly log job) -----------------------------------------------------
+def sweep(db: Session) -> int:
+    """Stage closing every open session past its cap; returns how many."""
+    now = datetime.now(timezone.utc)
+    result = db.execute(
+        update(DebugSession)
+        .where(DebugSession.ended_at.is_(None), DebugSession.until <= now)
+        .values(ended_at=DebugSession.until, ended_reason=EndReason.EXPIRED.value)
+        .execution_options(synchronize_session=False)
+    )
+    return result.rowcount or 0
+
+
+def gc(db: Session, retention_days: int) -> int:
+    """Stage deleting sessions (and by cascade their screenshots) started before the log retention
+    window — the rows they would bundle are gone by then too. Returns how many."""
+    cutoff = datetime.now(timezone.utc) - timedelta(days=max(1, retention_days))
+    result = db.execute(delete(DebugSession).where(DebugSession.started_at < cutoff))
+    return result.rowcount or 0

backend/app/diagnostics.py

@@ -0,0 +1,65 @@
+"""A user's diagnostics level (E22 S5): what they consented to, and for how long.
+
+Three levels, named in the UI Alap / Részletes / Teljes:
+
+  * INFO  — the base, for everyone, no consent. The browser keeps its log in memory; nothing extra
+            is sent or stored.
+  * DEBUG — consented. The server logs this user's requests in detail and stores every line of
+            them, and the browser streams its own log to `POST /api/diagnostics/events`.
+  * TRACE — consented. As DEBUG, plus what the user typed into a search.
+
+A raised level always lapses after `ELEVATION_HOURS`: the expiry is computed HERE at consent time,
+never taken from the client, and an expired level reads as INFO without anything having to clear it.
+The shared demo account cannot raise its level — its log would mix every visitor's session.
+
+Core layer: imports `models` and `sysconfig`.
+"""
+
+from datetime import datetime, timedelta, timezone
+from typing import Literal, get_args
+
+from sqlalchemy.orm import Session
+
+from app import sysconfig
+from app.models import User
+
+DiagnosticsLevel = Literal["INFO", "DEBUG", "TRACE"]
+LEVELS: tuple[str, ...] = get_args(DiagnosticsLevel)
+BASE: DiagnosticsLevel = "INFO"
+ELEVATION_HOURS = 24
+# How long a debug session records at most (E16, app/debugsession.py). Here, beside the other
+# consent window, because the start dialog's consent text states it and reads it from `state`.
+SESSION_MINUTES = 60
+
+
+def effective_level(user: User, now: datetime | None = None) -> DiagnosticsLevel:
+    """The level in force for `user` right now: the consented one until it expires, else BASE."""
+    level, until = user.diagnostics_level, user.diagnostics_until
+    if level not in LEVELS or level == BASE or until is None:
+        return BASE
+    if until <= (now or datetime.now(timezone.utc)):
+        return BASE
+    return level  # type: ignore[return-value] — narrowed by the LEVELS check above
+
+
+def state(db: Session, user: User) -> dict:
+    """The `/api/me` view of it: the level in force, when a raised one lapses, and how long stored
+    log lines are kept — the consent text states that number, so it must be the instance's own."""
+    level = effective_level(user)
+    return {
+        "level": level,
+        "until": user.diagnostics_until.isoformat() if level != BASE else None,
+        "retention_days": sysconfig.get_int(db, "log_retention_days"),
+        "session_minutes": SESSION_MINUTES,
+    }
+
+
+def set_level(user: User, level: DiagnosticsLevel) -> None:
+    """Record a consent (or its withdrawal). The caller commits. Raising the level again restarts
+    the clock — each consent is for the next ELEVATION_HOURS, not an extension of an older one."""
+    if level == BASE:
+        user.diagnostics_level = None
+        user.diagnostics_until = None
+        return
+    user.diagnostics_level = level
+    user.diagnostics_until = datetime.now(timezone.utc) + timedelta(hours=ELEVATION_HOURS)

backend/app/discover/hero.py

@@ -48,6 +48,7 @@ def pick(
     shelf: list[dict] | None = None,
     exclude: set[str] | None = None,
     start: int = 0,
+    lang: str,
 ) -> dict | None:
     """`{"tier": ..., "card": ...}`, or None when nothing at all fits.
 
@@ -70,8 +71,8 @@ def pick(
     skip = exclude or set()
 
     def _resume() -> dict | None:
-        for card in rails.continue_watching(db, user) if resume is None else resume:
+        for card in rails.continue_watching(db, user, lang=lang) if resume is None else resume:
-            playable = rails.playable(db, user.id, card)
+            playable = rails.playable(db, user.id, card, lang=lang)
             # Excluded on the card the USER saw (the show), not only on what it resolves to, so a
             # rerolled-past show does not come straight back as its own next episode.
             if cards.key(card) in skip or (playable is not None and cards.key(playable) in skip):
@@ -81,7 +82,11 @@ def pick(
         return None
 
     def _comfort() -> dict | None:
-        for card in pins.playables(pins.shelf(db, user) if shelf is None else shelf):
+        if shelf is None:
+            own = pins.shelf(db, user, include_plex=rails.plex_available(db), lang=lang)
+        else:
+            own = shelf
+        for card in pins.playables(own):
             if cards.key(card) not in skip and fits(card["remaining_seconds"], budget):
                 return card
         return None
@@ -89,7 +94,7 @@ def pick(
     def _surprise() -> dict | None:
         # A wider draw than the one card we need: the shuffle is seeded, so without the headroom an
         # excluded pick would leave the tier empty rather than move past it.
-        for card in rails.surprise(db, user, budget, seed, limit=SURPRISE_DRAW):
+        for card in rails.surprise(db, user, budget, seed, limit=SURPRISE_DRAW, lang=lang):
             if cards.key(card) not in skip:
                 return card
         return None

backend/app/discover/pins.py

@@ -8,14 +8,16 @@ second one.
 
 Every shelf card carries a resolved `next`: the thing that actually plays when you click the tile —
 a show's in-progress-or-next episode, a channel's newest unwatched video, a playlist's first
-unwatched item. It costs a query per non-leaf pin, which is affordable precisely because this shelf is
+unwatched item. That buys the property the whole surface exists for: one click, no intermediate page,
-short and hand-curated, and it buys the property the whole surface exists for: one click, no
+no second decision. Channels and playlists are resolved for the whole shelf at once (`_Batch`), so the
-intermediate page, no second decision.
+cost does not grow a query or two per pin.
 
 Domain module, so nothing here raises HTTP. `resolve_ref` returns None for "no such thing, or not
 yours" and the route turns that into its 404 — core never imports routes.
 """
 
+from dataclasses import dataclass, field
+
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
@@ -62,6 +64,8 @@ def resolve_ref(db: Session, user: User, kind: str, ref: str) -> tuple[str, obje
         return (col, sh.id) if sh is not None else None
     try:
         pid = int(ref)
+    # silent-ok: validating a reference the client supplied. "Not an integer" is answered as "no
+    # such pin target", which is the same answer this function gives for one that does not exist.
     except ValueError:
         return None
     pl = db.get(Playlist, pid)
@@ -71,6 +75,44 @@ def resolve_ref(db: Session, user: User, kind: str, ref: str) -> tuple[str, obje
     return (col, pl.id) if pl is not None and pl.user_id == user.id else None
 
 
+def reorder(db: Session, user: User, ids: list[int]) -> None:
+    """Put the listed pins in the listed order, WITHIN the slots they already hold.
+
+    The client sends the order it can SEE, which is not always the whole shelf: with Plex off the Home
+    rail hides the Plex pins, and a stale tab may not know a pin added elsewhere. Numbering only the
+    listed ids 0..n-1 collided with every pin left out — two tiles on one position, and the hidden
+    ones jumped about when they came back. So the pins left out keep their slots, and the listed ones
+    are dealt into the slots the listed set occupies. Ids that are not this user's, or repeated, are
+    ignored rather than failing a stale replay wholesale. Flushes; the route commits."""
+    # Locked: the rows it leaves alone are part of the answer too. Two overlapping reorders each
+    # skipped the rows that already matched in their own snapshot and interleaved into two pins on
+    # one slot; the second now waits and deals from the first one's result. Locked in id order and
+    # sorted HERE: with `ORDER BY position … FOR UPDATE` Postgres sorts before it locks, so the one
+    # that waited got the fresh rows in its stale order and dealt the pins it left out back to their
+    # old slots — and one global lock order means two reorders cannot deadlock.
+    locked = db.execute(
+        select(ComfortPin)
+        .where(ComfortPin.user_id == user.id)
+        .order_by(ComfortPin.id)
+        .with_for_update()
+        .execution_options(populate_existing=True)
+    ).scalars()
+    current = sorted(locked, key=lambda p: (p.position, p.id))
+    owned = {p.id for p in current}
+    listed: list[int] = []
+    for i in ids:
+        if i in owned and i not in listed:
+            listed.append(i)
+    chosen = set(listed)
+    queue = iter(listed)
+    order = [next(queue) if p.id in chosen else p.id for p in current]
+    by_id = {p.id: p for p in current}
+    for position, pin_id in enumerate(order):
+        if by_id[pin_id].position != position:
+            by_id[pin_id].position = position
+    db.flush()
+
+
 def list_pins(db: Session, user: User) -> list[ComfortPin]:
     return list(
         db.execute(
@@ -101,7 +143,34 @@ def next_position(db: Session, user: User) -> int:
     )
 
 
-def _shelf_card(db: Session, user: User, pin: ComfortPin) -> dict | None:
+@dataclass
+class _Batch:
+    """What the shelf resolves for every pinned channel and playlist in one go."""
+
+    channels: dict[str, Channel] = field(default_factory=dict)
+    channel_summary: dict[str, tuple[int, dict | None]] = field(default_factory=dict)
+    playlists: dict[int, Playlist] = field(default_factory=dict)
+    playlist_summary: dict[int, tuple[int, dict | None]] = field(default_factory=dict)
+
+    @classmethod
+    def load(cls, db: Session, user: User, pins: list[ComfortPin]) -> "_Batch":
+        ch_ids = [p.channel_id for p in pins if p.channel_id is not None]
+        pl_ids = [p.playlist_id for p in pins if p.playlist_id is not None]
+        b = cls()
+        if ch_ids:
+            b.channels = {
+                c.id: c for c in db.scalars(select(Channel).where(Channel.id.in_(ch_ids)))
+            }
+            b.channel_summary = queries.channels_summary(db, user, ch_ids)
+        if pl_ids:
+            b.playlists = {
+                p.id: p for p in db.scalars(select(Playlist).where(Playlist.id.in_(pl_ids)))
+            }
+            b.playlist_summary = queries.playlists_summary(db, user, pl_ids)
+        return b
+
+
+def _shelf_card(db: Session, user: User, pin: ComfortPin, batch: _Batch, lang: str) -> dict | None:
     """One pin as a unified card with its resolved `next`. None when the pinned row has gone (the FK
     cascade removes the pin itself, so this is only the window between a delete and a refetch)."""
     card: dict | None = None
@@ -115,11 +184,11 @@ def _shelf_card(db: Session, user: User, pin: ComfortPin) -> dict | None:
         nxt = dict(card)  # a video plays itself
 
     elif pin.channel_id is not None:
-        ch = db.get(Channel, pin.channel_id)
+        ch = batch.channels.get(pin.channel_id)
         if ch is None:
             return None
-        card = cards.from_channel(ch, queries.channel_unwatched_count(db, user, ch.id))
+        count, row = batch.channel_summary.get(ch.id, (0, None))
-        row = queries.channel_next_video(db, user, ch.id)
+        card = cards.from_channel(ch, count)
         nxt = cards.from_video(row) if row else None
 
     elif pin.plex_item_id is not None:
@@ -127,7 +196,7 @@ def _shelf_card(db: Session, user: User, pin: ComfortPin) -> dict | None:
         if it is None:
             return None
         st = rollup.item_state_map(db, user.id, [it.id]).get(it.id)
-        card = cards.from_plex(leaf_card(it, st))
+        card = cards.from_plex(leaf_card(it, st, lang=lang))
         nxt = dict(card)  # a movie or a single episode plays itself
 
     elif pin.plex_show_id is not None:
@@ -135,15 +204,15 @@ def _shelf_card(db: Session, user: User, pin: ComfortPin) -> dict | None:
         if sh is None:
             return None
         status = rollup.state_map(db, user.id, [sh.id]).get(sh.id, "new")
-        card = cards.from_plex(show_card(sh, status))
+        card = cards.from_plex(show_card(sh, status, lang=lang))
-        ep = queries.next_episode(db, user.id, sh.id)
+        ep = queries.next_episode(db, user.id, sh.id, lang=lang)
         nxt = cards.from_plex(ep) if ep else None
 
     elif pin.playlist_id is not None:
-        pl = db.get(Playlist, pin.playlist_id)
+        pl = batch.playlists.get(pin.playlist_id)
         if pl is None:
             return None
-        count, row = queries.playlist_summary(db, user, pl.id)
+        count, row = batch.playlist_summary.get(pl.id, (0, None))
         card = cards.from_playlist(pl, count, row.get("thumbnail_url") if row else None)
         nxt = cards.from_video(row) if row else None
 
@@ -157,11 +226,24 @@ def _shelf_card(db: Session, user: User, pin: ComfortPin) -> dict | None:
     return card
 
 
-def shelf(db: Session, user: User) -> list[dict]:
+_PLEX_COLUMNS = ("plex_item_id", "plex_show_id")
-    """The whole shelf, in the user's own order."""
+
+
+def shelf(db: Session, user: User, *, include_plex: bool = True, lang: str) -> list[dict]:
+    """The whole shelf, in the user's own order.
+
+    `include_plex=False` is the Home surface with the deployment's Plex toggle off (the owner's call,
+    2026-09-19): the rails beside it go YouTube-only, and a Plex tile there would offer a Play that
+    lands on a gated module. The pins themselves are kept — the pin menus still read the full shelf
+    (`GET /pins`), and switching Plex back on brings the tiles back in their old slots
+    (`reorder` keeps a hidden pin's place)."""
     out = []
-    for pin in list_pins(db, user):
+    pins = list_pins(db, user)
-        card = _shelf_card(db, user, pin)
+    if not include_plex:
+        pins = [p for p in pins if all(getattr(p, c) is None for c in _PLEX_COLUMNS)]
+    batch = _Batch.load(db, user, pins)
+    for pin in pins:
+        card = _shelf_card(db, user, pin, batch, lang)
         if card is not None:
             out.append(card)
     return out
@@ -169,5 +251,17 @@ def shelf(db: Session, user: User) -> list[dict]:
 
 def playables(shelf_cards: list[dict]) -> list[dict]:
     """The `next` of every shelf entry that has one — what the hero may draw from. A pinned channel
-    with nothing unwatched left, or an empty playlist, simply is not an answer right now."""
+    with nothing unwatched left, or an empty playlist, simply is not an answer right now.
-    return [c["next"] for c in shelf_cards if c.get("next")]
+
+    Each carries `pinned_from`: the pin it was resolved from. A show's next EPISODE is not itself
+    pinned, so without it the hero's menu looked the episode up, missed, and offered "Pin" under a
+    kicker that says "From your shelf" — and taking it inserted a second pin. With it, the hero's
+    pin row acts on the shelf entry the card came from (the owner's call, 2026-09-19: "Unpin")."""
+    return [
+        {
+            **c["next"],
+            "pinned_from": {k: c[k] for k in ("source", "kind", "id", "title")},
+        }
+        for c in shelf_cards
+        if c.get("next")
+    ]

backend/app/discover/queries.py

@@ -40,11 +40,12 @@ from app.models import (
     VideoState,
 )
 from app.plex import rollup
-from app.plex.cards import leaf_card, movie_card, show_card
+from app.plex.cards import display_title, leaf_card, movie_card, show_card
 from app.plex.query import apply_show_state_filter, content_libs
 
-# How many rows a single rail may carry. Matches `routes/plex/discover.py`'s RAIL_LIMIT on purpose —
+# How many rows a single rail may carry — Home's rails and the Plex module's Discover rails alike
-# the two surfaces should feel like the same app, and a rail is a rail.
+# (`routes/plex/discover.py` imports it): the two surfaces should feel like the same app, and a rail
+# is a rail. One definition, because a comment was all that held two copies in step.
 RAIL_LIMIT = 20
 
 # The feed parameters a Home rail runs with. Spelled out rather than borrowed from the route's
@@ -129,18 +130,6 @@ def yt_surprise(db: Session, user: User, budget: str, seed: int, limit: int) ->
     return [r for r in out if not r["position_seconds"]][:limit]
 
 
-def channel_next_video(db: Session, user: User, channel_id: str) -> dict | None:
-    """The newest unwatched video of a pinned channel — what "put on that channel" resolves to."""
-    # `library_source="all"`: a pinned channel means the whole channel, including videos that only
-    # ever arrived through a search. The feed's "organic" default is a feed-scoping choice, not a
-    # statement about what belongs to a channel.
-    params = dict(_FEED_BASE, channel_ids=[channel_id], scope="all", library_source="all")
-    query, rank_expr = filtered_query(db, user, **params)
-    keys = sort_keys("newest", 0, rank_expr)
-    row = db.execute(query.order_by(*order_by(keys)).limit(1)).first()
-    return serialize(row) if row is not None else None
-
-
 def yt_video_row(db: Session, user: User, video_id: str) -> dict | None:
     """One video as a feed row, with this user's watch state.
 
@@ -158,31 +147,55 @@ def yt_video_row(db: Session, user: User, video_id: str) -> dict | None:
     return serialize(row) if row is not None else None
 
 
-def channel_unwatched_count(db: Session, user: User, channel_id: str) -> int:
+def channels_summary(
-    """How many unwatched videos a pinned channel is holding — the tile's only honest promise, since a
+    db: Session, user: User, channel_ids: list[str]
-    channel has no runtime of its own."""
+) -> dict[str, tuple[int, dict | None]]:
-    params = dict(_FEED_BASE, channel_ids=[channel_id], scope="all", library_source="all")
+    """For each pinned channel: (unwatched count, its newest unwatched video) — the tile's only honest
-    query, _ = filtered_query(db, user, **params)
+    promise, since a channel has no runtime of its own, and what "put on that channel" resolves to.
-    return int(db.execute(select(func.count()).select_from(query.subquery())).scalar() or 0)
+
+    ONE query for the whole shelf, not two per channel: the count and the first row per channel are
+    window functions over the same filtered set. `library_source="all"`: a pinned channel means the
+    whole channel, including videos that only ever arrived through a search — the feed's "organic"
+    default is a feed-scoping choice, not a statement about what belongs to a channel."""
+    if not channel_ids:
+        return {}
+    params = dict(_FEED_BASE, channel_ids=list(channel_ids), scope="all", library_source="all")
+    query, rank_expr = filtered_query(db, user, **params)
+    newest = order_by(sort_keys("newest", 0, rank_expr))
+    per_channel = {"partition_by": Video.channel_id}
+    sub = query.add_columns(
+        func.row_number().over(**per_channel, order_by=newest).label("_rn"),
+        func.count().over(**per_channel).label("_total"),
+    ).subquery()
+    rows = db.execute(select(sub).where(sub.c._rn == 1)).all()
+    return {r.channel_id: (int(r._total), serialize(r)) for r in rows}
 
 
-def playlist_summary(db: Session, user: User, playlist_id: int) -> tuple[int, dict | None]:
+def playlists_summary(
-    """(item count, the row to play next) for a pinned playlist — its first not-yet-watched item, or
+    db: Session, user: User, playlist_ids: list[int]
-    its first item once the whole list is watched. The same row doubles as the tile's artwork, so a
+) -> dict[int, tuple[int, dict | None]]:
-    shelf of playlists is scannable instead of a row of blank boxes."""
+    """For each pinned playlist: (item count, the row to play next) — its first not-yet-watched item,
+    or its first item once the whole list is watched. The same row doubles as the tile's artwork, so
+    a shelf of playlists is scannable instead of a row of blank boxes. One query for the whole shelf."""
+    if not playlist_ids:
+        return {}
     st = aliased(VideoState)
     rows = db.execute(
-        select(*feed_columns(user, st), PlaylistItem.position)
+        select(*feed_columns(user, st), PlaylistItem.playlist_id.label("_playlist"))
         .join(PlaylistItem, PlaylistItem.video_id == Video.id)
         .join(Channel, Channel.id == Video.channel_id)
         .outerjoin(st, and_(st.video_id == Video.id, st.user_id == user.id))
-        .where(PlaylistItem.playlist_id == playlist_id)
+        .where(PlaylistItem.playlist_id.in_(playlist_ids))
-        .order_by(PlaylistItem.position.asc())
+        .order_by(PlaylistItem.playlist_id, PlaylistItem.position.asc())
     ).all()
-    if not rows:
+    grouped: dict[int, list] = {}
-        return 0, None
+    for r in rows:
-    nxt = next((r for r in rows if (r.status or "new") not in ("watched", "hidden")), rows[0])
+        grouped.setdefault(r._playlist, []).append(r)
-    return len(rows), serialize(nxt)
+    out: dict[int, tuple[int, dict | None]] = {}
+    for pid, items in grouped.items():
+        nxt = next((r for r in items if (r.status or "new") not in ("watched", "hidden")), items[0])
+        out[pid] = (len(items), serialize(nxt))
+    return out
 
 
 # ── Plex ──────────────────────────────────────────────────────────────────────────────────────────
@@ -208,7 +221,9 @@ def _unseen_movie_conds(uid: int):
     return ~seen
 
 
-def plex_in_progress(db: Session, uid: int, limit: int = RAIL_LIMIT) -> list[tuple[object, dict]]:
+def plex_in_progress(
+    db: Session, uid: int, limit: int = RAIL_LIMIT, *, lang: str
+) -> list[tuple[object, dict]]:
     """In-progress movies + partially-watched shows as (progressed_at, card).
 
     Mirrors `routes/plex/discover.py::_continue_watching` — same `PlexState` predicate for movies (a
@@ -236,7 +251,7 @@ def plex_in_progress(db: Session, uid: int, limit: int = RAIL_LIMIT) -> list[tup
             .limit(limit)
         )
         for it, st in db.execute(mq).all():
-            out.append((st.progress_updated_at, movie_card(it, st)))
+            out.append((st.progress_updated_at, movie_card(it, st, lang=lang)))
 
     if show_libs:
         prog = rollup.agg_subquery(db, uid, include_resumed=True).subquery()
@@ -250,12 +265,14 @@ def plex_in_progress(db: Session, uid: int, limit: int = RAIL_LIMIT) -> list[tup
         for sh, resumed_at in db.execute(
             sq.order_by(prog.c.resumed_at.desc().nullslast()).limit(limit)
         ).all():
-            out.append((resumed_at, show_card(sh, "in_progress")))
+            out.append((resumed_at, show_card(sh, "in_progress", lang=lang)))
 
     return out
 
 
-def plex_surprise(db: Session, uid: int, budget: str, seed: int, limit: int) -> list[dict]:
+def plex_surprise(
+    db: Session, uid: int, budget: str, seed: int, limit: int, *, lang: str
+) -> list[dict]:
     """Unseen playable movies inside the budget, in seeded-shuffle order.
 
     Movies only. A show is not a thing you can start playing, and resolving every candidate show to
@@ -287,14 +304,14 @@ def plex_surprise(db: Session, uid: int, budget: str, seed: int, limit: int) ->
         q = q.where(PlexItem.duration_s <= hi)
     items = db.execute(q).scalars().all()
     states = rollup.item_state_map(db, uid, [it.id for it in items])
-    return [movie_card(it, states.get(it.id)) for it in items]
+    return [movie_card(it, states.get(it.id), lang=lang) for it in items]
 
 
 # Watch states that take an episode out of the running for "what would play next".
 _DONE = ("watched", "hidden")
 
 
-def next_episode(db: Session, uid: int, show_id: int) -> dict | None:
+def next_episode(db: Session, uid: int, show_id: int, *, lang: str) -> dict | None:
     """What "put that show on" means: the episode in progress, else the first unwatched one, else the
     first — over the same `(season, episode)` broadcast ordering `detail.py::_episode_nav` uses.
 
@@ -344,5 +361,9 @@ def next_episode(db: Session, uid: int, show_id: int) -> dict | None:
     ep, st = pick
     show = db.get(PlexShow, show_id)
     return leaf_card(
-        ep, st, show_title=show.title if show else None, show_rk=show.rating_key if show else None
+        ep,
+        st,
+        show_title=display_title(show, lang) if show else None,
+        show_rk=show.rating_key if show else None,
+        lang=lang,
     )

backend/app/discover/rails.py

@@ -22,7 +22,7 @@ def plex_available(db: Session) -> bool:
     return sysconfig.get_bool(db, "plex_enabled")
 
 
-def playable(db: Session, uid: int, card: dict) -> dict | None:
+def playable(db: Session, uid: int, card: dict, *, lang: str) -> dict | None:
     """The thing that would actually START from this card — itself for anything with a runtime, and
     for a Plex SHOW the episode that is next (in progress, else the first unwatched, else the first).
 
@@ -40,11 +40,11 @@ def playable(db: Session, uid: int, card: dict) -> dict | None:
     sh = by_rating_key(db, PlexShow, card["id"])
     if sh is None:
         return None
-    ep = queries.next_episode(db, uid, sh.id)
+    ep = queries.next_episode(db, uid, sh.id, lang=lang)
     return cards.from_plex(ep) if ep else None
 
 
-def continue_watching(db: Session, user: User, limit: int = RAIL_LIMIT) -> list[dict]:
+def continue_watching(db: Session, user: User, limit: int = RAIL_LIMIT, *, lang: str) -> list[dict]:
     """One resume rail across both sources, most recently progressed first.
 
     A Plex SHOW enters as a show CARD — that is what the Plex rail shows, and it is the artwork the
@@ -57,17 +57,22 @@ def continue_watching(db: Session, user: User, limit: int = RAIL_LIMIT) -> list[
         (ts, cards.from_video(r)) for ts, r in queries.yt_in_progress(db, user, limit)
     ]
     if plex_available(db):
-        rows += [(ts, cards.from_plex(c)) for ts, c in queries.plex_in_progress(db, user.id, limit)]
+        rows += [
+            (ts, cards.from_plex(c))
+            for ts, c in queries.plex_in_progress(db, user.id, limit, lang=lang)
+        ]
     rows.sort(key=lambda r: r[0] or NO_TS, reverse=True)
     out = [card for _, card in rows[:limit]]
     # Resolve AFTER the slice: only the rows that survive it are worth a query.
     for card in out:
         if card["kind"] == "show":
-            card["next"] = playable(db, user.id, card)
+            card["next"] = playable(db, user.id, card, lang=lang)
     return out
 
 
-def surprise(db: Session, user: User, budget: str, seed: int, limit: int = 3) -> list[dict]:
+def surprise(
+    db: Session, user: User, budget: str, seed: int, limit: int = 3, *, lang: str
+) -> list[dict]:
     """The hero's draw pool: unseen things that fit the budget, from both sources under ONE seed.
 
     Interleaved rather than concatenated — the same alternation `routes/plex/discover.py` uses for its
@@ -83,7 +88,9 @@ def surprise(db: Session, user: User, budget: str, seed: int, limit: int = 3) ->
         (cards.from_video, queries.yt_surprise(db, user, budget, seed, limit)),
         (
             cards.from_plex,
-            queries.plex_surprise(db, user.id, budget, seed, limit) if plex_available(db) else [],
+            queries.plex_surprise(db, user.id, budget, seed, limit, lang=lang)
+            if plex_available(db)
+            else [],
         ),
     ]
     if seed % 2:

backend/app/downloads/adapters.py

@@ -18,6 +18,7 @@ import yt_dlp
 from yt_dlp.networking.impersonate import ImpersonateTarget
 
 from app.config import settings
+from app.downloads.stated import StatedFailure
 
 # Facebook's "Cannot parse data": the page yt-dlp got was a wall, not the content. Both the Chrome retry
 # below and the failure catalog (errors.py) decide on it, so the wording exists once.
@@ -93,12 +94,46 @@ class YouTubeAdapter(SourceAdapter):
         "Requested format is not available",
         "Only images are available",
         "Sign in to confirm",
+        # Measured 2026-09-16: a gated video whose EXTRACTION succeeds and whose media fetch is then
+        # refused by the CDN ("unable to download video data: HTTP Error 403"). The wall moves
+        # between the player response and the stream, so this is the same per-client flakiness one
+        # step later — and it used to end the job on attempt 1, because the text classifies as
+        # `network` and nothing here matched it. The format URLs belong to the client that minted
+        # them, so the next client set is a real second chance.
+        "unable to download video data",
     )
+    # Measured in the worker against a REAL account cookie (2026-09-16, yt-dlp 2026.07.04): with
+    # `player_client=android_vr` yt-dlp logs `Skipping client "android_vr" since it does not support
+    # cookies`, drops it, and the request that is left with no client fails with "Requested format is
+    # not available" — which is already a retriable marker above, so the cost is a wasted attempt
+    # rather than a hard failure. Our primary set is android_vr ALONE, so without this filter every
+    # signed-in download would burn its first attempt that way before reaching the POT-backed web
+    # fallback the cookie exists for. The list is yt-dlp's own INNERTUBE_CLIENTS[...]
+    # ["SUPPORTS_COOKIES"], pinned by test_adapters against the installed yt-dlp, so a version bump
+    # that moves the table turns the gate red instead of silently un-authing every YouTube download.
+    _COOKIE_CAPABLE = frozenset(
+        {
+            "mweb",
+            "tv",
+            "tv_downgraded",
+            "web",
+            "web_creator",
+            "web_embedded",
+            "web_music",
+            "web_safari",
+        }
+    )
+    # Where a signed-in download lands when the admin's knobs hold no cookie-capable client at all.
+    # Both clients in ONE set: yt-dlp queries each and merges their formats, which is the cover for
+    # the per-client flakiness that the two-tier ladder gives a signed-out download.
+    _AUTHED_DEFAULT = ("web_safari", "web")
 
     def client_sets(self, authed: bool = False) -> list:
-        return [
+        sets = [c for c in (settings.player_client_list, settings.player_client_fallback_list) if c]
-            c for c in (settings.player_client_list, settings.player_client_fallback_list) if c
+        if authed:
-        ] or [None]
+            kept = [[c for c in s if c in self._COOKIE_CAPABLE] for s in sets]
+            return [s for s in kept if s] or [list(self._AUTHED_DEFAULT)]
+        return sets or [None]
 
     def ydl_opts_extra(self, clients) -> dict:
         # == --extractor-args "youtube:player_client=…;youtubepot-bgutilhttp:base_url=…" (list values).
@@ -144,11 +179,11 @@ class FacebookAdapter(SourceAdapter):
         # Measured (E11.5 review): a refused or expired cookie gets the login page (~450 KB, a login
         # form, no post id at all), which is not a post without a video — the user's cookie is the cause.
         if _quoted(post_id) not in page:
-            raise RuntimeError(POST_NOT_SHOWN_MESSAGE)
+            raise StatedFailure(POST_NOT_SHOWN_MESSAGE)
         video_id = facebook_post_video_id(page, post_id)
         if video_id is None:
             # Never fall back to the post link: that is exactly where the wrong video comes from.
-            raise RuntimeError(POST_WITHOUT_VIDEO_MESSAGE)
+            raise StatedFailure(POST_WITHOUT_VIDEO_MESSAGE)
         return f"https://www.facebook.com/watch/?v={video_id}"
 
     def client_sets(self, authed: bool = False) -> list:

backend/app/downloads/backfill.py

@@ -11,11 +11,12 @@ Run once per environment:
     docker exec siftlode-worker-1 python -c "from app.downloads import backfill; print(backfill.run())"
 """
 
-import logging
+from urllib.parse import urlsplit
 
 import yt_dlp
 from sqlalchemy import select
 
+from app import logs
 from app.config import settings
 from app.db import SessionLocal
 from app.downloads import adapters
@@ -23,12 +24,17 @@ from app.downloads import edit as editmod
 from app.downloads import formats, service
 from app.models import MediaAsset
 
-log = logging.getLogger("siftlode.backfill")
+log = logs.get_logger(__name__)
 
 
-def _extract_meta(url: str, adapter: adapters.SourceAdapter) -> dict | None:
+def _extract_meta(url: str, adapter: adapters.SourceAdapter, asset_id: int) -> dict | None:
     """Metadata-only yt-dlp extraction, using the source adapter's client sets + options (YouTube
     needs the POT/player-client strategy even just to extract)."""
+    with formats.hiding(url):
+        return _extract_meta_scoped(url, adapter, asset_id)
+
+
+def _extract_meta_scoped(url: str, adapter: adapters.SourceAdapter, asset_id: int) -> dict | None:
     for clients in adapter.client_sets():
         opts = formats.build_ydl_opts(
             {}, "%(id)s.%(ext)s", lambda d: None, extra=adapter.ydl_opts_extra(clients)
@@ -40,7 +46,19 @@ def _extract_meta(url: str, adapter: adapters.SourceAdapter) -> dict | None:
             if info:
                 return info
         except Exception as exc:  # noqa: BLE001 — best-effort backfill, keep going
-            log.info("meta extract failed for %s (%s): %s", url, clients, str(exc)[:100])
+            # The asset id, not the URL. A source URL is a row of this user's download
+            # history, and `redact_text` only takes credential params out of it — the rest
+            # of it goes in whole. The id names the same asset to anyone holding the
+            # database and nothing at all to anyone who is not; the host is kept because it
+            # is what decides which client set was tried. yt-dlp's message goes through
+            # `log_safe` for the same reason: it opens with the URL or the video id.
+            log.info(
+                "meta extract failed for asset %s on %s (%s): %s",
+                asset_id,
+                urlsplit(url).hostname or "?",  # log-pii-ok: the host, never the path or query
+                clients,
+                formats.log_safe(exc)[:100],
+            )
     return None
 
 
@@ -60,12 +78,14 @@ def run() -> dict:
             adapter = adapters.resolve_adapter(a.source_kind, a.source_ref)
             if a.uploader_url is None and adapter.wants_metadata_backfill():
                 url = a.source_webpage_url or service.source_url(a.source_kind, a.source_ref)
-                info = _extract_meta(url, adapter)
+                info = _extract_meta(url, adapter, a.id)
                 uu = adapter.uploader_url(info) if info else None
                 if uu:
                     a.uploader_url = uu
                     filled_url += 1
-                    log.info("asset %s uploader_url <- %s", a.id, uu)
+                    log.info(
+                        "asset %s uploader_url filled", a.id
+                    )  # the URL names what they downloaded
             if a.poster_path is None:
                 # Record a self-hosted poster for every download (the reliable og:image source):
                 # ensure_poster returns the existing <base>.jpg thumbnail sidecar, or cuts a frame.
@@ -73,7 +93,7 @@ def run() -> dict:
                 if rel:
                     a.poster_path = rel
                     filled_poster += 1
-                    log.info("asset %s poster <- %s", a.id, rel)
+                    log.info("asset %s poster filled", a.id)  # the path carries the title
             db.commit()
     result = {
         "assets": len(assets),

backend/app/downloads/credentials.py

@@ -23,19 +23,19 @@ Security controls enforced here:
     logged, or written to a job's error/snapshot.
 """
 
-import logging
 import os
 from pathlib import Path
 
 from sqlalchemy import select
 from sqlalchemy.orm import Session
 
+from app import logs
 from app.downloads import formats
 from app.downloads.adapters import _host_is
 from app.models import DownloadCredential, MediaAsset
 from app.security import decrypt_optional, encrypt
 
-log = logging.getLogger("siftlode.downloads.credentials")
+log = logs.get_logger(__name__)
 
 
 class HostSpec:
@@ -55,6 +55,11 @@ class HostSpec:
 # all derive from this, so there is nothing else to wire up.
 SUPPORTED_HOSTS: tuple[HostSpec, ...] = (
     HostSpec("facebook", "Facebook", ("facebook.com", "fb.watch")),
+    # YouTube gates INDIVIDUAL videos behind "Sign in to confirm you're not a bot" (measured on prod
+    # job 273: every player client fails on that one video while its neighbours download fine, and
+    # the newest yt-dlp fails the same way — a signed-in session is the only way through). Without a
+    # row here the `login_required` copy sent the user to a dropdown that could not help them.
+    HostSpec("youtube", "YouTube", ("youtube.com", "youtu.be", "youtube-nocookie.com")),
     HostSpec("instagram", "Instagram", ("instagram.com",)),
     HostSpec("twitter", "X (Twitter)", ("twitter.com", "x.com")),
     HostSpec("tiktok", "TikTok", ("tiktok.com",)),

backend/app/downloads/edit.py

@@ -23,7 +23,6 @@ for a player hover-preview later.
 
 import hashlib
 import json
-import logging
 import math
 import os
 import queue
@@ -33,9 +32,11 @@ import threading
 import time
 from pathlib import Path
 
+from app import logs
 from app.config import settings
+from app.downloads.stated import StatedFailure
 
-log = logging.getLogger("siftlode.edit")
+log = logs.get_logger(__name__)
 
 
 def encode_thread_cap() -> int:
@@ -71,7 +72,13 @@ def normalize_edit_spec(spec: dict | None) -> dict:
     if crop:
         try:
             c = {k: int(round(float(crop[k]))) for k in ("x", "y", "w", "h")}
-        except (KeyError, TypeError, ValueError):
+        except (KeyError, TypeError, ValueError) as exc:
+            # The clip still renders — uncropped. The user asked for a crop, got a whole frame,
+            # and nothing anywhere called it a failure; this line is the only trace there is.
+            log.warning(
+                "the crop in an edit spec could not be read and was dropped",
+                extra=logs.event("edit.spec.crop_dropped", reason=type(exc).__name__),
+            )
             c = None
         if c and c["w"] > 0 and c["h"] > 0 and c["x"] >= 0 and c["y"] >= 0:
             out["crop"] = c
@@ -80,6 +87,7 @@ def normalize_edit_spec(spec: dict | None) -> dict:
     segs = spec.get("segments")
     if isinstance(segs, list) and segs:
         norm: list[dict] = []
+        unreadable = 0
         for s in segs:
             try:
                 st = max(0.0, float((s or {}).get("start_s") or 0.0))
@@ -88,9 +96,18 @@ def normalize_edit_spec(spec: dict | None) -> dict:
                     continue
                 en = float(end_raw)
             except (TypeError, ValueError):
+                # silent-ok: counted, not logged here — one cut-list can drop several, and the
+                # join that comes out is simply shorter than the one the user built. Reported
+                # once after the loop.
+                unreadable += 1
                 continue
             if en > st:
                 norm.append({"start_s": round(st, 3), "end_s": round(en, 3)})
+        if unreadable:
+            log.warning(
+                "some segments of a cut list could not be read and were dropped",
+                extra=logs.event("edit.spec.segments_dropped", dropped=unreadable, kept=len(norm)),
+            )
         norm.sort(key=lambda x: x["start_s"])
         # SEC-13: an accurate join builds one ffmpeg `filter_complex` with two filters per segment,
         # so an unbounded list becomes an unbounded command line and an ffmpeg that never finishes.
@@ -114,7 +131,13 @@ def normalize_edit_spec(spec: dict | None) -> dict:
             start = max(0.0, float(trim.get("start_s") or 0.0))
             end_raw = trim.get("end_s")
             end = float(end_raw) if end_raw is not None else None
-        except (TypeError, ValueError):
+        except (TypeError, ValueError) as exc:
+            # Either a 400 the user cannot act on, or — beside a crop — a full-length clip where
+            # they asked for a cut. Both are worth one line saying the spec was the problem.
+            log.warning(
+                "the trim in an edit spec could not be read and was dropped",
+                extra=logs.event("edit.spec.trim_dropped", reason=type(exc).__name__),
+            )
             start, end = 0.0, None
         t: dict = {"start_s": round(start, 3)}
         if end is not None and end > start:
@@ -306,11 +329,14 @@ def _parse_out_time(line: str) -> float | None:
     if line.startswith("out_time_us="):
         try:
             return int(line.split("=", 1)[1]) / 1_000_000
+        # silent-ok: one line of ffmpeg's `-progress` stream, several per second. None means "this
+        # line carried no usable time", which the caller already treats as "no progress yet".
         except ValueError:
             return None
     if line.startswith("out_time_ms="):  # (some builds mislabel this as microseconds)
         try:
             return int(line.split("=", 1)[1]) / 1_000_000
+        # silent-ok: as above, the other spelling of the same line.
         except ValueError:
             return None
     return None
@@ -329,7 +355,9 @@ def _pump(stream, sink: "queue.Queue[str | None]") -> None:
     try:
         for line in stream:
             sink.put(line)
-    except (ValueError, OSError):  # pipe closed under us by the kill path
+    # silent-ok: the pipe was closed under us by our OWN kill path — the expected end of a
+    # cancelled or timed-out run, and the code that did the killing is the code that reports it.
+    except (ValueError, OSError):
         pass
     finally:
         sink.put(None)
@@ -347,6 +375,8 @@ def _drain_tail(stream, sink: list[str]) -> None:
             sink.append(line)
             if len(sink) > 200:
                 del sink[:100]
+    # silent-ok: same closed pipe as `_pump` above, and the tail collected so far is still handed
+    # to whoever reports the failure.
     except (ValueError, OSError):
         pass
 
@@ -362,13 +392,17 @@ def escalate_kill(proc: subprocess.Popen, terminate, kill, grace_s: float, label
     try:
         proc.wait(timeout=grace_s)
         return
+    # silent-ok: the escalation below IS the handling of this timeout, and surviving THAT is what
+    # the line under the second wait reports.
     except subprocess.TimeoutExpired:
         pass
     kill()
     try:
         proc.wait(timeout=grace_s)
     except subprocess.TimeoutExpired:
-        log.error("%s pid %s survived SIGKILL", label, proc.pid)
+        # TRY400: the exception IS the message here. A TimeoutExpired's traceback is this
+        # `wait` call and nothing else — it would add three frames and no fact the line lacks.
+        log.error("%s pid %s survived SIGKILL", label, proc.pid)  # noqa: TRY400
 
 
 def _stop(proc: subprocess.Popen) -> None:
@@ -404,6 +438,8 @@ def run_ffmpeg(
         while True:
             try:
                 line = lines.get(timeout=1.0)
+            # silent-ok: not a failure — the one-second timeout is how this loop gets a turn to
+            # poll for cancellation while ffmpeg is quiet.
             except queue.Empty:
                 line = ""
             if line is None:  # stdout hit EOF — ffmpeg is done (or dying)
@@ -416,11 +452,11 @@ def run_ffmpeg(
             if should_cancel():
                 raise EditAborted
             if time.monotonic() - last_output > stall_timeout_s:
-                raise RuntimeError(f"ffmpeg stalled: no progress for {int(stall_timeout_s)}s")
+                raise StatedFailure(f"ffmpeg stalled: no progress for {int(stall_timeout_s)}s")
         try:
             ret = proc.wait(timeout=_FFMPEG_EXIT_GRACE_S)
         except subprocess.TimeoutExpired:  # pipes closed but the process lingers
-            raise RuntimeError("ffmpeg did not exit after closing its output")
+            raise StatedFailure("ffmpeg did not exit after closing its output") from None
     except BaseException:
         # ONE exit path stops the process, so the invariant is unconditional: nothing leaves this
         # function with ffmpeg still running. Not just our own EditAborted/stall — `on_progress`
@@ -435,7 +471,7 @@ def run_ffmpeg(
             if pipe:
                 pipe.close()
     if ret != 0:
-        raise RuntimeError(f"ffmpeg failed ({ret}): {''.join(err_tail)[-300:]}")
+        raise StatedFailure(f"ffmpeg failed ({ret}): {''.join(err_tail)[-300:]}")
 
 
 # --- ffmpeg: universal browser playability --------------------------------------------------
@@ -471,7 +507,13 @@ def run_ffprobe_json(
     try:
         out = subprocess.run(args, capture_output=True, text=True, timeout=30, check=True).stdout
         return json.loads(out)
-    except (subprocess.TimeoutExpired, subprocess.CalledProcessError, OSError, ValueError):
+    except (subprocess.TimeoutExpired, subprocess.CalledProcessError, OSError, ValueError) as exc:
+        # None makes every field it would have filled read as absent, and a file with no streams
+        # and no duration is indistinguishable here from one that is not media at all.
+        log.warning(
+            "ffprobe could not read a file",
+            extra=logs.event("media.ffprobe_failed", reason=type(exc).__name__),
+        )
         return None
 
 
@@ -525,7 +567,13 @@ def probe_duration(path: Path) -> int:
             check=True,
         ).stdout.strip()
         return int(float(out)) if out else 0
-    except (subprocess.TimeoutExpired, subprocess.CalledProcessError, OSError, ValueError):
+    except (subprocess.TimeoutExpired, subprocess.CalledProcessError, OSError, ValueError) as exc:
+        # Zero is the denominator of the re-encode progress bar, so a failed probe leaves the user
+        # watching a bar that never moves rather than a job that says it went wrong.
+        log.info(
+            "the duration probe failed; progress for this job has no total",
+            extra=logs.event("media.duration_probe_failed", reason=type(exc).__name__),
+        )
         return 0
 
 
@@ -670,15 +718,28 @@ def _run_ffmpeg_best_effort(cmd: list[str], timeout_s: int) -> Exception | None:
     try:
         subprocess.run(cmd, capture_output=True, timeout=timeout_s, check=True)
         return None
+    # silent-ok: the exception is the RETURN VALUE. This helper exists so the three callers share
+    # one run/timeout policy, and each logs the failure with the context only it has.
     except (subprocess.TimeoutExpired, subprocess.CalledProcessError, OSError) as exc:
         return exc
 
 
+def _step_failure(err: BaseException) -> str:
+    """What a best-effort image step's failure line says: the kind and, for ffmpeg, its exit
+    status. Never `str(err)`: a CalledProcessError or TimeoutExpired quotes ffmpeg's argv, and an
+    OSError its file name — the source path either way, which is built from the channel and the
+    title."""
+    code = getattr(err, "returncode", None)
+    return f"{type(err).__name__} {code}" if code is not None else type(err).__name__
+
+
 def _is_jpeg(path: Path) -> bool:
     """True if the file starts with the JPEG SOI marker (FF D8 FF)."""
     try:
         with path.open("rb") as f:
             return f.read(3) == b"\xff\xd8\xff"
+    # silent-ok: False only means "do not copy this straight through" — the caller then goes the
+    # ffmpeg route, which opens the same file and reports it properly if it really is unreadable.
     except OSError:
         return False
 
@@ -699,13 +760,15 @@ def thumbnail_to_jpg(src: Path | None, dest: Path, timeout_s: int = 30) -> Path
             shutil.copyfile(src, dest)
             return dest
         except OSError as exc:
-            log.info("thumbnail copy failed (%s); poster falls back to a cut frame", str(exc)[:120])
+            log.info(
+                "thumbnail copy failed (%s); poster falls back to a cut frame", _step_failure(exc)
+            )
             return None
     err = _run_ffmpeg_best_effort(build_thumbnail_jpg_cmd(src, dest), timeout_s)
     if err is not None:
         log.info(
             "thumbnail->jpg conversion failed (%s); poster falls back to a cut frame",
-            str(err)[:120],
+            _step_failure(err),
         )
         return None
     return dest if dest.exists() else None
@@ -730,7 +793,7 @@ def ensure_poster(asset, download_root: str, timeout_s: int = 30) -> str | None:
     at = min(3.0, (asset.duration_s or 10) * 0.1)
     err = _run_ffmpeg_best_effort(build_poster_cmd(src, dest, at), timeout_s)
     if err is not None:
-        log.info("poster gen skipped for asset %s: %s", asset.id, str(err)[:120])
+        log.info("poster gen skipped for asset %s: %s", asset.id, _step_failure(err))
         return None
     return rel if dest.exists() else None
 
@@ -761,7 +824,7 @@ def ensure_storyboard(asset, download_root: str, timeout_s: int = 90) -> dict |
     cmd = build_storyboard_cmd(src, dest, asset.duration_s, geom)
     err = _run_ffmpeg_best_effort(cmd, timeout_s)
     if err is not None:
-        log.info("storyboard gen skipped for asset %s: %s", asset.id, str(err)[:120])
+        log.info("storyboard gen skipped for asset %s: %s", asset.id, _step_failure(err))
         return None
     if not dest.exists():
         return None

backend/app/downloads/errors.py

@@ -212,6 +212,43 @@ _CATALOG: tuple[_Entry, ...] = (
         cookie=False,
         link=facebook_group_post_id,
     ),
+    # YouTube gates INDIVIDUAL videos behind "Sign in to confirm you're not a bot" (measured on prod
+    # job 273, a festival set: all 12 of yt-dlp's player clients fail on that one video while its
+    # neighbours download fine, the POT sidecar is live and answering, and the newest yt-dlp fails
+    # identically — so this is the video's own signed-out gate, not a broken stack or a stale yt-dlp).
+    # The marker is quote-agnostic on purpose: YouTube writes the apostrophe as U+2019, so
+    # "you're not a bot" with a straight quote never matches the real text.
+    _Entry(
+        host="youtube",
+        markers=("[youtube]", "not a bot"),
+        code="login_required",
+        detail="youtube.bot_wall",
+        cookie=False,
+    ),
+    # The same wall WITH a cookie. Not yet reproduced with a real YouTube cookie (this user's first
+    # one is being added now) — but "add cookies" is provably the wrong advice for someone who has,
+    # and yt-dlp names the likeliest cause itself: "The provided YouTube account cookies are no
+    # longer valid. They have likely been rotated in the browser as a security measure."
+    _Entry(
+        host="youtube",
+        markers=("[youtube]", "not a bot"),
+        code="login_required",
+        detail="youtube.cookie_rejected",
+        cookie=True,
+    ),
+    # The same gate one step later: the player response comes back, and then the CDN refuses the
+    # stream itself ("unable to download video data: HTTP Error 403"). Measured 2026-09-16 on the
+    # very same video, ~90 minutes after it was answering with the bot message — signed out it
+    # 403'd on every attempt; with the owner's cookie attached it downloaded in full (519 MB). So
+    # the category this text carries generically, `network`, is a lie here: nothing was wrong with
+    # the network. Scoped to a cookie-less YouTube download, where that is what we measured.
+    _Entry(
+        host="youtube",
+        markers=("unable to download video data", "403"),
+        code="login_required",
+        detail="youtube.media_forbidden",
+        cookie=False,
+    ),
 )
 
 

backend/app/downloads/formats.py

@@ -17,8 +17,134 @@ spec shape (all keys optional; defaults + browser-playability clamp applied by `
 `container`/`vcodec` keys survive only so an old stored profile deserializes without error.
 """
 
+import contextlib
+import contextvars
 import hashlib
 import json
+import logging
+import os
+import re
+import urllib.parse
+from collections.abc import Iterator
+
+from app import logs
+
+log = logs.get_logger(__name__)
+
+# yt-dlp names what it is working on in these shapes, and each one is a row of someone's download
+# history: the source URL, the `[extractor] <id>:` prefix on nearly every warning and error, the
+# file path it writes to (`storage.rel_path` builds it from the channel and the title), a
+# playlist's title, and an id quoted mid-sentence. The job id is already on every line
+# (`logs.set_job` in the download child), so none of them adds anything an operator needs. What
+# stays: the host (it says which extractor is failing) and the extractor tag. Every pattern below
+# is a shape measured in yt-dlp's own source (2026.07.04), pinned in test_ytdlp_logging.py.
+_URL_RE = re.compile(r"https?://([^/\s]+)\S*")
+# `Destination:` is yt-dlp's own word in `[download] Destination: <path>`, not an id — the path
+# after it is what `_PATH_RE` takes out.
+_ID_PREFIX_RE = re.compile(r"(\[[\w:.-]+\] )(?!Destination:)[^\s:\[\]]+:")
+# The generic extractor's "id" is the URL's basename, percent-DECODED: `My%20Holiday.mp4` is
+# `My Holiday` — spaces, brackets, colons and all, which no pattern can bound (`Our Trip: Day 1`
+# ends at which `: `?). The job knows its own URL, so the exact name is taken out as a literal
+# (`hiding`). This pattern is the fallback for a line logged outside that scope: up to the first
+# `: `. Only this extractor: every other one names a site id, and letting the prefix run to `: `
+# everywhere would take the words out of lines like `[VideoConvertor] Converting …; Destination:`.
+_GENERIC_ID_RE = re.compile(r"(\[generic\] )(?!Destination:)[^\[\]\n]+?:(?= |$)", re.M)
+# `[youtube:tab] Playlist <title>: Downloading 5 items` — the title is not an id, and has spaces.
+_PLAYLIST_TITLE_RE = re.compile(r"(\] Playlist ).+?(?=: Downloading )")
+# `[download] Finished downloading playlist: <title>`, `[download] Downloading playlist: <title>`,
+# `… Skipping already downloaded playlist: <title>` — the title runs to the end of the line.
+_TITLE_TAIL_RE = re.compile(r"(playlist: |\[download\] Downloading \w+: ).*", re.I)
+# An id quoted mid-sentence, not as a prefix: `… not available for <id>`, and `--no-playlist`'s
+# `Downloading just the video <id>` (`_yes_playlist`) / `Downloading just video <id>` (multifeed).
+_BARE_ID_RE = re.compile(
+    r"((?:subtitles not available for|Downloading just (?:the )?\w+) )[^\s,;]+"
+)
+# Two segments or more, and not glued to a word: `www.youtube.com/…` is what the URL pattern
+# leaves behind, and it must survive this one.
+_PATH_RE = re.compile(r"(?<![\w.…])/(?:[^/\s\"']+/)+[^/\s\"']*")
+
+
+_HIDDEN: contextvars.ContextVar[tuple[str, ...]] = contextvars.ContextVar(
+    "formats_hidden", default=()
+)
+
+
+def generic_id(url: str) -> str:
+    """The name yt-dlp's generic extractor gives a URL: its basename, extension off, percent-
+    decoded — `InfoExtractor._generic_id` (2026.07.04), pinned against it in the tests."""
+    return urllib.parse.unquote(os.path.splitext(url.rstrip("/").split("/")[-1])[0])
+
+
+@contextlib.contextmanager
+def hiding(url: str | None) -> Iterator[None]:
+    """For the work on one URL: `log_safe` also takes out the name yt-dlp's generic extractor
+    gives it, exactly, where yt-dlp puts it (`[generic] <name>:`). A context variable, so it
+    holds for yt-dlp's logger calls too — they run on the same thread as the extraction."""
+    name = generic_id(url) if url else ""
+    token = _HIDDEN.set((*_HIDDEN.get(), name) if name else _HIDDEN.get())
+    try:
+        yield
+    finally:
+        _HIDDEN.reset(token)
+
+
+def log_safe(text: object) -> str:
+    """yt-dlp's text (its own lines, or a DownloadError's message) with the parts that name a
+    person's download taken out. Every log line that carries yt-dlp's words goes through here,
+    not just the adapter's: yt-dlp raises the same text it has just logged, so a line that skipped
+    it would put back one line later exactly what the adapter took out."""
+    text = str(text)
+    for name in _HIDDEN.get():
+        text = text.replace(f"[generic] {name}:", "[generic] …:")
+    text = _URL_RE.sub(lambda m: f"{m.group(1)}/…", text)
+    text = _ID_PREFIX_RE.sub(r"\1…:", text)
+    text = _GENERIC_ID_RE.sub(r"\1…:", text)
+    text = _PLAYLIST_TITLE_RE.sub(r"\1…", text)
+    text = _TITLE_TAIL_RE.sub(r"\1…", text)
+    text = _BARE_ID_RE.sub(r"\1…", text)
+    return _PATH_RE.sub("…", text)
+
+
+class _YdlLog:
+    """yt-dlp's logger protocol, pointed at ours.
+
+    Without it yt-dlp explains itself to nobody: `quiet` sends its output to a stdout we do not
+    read, so "Requested format is not available", a signature-extraction failure or a POT refusal
+    reached us only as whatever text ended up in the raised exception. Measured against yt-dlp
+    2026.07.04: with `logger` set it calls `debug` for extraction and progress lines and
+    `warning`/`error` for the rest, and `quiet`/`no_warnings` do NOT gate those calls.
+
+    `debug` goes to TRACE: it is several lines per second during a download, so it shows only
+    when the whole process runs at TRACE. (Not under a user's own Teljes consent — that lives in
+    the API process's request scope, and the download child that runs yt-dlp has none.)
+
+    `error` goes to WARNING, not ERROR. yt-dlp calls it from `trouble()` just before it raises,
+    on EVERY failed attempt — including the ones the pipeline then recovers from with its fallback
+    client set. Whether the job failed is the pipeline's verdict (`download.job.failed`), and that
+    is the one line whose level should say so."""
+
+    def _emit(self, level: int, msg: str) -> None:
+        # The line goes as an ARGUMENT, not as the format string. yt-dlp writes progress as
+        # `[download]  53.2% of 12.00MiB`, and while stdlib logging only applies `%` formatting to
+        # a record that has args, this way the question never arises — and it is the shape ruff's
+        # `G` rules ask for. (The sink's flood limiter groups a bare `%s` by the text itself, so
+        # distinct yt-dlp warnings do not share one bucket — `logsink._template`.)
+        log.log(level, "%s", log_safe(msg), extra=logs.event("download.ytdlp"))
+
+    def debug(self, msg):
+        self._emit(logs.TRACE, msg)
+
+    def info(self, msg):
+        self._emit(logging.DEBUG, msg)
+
+    def warning(self, msg):
+        self._emit(logging.WARNING, msg)
+
+    def error(self, msg):
+        self._emit(logging.WARNING, msg)
+
+
+_YDL_LOG = _YdlLog()
 
 # Fields that affect the produced file — and therefore the cache identity.
 _SIG_KEYS = (
@@ -78,7 +204,13 @@ def normalize(spec: dict) -> dict:
         # rather than raising an unhandled 500 at enqueue.
         try:
             s["max_height"] = int(s["max_height"])
-        except (TypeError, ValueError):
+        except (TypeError, ValueError) as exc:
+            # "best" is the opposite of what a height cap asks for: the user set a ceiling to save
+            # bandwidth or disk, and the download quietly takes the largest rendition there is.
+            log.warning(
+                "the height cap on a download could not be read and was dropped",
+                extra=logs.event("download.format.max_height_dropped", reason=type(exc).__name__),
+            )
             s["max_height"] = None
     for b in ("embed_subs", "embed_chapters", "embed_thumbnail", "sponsorblock"):
         s[b] = bool(s[b])
@@ -181,6 +313,12 @@ def build_ydl_opts(
     if extra:
         opts.update(extra)
 
+    # With a logger attached, `quiet` and `no_warnings` stop deciding whether yt-dlp explains itself
+    # and only decide where — so both stay on above and the explanation comes here instead of to a
+    # stdout nothing reads. AFTER the adapter merge, like `cookiefile` and the size cap: an adapter
+    # must not be able to silence the one account of what went wrong that a failed job leaves.
+    opts["logger"] = _YDL_LOG
+
     # SEC-13: hard ceiling on a single download. AFTER the adapter merge, for the same reason
     # `cookiefile` is — a source adapter must not be able to raise a limit the quota layer set.
     #

backend/app/downloads/gc.py

@@ -16,7 +16,6 @@ Five passes each run:
 Runs in the API process (which mounts DOWNLOAD_ROOT). Pure disk/DB work — no YouTube quota.
 """
 
-import logging
 import time
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
@@ -24,14 +23,13 @@ from pathlib import Path
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.config import settings
 from app.downloads import storage
 from app.models import DownloadJob, MediaAsset
 from app.notifications import create_notification
 
-log = logging.getLogger("siftlode.downloads.gc")
+log = logs.get_logger(__name__)
-
 # Dot-prefixed top-level trees that are NOT part of the Plex/flat asset layout — the orphan sweep
 # never touches them (user clips, Plex HLS/img caches, in-progress staging, editor storyboards).
 _SYSTEM_TREE = {".edits", ".plex-hls", ".plex-img-cache", ".staging", ".storyboards"}
@@ -99,6 +97,7 @@ def _sweep_orphan_files(db: Session, download_root: str) -> int:
         if pause_s and touched % batch == 0:
             time.sleep(pause_s)
 
+    unreclaimed = 0
     for entry in root.iterdir():
         if entry.name.startswith("."):  # .edits / .plex-* / .staging / .storyboards + any dotfile
             continue
@@ -111,7 +110,9 @@ def _sweep_orphan_files(db: Session, download_root: str) -> int:
                 f.unlink()
                 removed += 1
             except OSError:
-                pass
+                # silent-ok: counted, not logged per file — a broken mount would put one line
+                # under every orphan. The total is reported once at the end of the sweep.
+                unreclaimed += 1
         if entry.is_dir():  # prune emptied dirs (deepest first), then the channel dir itself
             for d in sorted(
                 (p for p in entry.rglob("*") if p.is_dir()),
@@ -120,12 +121,20 @@ def _sweep_orphan_files(db: Session, download_root: str) -> int:
             ):
                 try:
                     d.rmdir()
+                # silent-ok: `rmdir` refusing means the directory still holds files, which is
+                # the ordinary answer here — this prune only takes the ones that emptied.
                 except OSError:
                     pass
             try:
                 entry.rmdir()
+            # silent-ok: the same prune one level up, at the channel directory.
             except OSError:
                 pass
+    if unreclaimed:
+        log.warning(
+            "some orphaned download files could not be reclaimed",
+            extra=logs.event("download.gc.unreclaimed", files=unreclaimed, removed=removed),
+        )
     return removed
 
 

backend/app/downloads/links.py

@@ -49,6 +49,8 @@ def check_grant(token: str, grant: str | None, password_version: int) -> bool:
     exp_s, sig = grant.split(".", 1)
     try:
         exp = int(exp_s)
+    # silent-ok: an untrusted grant that does not parse is simply not a valid grant, and the
+    # honest answer to it — a password prompt — is the same as for one that expired.
     except ValueError:
         return False
     if exp < datetime.now(timezone.utc).timestamp():

backend/app/downloads/og.py

@@ -66,6 +66,8 @@ def _jpeg_size(path: Path) -> tuple[int, int] | None:
                     width = (data[3] << 8) + data[4]
                     return width, height
                 f.seek(length - 2, 1)  # skip this segment's payload
+    # silent-ok: None omits width/height from the unfurl card, which is what the card does for
+    # every poster whose dimensions were never asked for. Nothing is claimed that is not true.
     except OSError:
         return None
 

backend/app/downloads/pipeline.py

@@ -1,6 +1,5 @@
 """Child-side download/edit pipeline — the yt-dlp download + ffmpeg edit that runs in the spawned subprocess."""
 
-import logging
 import re
 import shutil
 import time
@@ -9,7 +8,7 @@ from pathlib import Path
 
 import yt_dlp
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.config import settings
 from app.db import SessionLocal
 from app.downloads import edit as editmod
@@ -24,12 +23,13 @@ from app.downloads.jobdb import (
     _staging_dir,
     _touch_job,
 )
+from app.downloads.stated import StatedFailure
 from app.downloads.status import JobStatus
-from app.models import DownloadJob, MediaAsset
+from app.models import DISPLAY_TEXT_MAX, DownloadJob, MediaAsset
 from app.notifications import create_notification
 from app.titles import normalize_title
 
-log = logging.getLogger("siftlode.worker")
+log = logs.get_logger(__name__)
 
 
 class _Aborted(Exception):
@@ -98,6 +98,8 @@ def _parse_upload_date(raw) -> date | None:
         return None
     try:
         return datetime.strptime(str(raw), "%Y%m%d").date()
+    # silent-ok: a date we cannot read becomes a date the card does not show, which is what the
+    # card shows for every source that never supplies one. Absent is the honest rendering.
     except ValueError:
         return None
 
@@ -138,7 +140,14 @@ def _make_progress_hook(job_id: int, asset_id: int):
             try:
                 _fill_asset_meta_early(asset_id, d["info_dict"])
             except Exception:  # noqa: BLE001 — metadata is best-effort, never fail a download
-                pass
+                # Best-effort is about the DOWNLOAD, not about knowing. What this swallows is
+                # anything from a shape change in yt-dlp's info dict to the database being down,
+                # and all the user sees is a card with an id on it until the job finishes.
+                log.info(
+                    "early metadata could not be filled in for this download",
+                    extra=logs.event("download.meta.early_fill_failed", asset_id=asset_id),
+                    exc_info=True,
+                )
         if d.get("status") != "downloading":
             return
         if not allow_write():
@@ -299,7 +308,7 @@ def _produced_file(info: dict, staging: Path) -> Path:
         and not _YTDLP_TEMP_RE.search(p.name)
     ]
     if not cands:
-        raise RuntimeError("yt-dlp produced no output file")
+        raise StatedFailure("yt-dlp produced no output file")
     return max(cands, key=lambda p: p.stat().st_size)
 
 
@@ -375,13 +384,22 @@ def _extract_with_fallback(
                     "job %s: clients %s failed (%s); retrying with fallback",
                     job_id,
                     clients,
-                    str(exc)[:70],
+                    formats.log_safe(exc)[:70],
                 )
                 for p in list(staging.iterdir()):  # drop partials before the retry
                     try:
                         p.unlink()
-                    except OSError:
+                    except OSError as exc:
-                        pass
+                        # A partial left behind is not inert: the next client set resumes into it,
+                        # so the retry this loop exists to make clean starts dirty instead.
+                        log.info(
+                            "a partial file could not be removed before the retry",
+                            extra=logs.event(
+                                "download.retry.partial_unlink_failed",
+                                job_id=job_id,
+                                reason=type(exc).__name__,
+                            ),
+                        )
                 _set_job(job_id, progress=0, phase="downloading")
                 continue
             raise
@@ -464,7 +482,7 @@ def _download(
             try:
                 urlguard.check_url(url)
             except urlguard.BlockedUrl as e:
-                raise RuntimeError(f"Refusing to download from this address ({e.reason}).") from e
+                raise StatedFailure(f"Refusing to download from this address ({e.reason}).") from e
         # S6: attach the acting user's per-host cookie ONLY when it's host-scoped to this URL AND
         # this asset is that user's private, salted identity (the double gate lives in credentials).
         remaining_bytes: int | None = None
@@ -579,20 +597,30 @@ def _download(
             }
             if job and not job.display_name and meta.title:
                 # display_name is VARCHAR(255) — clip a long title (see _fill_asset_meta_early).
-                fields["display_name"] = meta.title[:255]
+                fields["display_name"] = meta.title[:DISPLAY_TEXT_MAX]
             # Atomic terminal write: finalizes ONLY while the row is still running, so a child that
             # finishes just after a route cancel/pause can't resurrect it (see finalize_if_running).
             finalized = service.finalize_if_running(db, job_id, **fields)
             db.commit()
             if finalized:
                 _notify_download_done(db, owner_id, meta.title, job_id, kind="download")
-                log.info("job %s done: %s", job_id, rel)
+                # The asset id, not `rel`: the path is built from the channel and the title.
+                log.info(
+                    "job %s done: asset %s",
+                    job_id,
+                    asset_id,
+                    extra=logs.event("download.job.done", job_id=job_id, asset_id=asset_id),
+                )
 
     except (_Aborted, editmod.EditAborted):
         # Paused/canceled mid-download or mid-transcode: free the asset for a later retry; keep the
         # job's status (the route set it to paused/canceled). ref_count is adjusted by the route.
         _reset_asset_pending(asset_id)
-        log.info("job %s aborted (pause/cancel)", job_id)
+        log.info(
+            "job %s aborted (pause/cancel)",
+            job_id,
+            extra=logs.event("download.job.aborted", job_id=job_id),
+        )
     except _OwnerRefused as exc:
         # Fail this job only. The asset is never marked `error`; it goes back to `pending` (below),
         # so another user's job for the same video is claimed and downloaded normally instead of
@@ -613,7 +641,12 @@ def _download(
         # it, and this one must already be out of the way. (The staging tree is still torn down in
         # `finally` below — the same narrow window the pause/cancel path above has always had.)
         _reset_asset_pending(asset_id)
-        log.info("job %s refused for its owner: %s", job_id, msg)
+        log.info(
+            "job %s refused for its owner: %s",
+            job_id,
+            formats.log_safe(msg),
+            extra=logs.event("download.job.refused", job_id=job_id),
+        )
     except Exception as exc:  # noqa: BLE001 — surface any failure to the user, keep worker alive
         msg = str(exc)[:500]
         # The catalog needs the link and whether the user's cookie went with it: the same yt-dlp
@@ -626,8 +659,9 @@ def _download(
             if asset:
                 asset.status = "error"
                 asset.error = msg
-            # Atomic fail: only while still running, so a route-set cancel/pause is never overwritten
+            # Atomic fail: only while still running, so a route-set cancel/pause is never
-            # with `error` (error also holds the asset → would re-open the double-release class).
+            # overwritten with `error` (error also holds the asset → would re-open the
+            # double-release class).
             service.finalize_if_running(
                 db,
                 job_id,
@@ -638,7 +672,10 @@ def _download(
                 phase=None,
             )
             db.commit()
-        log.warning("job %s failed: %s", job_id, msg)
+        # After the write: a job the write could not fail is not a failed job. When it raises,
+        # the runner reports this failure with it (`runner._run`) and the worker fails the job.
+        # silent-ok: not silent — `_log_download_failed` is the report.
+        _log_download_failed(job_id, msg, diagnosis.code, exc)
     finally:
         # Unlink the decrypted cookie file explicitly (the rmtree below removes the whole staging
         # tree too, but never leave a plaintext session cookie on disk a moment longer than needed).
@@ -647,7 +684,50 @@ def _download(
         shutil.rmtree(staging, ignore_errors=True)
 
 
-def _fail_edit(job_id: int, asset_id: int, msg: str) -> None:
+def _failure_stack(exc: BaseException) -> dict[str, str]:
+    """The `extra=` fragment that puts a failure's traceback on its line — or nothing.
+
+    Nothing for yt-dlp's DownloadError and for our own `StatedFailure` (`ffmpeg failed (1): …`):
+    their message is the whole story, and their stack is only the `except` that caught them.
+    Everything else keeps it — a TypeError on an odd info dict, a plain RuntimeError, an OSError,
+    a database error: that is our code or our platform failing, and the stack is the lead.
+
+    Scrubbed, link by link (`logs.scrubbed_stack`): a stack's messages are data — an OSError names
+    its file, and a database error raised inside `except DownloadError` carries yt-dlp's
+    unscrubbed text as its context."""
+    if isinstance(exc, (yt_dlp.utils.DownloadError, StatedFailure)):
+        return {}
+    return logs.scrubbed_stack(exc, formats.log_safe)
+
+
+def _log_download_failed(job_id: int, msg: str, code: str | None, exc: BaseException) -> None:
+    """The child's `download.job.failed` line. `msg` is the user's copy and keeps their URL; the
+    log's copy goes through `log_safe`, or it would put back one line later the URL and id the
+    yt-dlp adapter took out of the line before it. The stack only when the classifier could not
+    place the failure: a recognised one is fully described by its code."""
+    unplaced = code in (None, "", "unknown")
+    log.warning(
+        "job %s failed: %s",
+        job_id,
+        formats.log_safe(msg),
+        extra={
+            **logs.event(
+                "download.job.failed",
+                job_id=job_id,
+                error_code=code,
+                error_type=type(exc).__name__,
+            ),
+            **(_failure_stack(exc) if unplaced else {}),
+        },
+    )
+
+
+def _fail_edit(job_id: int, asset_id: int, msg: str, exc: BaseException | None = None) -> None:
+    """Fail an edit job, and say so ONCE — every way an edit fails comes through here, so this is
+    where its line and its event code live. `exc` is the exception when there was one.
+
+    Its traceback goes with the line unless the message already tells it all (`_failure_stack`).
+    Logged after the write, as on the download error path."""
     with SessionLocal() as db:
         asset = db.get(MediaAsset, asset_id)
         if asset:
@@ -663,7 +743,22 @@ def _fail_edit(job_id: int, asset_id: int, msg: str) -> None:
             phase=None,
         )
         db.commit()
-    log.warning("edit job %s failed: %s", job_id, msg)
+    # `msg` is the user's copy; ffmpeg's stderr tail in it can name the source file, whose path is
+    # built from the channel and the title.
+    log.warning(
+        "edit job %s failed: %s",
+        job_id,
+        formats.log_safe(msg),
+        extra={
+            **logs.event(
+                "edit.job.failed",
+                job_id=job_id,
+                asset_id=asset_id,
+                **({"error_type": type(exc).__name__} if exc is not None else {}),
+            ),
+            **(_failure_stack(exc) if exc is not None else {}),
+        },
+    )
 
 
 def _process_edit(job_id: int, asset_id: int) -> None:
@@ -715,7 +810,7 @@ def _process_edit(job_id: int, asset_id: int) -> None:
             should_cancel=_cancel_poll(job_id),
         )
         if not dest_staging.exists():
-            raise RuntimeError("ffmpeg produced no output file")
+            raise StatedFailure("ffmpeg produced no output file")
 
         rel = storage.edit_rel_path(user_id, asset_id, out_ext)
         storage.place_file(dest_staging, settings.download_root, rel)
@@ -751,13 +846,24 @@ def _process_edit(job_id: int, asset_id: int) -> None:
             db.commit()
             if finalized:
                 _notify_download_done(db, owner_id, dl_title, job_id, kind="edit")
-                log.info("edit job %s done: %s", job_id, rel)
+                # The asset id, not `rel`: the path is built from the channel and the title.
+                log.info(
+                    "edit job %s done: asset %s",
+                    job_id,
+                    asset_id,
+                    extra=logs.event("edit.job.done", job_id=job_id, asset_id=asset_id),
+                )
 
     except editmod.EditAborted:
         # Paused/canceled mid-encode: free the asset for a later retry (ref_count is the route's job).
         _reset_asset_pending(asset_id)
-        log.info("edit job %s aborted (pause/cancel)", job_id)
+        log.info(
+            "edit job %s aborted (pause/cancel)",
+            job_id,
+            extra=logs.event("edit.job.aborted", job_id=job_id),
+        )
     except Exception as exc:  # noqa: BLE001 — surface the failure, keep the worker alive
-        _fail_edit(job_id, asset_id, str(exc)[:500])
+        # silent-ok: not silent — `_fail_edit` is the report, and it takes `exc` for its stack.
+        _fail_edit(job_id, asset_id, str(exc)[:500], exc)
     finally:
         shutil.rmtree(staging, ignore_errors=True)

backend/app/downloads/probe.py

@@ -64,6 +64,8 @@ def _facts_from_probe(
         raw = fmt.get("duration")
         if raw is not None:
             duration = int(float(raw))
+    # silent-ok: an unusable duration is stored as no duration, which the UI renders as unknown
+    # rather than as zero — the honest state, and the one a re-probe can still fill in.
     except (TypeError, ValueError):
         duration = None
     return vcodec, acodec, width, height, duration

backend/app/downloads/runner.py

@@ -11,18 +11,20 @@ writes the terminal job/asset state itself (done / error, or a cooperative pause
 progress hook), then exits 0. An unexpected crash exits non-zero and the parent fails the job.
 """
 
-import logging
 import os
 import shutil
 import subprocess
 import sys
 
+from app import logs
+from app import logsink
 from app.config import settings
 from app.db import SessionLocal
 from app.downloads import edit as editmod
+from app.downloads import formats
 from app.models import DownloadJob
 
-log = logging.getLogger("siftlode.worker.runner")
+log = logs.get_logger(__name__)
 
 
 def _deprioritize() -> None:
@@ -52,12 +54,21 @@ def _deprioritize() -> None:
 
 
 def main() -> None:
-    logging.basicConfig(level=logging.INFO)
+    logs.configure_logging("runner")
     _deprioritize()
     if len(sys.argv) != 2:
-        print("usage: python -m app.downloads.runner <job_id>", file=sys.stderr)
+        # T201: a usage line is for the person who typed the command. Logging is already
+        # configured here, and in a deployment it writes JSON records — the one reader of this
+        # line wants the bare sentence, not a record with a request id they do not have.
+        print(  # noqa: T201
+            "usage: python -m app.downloads.runner <job_id>", file=sys.stderr
+        )
         raise SystemExit(2)
     job_id = int(sys.argv[1])
+    # This whole process exists to run ONE job — tag every line it prints with it, so the child's
+    # output joins up with the worker's and the API's view of the same download.
+    logs.set_job(job_id)
+    logsink.start()  # flushed at exit (atexit) — this process has no shutdown hook of its own
 
     with SessionLocal() as db:
         job = db.get(DownloadJob, job_id)
@@ -71,14 +82,40 @@ def main() -> None:
     if asset_id is None:
         return
 
+    _run(job_id, asset_id, job_kind, source_kind, source_ref, spec, user_id)
+
+
+def _run(job_id, asset_id, job_kind, source_kind, source_ref, spec, user_id) -> None:
+    """Run the job, and be the last stop before Python's own exception hook.
+
+    What gets past the pipeline is a failure it could not even record (the database refused the
+    write that fails the job). Python's hook would print it to stderr raw — and its context is the
+    download's own failure, whose text is the URL or ffmpeg's stderr naming the file. Here it is
+    one line with the whole chain scrubbed; the non-zero exit then has the worker fail the job
+    (`cause=child_exit`)."""
     # Imported here (not at module top) so the child pays the yt-dlp import cost only when it has
     # real work, and to keep the import graph obvious: runner is a thin entry over the pipeline.
     from app.downloads.pipeline import _download, _process_edit
 
-    if job_kind == "edit":
+    with formats.hiding(source_ref):
-        _process_edit(job_id, asset_id)
+        try:
-    else:
+            if job_kind == "edit":
-        _download(job_id, asset_id, source_kind, source_ref, spec, user_id)
+                _process_edit(job_id, asset_id)
+            else:
+                _download(job_id, asset_id, source_kind, source_ref, spec, user_id)
+        except Exception as exc:  # noqa: BLE001 — whatever it is, it is reported here or raw
+            # TRY400: the stack goes with it scrubbed (`scrubbed_stack`); `exc_info` is raw.
+            log.error(  # noqa: TRY400
+                "the download child for job %s failed",
+                job_id,
+                extra={
+                    **logs.event(
+                        "download.child.failed", job_id=job_id, error_type=type(exc).__name__
+                    ),
+                    **logs.scrubbed_stack(exc, formats.log_safe),
+                },
+            )
+            raise SystemExit(1) from None
 
 
 if __name__ == "__main__":

backend/app/downloads/service.py

@@ -20,7 +20,7 @@ from app.downloads import credentials
 from app.downloads import edit as editmod
 from app.downloads import formats, mediakind, quota, storage
 from app.downloads.status import HOLDING, JobStatus
-from app.models import Channel, DownloadJob, MediaAsset, Video
+from app.models import DISPLAY_TEXT_MAX, Channel, DownloadJob, MediaAsset, Video
 
 
 class EditError(Exception):
@@ -87,6 +87,8 @@ def get_or_create_asset(
         with db.begin_nested():
             db.add(asset)
             db.flush()
+    # silent-ok: the expected outcome of the race this savepoint exists for — two jobs wanting the
+    # same asset. Losing it means the other insert won, and re-selecting is the whole handling.
     except IntegrityError:
         asset = db.execute(stmt).scalar_one()
     return asset
@@ -177,7 +179,7 @@ def enqueue(
         profile_id=profile_id,
         profile_snapshot=spec,
         display_name=display_name
-        or (asset.title[:255] if asset.status == "ready" and asset.title else None),
+        or (asset.title[:DISPLAY_TEXT_MAX] if asset.status == "ready" and asset.title else None),
         status="queued",
         queue_pos=next_queue_pos(db),
     )
@@ -196,10 +198,13 @@ def enqueue(
 
 
 def _clip_title(source_title: str | None, display_name: str | None) -> str:
+    # Clipped to DISPLAY_TEXT_MAX, not to a literal: this feeds `DownloadJob.display_name` through
+    # `enqueue_edit`, the same column `display_text` guards. A machine-made title, so it is clipped
+    # rather than run through `display_text` — blank here means "no title", not None.
     if display_name:
-        return display_name[:255]
+        return display_name[:DISPLAY_TEXT_MAX]
     base = source_title or "Clip"
-    return f"{base} (clip)"[:255]
+    return f"{base} (clip)"[:DISPLAY_TEXT_MAX]
 
 
 def enqueue_edit(
@@ -253,7 +258,7 @@ def enqueue_edit(
         source_asset_id=src_asset.id,
         edit_spec=spec,
         profile_snapshot={},
-        display_name=display_name or asset.title,
+        display_name=display_name or asset.title[:DISPLAY_TEXT_MAX],
         status="queued",
         queue_pos=next_queue_pos(db),
     )

backend/app/downloads/stated.py

@@ -0,0 +1,13 @@
+"""The failures the download pipeline raises itself, with the whole story in the message.
+
+`ffmpeg failed (1): …`, `yt-dlp produced no output file`, a post that shows no video: the message
+says what happened, and the stack is only the `except` that caught it. A failure line leaves that
+stack off (`pipeline._failure_stack`) — but ONLY for these. A plain `RuntimeError` is also what
+Python raises for a bug of ours (a dict changed size during iteration, a generator's
+StopIteration, `RecursionError`, `NotImplementedError`), and that one's stack is the only lead.
+"""
+
+
+class StatedFailure(RuntimeError):
+    """A failure we raised on purpose, whose message is its full explanation. A `RuntimeError` so
+    every `except RuntimeError` that caught these before still does."""

backend/app/downloads/storage.py

@@ -19,8 +19,11 @@ from datetime import date
 from pathlib import Path
 from xml.sax.saxutils import escape
 
+from app import logs
 from app.downloads import mediakind
 
+log = logs.get_logger(__name__)
+
 # Reserve room for the "_id.ext" suffix + parent dirs within the 255-char component limit.
 _MAX_TITLE = 120
 # User-facing (Content-Disposition) names can be longer than an on-disk component.
@@ -294,7 +297,13 @@ def write_sidecars(download_root: str, rel: str, meta: MediaMeta, thumb_src: Pat
         if thumb_src and thumb_src.exists():
             shutil.copyfile(str(thumb_src), str(_sidecar(media, ".jpg")))
         return True
-    except OSError:
+    except OSError as exc:
+        # The download itself still completes, so the job reads as done while the library it was
+        # downloaded FOR has no .nfo to read and no poster to show.
+        log.warning(
+            "the sidecars beside a download could not be written",
+            extra=logs.event("download.sidecars.write_failed", reason=type(exc).__name__),
+        )
         return False
 
 
@@ -304,14 +313,23 @@ def delete_asset_files(download_root: str, rel: str) -> None:
     for p in (media, _sidecar(media, ".nfo"), _sidecar(media, ".jpg")):
         try:
             p.unlink()
-        except OSError:
+        except FileNotFoundError:
-            pass
+            pass  # silent-ok: already gone is the outcome this function wanted
+        except OSError as exc:
+            # The row goes and the bytes stay. Nothing else ever revisits this path, so the disk
+            # quietly keeps a file no account has any way to see, let alone delete.
+            log.info(
+                "a deleted download left a file behind",
+                extra=logs.event("download.asset.file_unlink_failed", reason=type(exc).__name__),
+            )
     # Prune empty Season/Channel dirs up to (but not including) the root.
     root = Path(download_root).resolve()
     parent = media.parent
     while parent != root and root in parent.parents:
         try:
             parent.rmdir()  # only succeeds if empty
+        # silent-ok: this `rmdir` failing is the loop's stop condition — the parent still
+        # holds another download's files, so the prune has gone as far up as it may.
         except OSError:
             break
         parent = parent.parent

backend/app/downloads/upload.py

@@ -25,7 +25,7 @@ from sqlalchemy.orm import Session
 from app.config import settings
 from app.downloads import edit as editmod
 from app.downloads import mediakind, probe, quota, service, storage
-from app.models import DownloadJob, MediaAsset
+from app.models import DownloadJob, MediaAsset, display_text
 
 
 def _title_from_filename(filename: str) -> str:
@@ -99,7 +99,7 @@ def ingest_upload(
             source_kind="upload",
             source_ref=ref,
             profile_snapshot={},
-            display_name=(display_name or "").strip() or None,
+            display_name=display_text(display_name),
             status="done",
             progress=100,
             queue_pos=service.next_queue_pos(db),

backend/app/email.py

@@ -11,7 +11,6 @@ to be a person telling them nothing arrived.
 """
 
 import html
-import logging
 import re
 import smtplib
 import ssl
@@ -21,15 +20,14 @@ from email.utils import formatdate
 from sqlalchemy import select
 from sqlalchemy.orm import Session
 
-from app import mailtext, sysconfig
+from app import logs, mailtext, sysconfig
 from app.config import settings
 from app.db import SessionLocal
 from app.models import User
 from app.notifications import create_notification
 from app.ratelimit import RateLimiter
 
-log = logging.getLogger("siftlode.email")
+log = logs.get_logger(__name__)
-
 # One admin notice an hour per reason. A relay that is down fails every send, and the inbox must
 # stay readable; the log keeps the full record either way.
 _alert_throttle = RateLimiter(max_events=1, window_seconds=3600)
@@ -155,7 +153,15 @@ def _send(to: list[str], subject: str, body: str, reply_to: str | None = None) -
         return False
     c = _smtp()
     if not (c["host"] and c["user"] and c["password"]):
-        log.info("SMTP not configured; skipping email %r to %s", subject, recipients)
+        # The count, not the addresses: `redact_text` rewrites every one of them to `[email]`
+        # before either sink sees the line, so the list only ever printed as noise. The subject
+        # stays — the templates are static, and the one that interpolates (`admin_new_request`)
+        # interpolates an address, which the same redactor covers.
+        log.info(
+            "SMTP not configured; skipping email %r to %d recipient(s)",
+            subject,
+            len(recipients),  # log-pii-ok: how many, not who
+        )
         _alert_admins("not_configured", subject, recipients)
         return False
     msg = EmailMessage()
@@ -176,10 +182,12 @@ def _send(to: list[str], subject: str, body: str, reply_to: str | None = None) -
             s.starttls(context=ssl.create_default_context())
             s.login(c["user"], c["password"])
             s.send_message(msg)
-        log.info("Sent email %r to %s", subject, recipients)
+        # log-pii-ok: how many, not who
+        log.info("Sent email %r to %d recipient(s)", subject, len(recipients))
         return True
     except Exception:
-        log.exception("SMTP send failed (%r to %s)", subject, recipients)
+        # log-pii-ok: how many, not who
+        log.exception("SMTP send failed (%r to %d recipient(s))", subject, len(recipients))
         _alert_admins("send_failed", subject, recipients)
         return False
 

backend/app/feedquery.py

@@ -12,9 +12,10 @@ import binascii
 import json
 from datetime import date, datetime, timedelta, timezone
 
-from sqlalchemy import BigInteger, Select, and_, cast, false, func, or_, select
+from sqlalchemy import BigInteger, Select, and_, case, cast, false, func, or_, select
 from sqlalchemy.orm import Session, aliased
 
+from app.collation import collated
 from app.models import (
     LIVE_OR_UPCOMING,
     BlockedChannel,
@@ -30,7 +31,7 @@ from app.models import (
     Video,
     VideoState,
 )
-from app.utils import iso, to_tsquery_str
+from app.utils import iso, like_escape, to_tsquery_str
 
 
 def _channel_url(channel_id: str, handle: str | None) -> str:
@@ -105,15 +106,30 @@ def _tag_visible(user: User):
     return or_(ChannelTag.user_id.is_(None), ChannelTag.user_id == user.id)
 
 
-def tags_by_category(db: Session, tag_ids: list[int]) -> dict[str, list[int]]:
+# The pseudo-category of a selected tag id no Tag row answers (a saved view or the feed state outlives
+# a deleted tag). It filters like a category whose tags no channel carries — to nothing, in the SQL and
+# in feed_summary's mirror alike — instead of the id being dropped, which showed the WHOLE feed under
+# the view's old name (E32 S2). No real category is called this (language/topic/other).
+MISSING_TAG_CATEGORY = "missing"
+
+
+def tags_by_category(db: Session, user: User, tag_ids: list[int]) -> dict[str, list[int]]:
     """Group tag ids by their category (language/topic/other). The single source for the category split
     shared by the feed's tag WHERE (filtered_query) and the facet/count summary (feed_summary), so the
-    two can't disagree on which selected tag belongs to which category."""
+    two can't disagree on which selected tag belongs to which category. An unknown id lands in
+    MISSING_TAG_CATEGORY — and so does another user's private tag: resolving it to its real category
+    told the caller which foreign ids exist (E32 review)."""
     if not tag_ids:
         return {}  # skip the pointless `WHERE id IN ()` round-trip on the no-tag-filter hot path
     out: dict[str, list[int]] = {}
-    for tid, cat in db.execute(select(Tag.id, Tag.category).where(Tag.id.in_(tag_ids))).all():
+    visible = or_(Tag.user_id.is_(None), Tag.user_id == user.id)
+    for tid, cat in db.execute(
+        select(Tag.id, Tag.category).where(Tag.id.in_(tag_ids), visible)
+    ).all():
         out.setdefault(cat, []).append(tid)
+    known = {tid for ids in out.values() for tid in ids}
+    if missing := [tid for tid in tag_ids if tid not in known]:
+        out[MISSING_TAG_CATEGORY] = missing
     return out
 
 
@@ -264,7 +280,8 @@ def filtered_query(
     # the queries that surfaced the video > description; see Video.search_vector). YouTube-like
     # "fuzzy" matching: word-order-independent, multi-word AND, prefix on the word being typed,
     # accent-insensitive (unaccent_simple). The channel name still matches as a substring so
-    # channel searches work. `rank_expr` (weight-aware ts_rank) drives the "relevance" sort.
+    # channel searches work. `rank_expr` (weight-aware ts_rank) drives the "relevance" sort, over the
+    # newest RANK_WINDOW matches only (see there).
     rank_expr = None
     if q:
         # Resolve the channels whose NAME matches up front (there are only a few hundred channels, so
@@ -272,8 +289,9 @@ def filtered_query(
         # `unaccent(channels.title) ILIKE '%q%'` straight into the per-video predicate defeated the
         # search_vector GIN — the planner detoasted every non-matching tsvector to test the `@@`
         # (~2.6s/search on the prod-sized set); channel_id.in_(...) keeps the GIN as a BitmapOr branch
-        # (PERF-12). Empty match -> false() so the OR collapses to the text match alone.
+        # (PERF-12). Empty match -> false() so the OR collapses to the text match alone. The query
-        like = func.unaccent(f"%{q}%")
+        # is a literal inside the pattern: `_` in `my_channel` is not a wildcard.
+        like = func.unaccent(f"%{like_escape(q)}%")
         matching_channel_ids = (
             db.execute(select(Channel.id).where(func.unaccent(Channel.title).ilike(like)))
             .scalars()
@@ -301,7 +319,7 @@ def filtered_query(
         # ALL its selected tags. ⚠️ feed_summary.passes() reimplements this exact rule in Python for the
         # sidebar count/facets — keep the two in lockstep (equivalence-tested by
         # test_feed_summary_matches_count_plus_facets).
-        by_category = tags_by_category(db, tags)
+        by_category = tags_by_category(db, user, tags)
         # Facet counting drops the category being counted so its own chips don't zero each
         # other out (standard drill-down faceting: a category's count ignores its own
         # selections but still honours the other categories' filters).
@@ -392,7 +410,7 @@ def feed_summary(db: Session, user: User, **params) -> dict:
         return {"count": 0, "facets": {}}
 
     # channel -> its visible tag ids, plus each tag's category (for grouping the facet output).
-    visible = or_(ChannelTag.user_id.is_(None), ChannelTag.user_id == user.id)
+    visible = _tag_visible(user)
     channel_tags: dict[str, set[int]] = {}
     tag_category: dict[int, str] = {}
     for cid, tid, cat in db.execute(
@@ -407,7 +425,7 @@ def feed_summary(db: Session, user: User, **params) -> dict:
     # base channel carries still filters everything out for its category — matching the SQL `IN (empty)`
     # behaviour).
     selected_by_cat: dict[str, set[int]] = {
-        cat: set(ids) for cat, ids in tags_by_category(db, params.get("tags") or []).items()
+        cat: set(ids) for cat, ids in tags_by_category(db, user, params.get("tags") or []).items()
     }
 
     def passes(ch_tags: set[int], exclude_cat: str | None) -> bool:
@@ -435,7 +453,7 @@ def feed_summary(db: Session, user: User, **params) -> dict:
         # Drop this category's own selections so its chips keep independent counts; a conjunctive
         # topic-AND keeps them (each remaining chip narrows to channels that ALSO have the selected).
         exclude = None if tag_modes.get(category, "or") == "and" else category
-        for cid, ch_tags in channel_tags.items():
+        for ch_tags in channel_tags.values():
             if not passes(ch_tags, exclude):
                 continue
             for tid in ch_tags:
@@ -477,8 +495,10 @@ def feed_years(db: Session, user: User, **params) -> list[dict]:
 # replaces OFFSET/LIMIT: deep scrolling no longer scans-and-discards skipped rows,
 # and the page boundary is stable even as the scheduler ingests new videos mid-scroll
 # (OFFSET would shift and duplicate/skip rows).
-def sort_keys(sort: str, seed: int, rank_expr=None) -> list[dict]:
+def sort_keys(sort: str, seed: int, rank_expr=None, lang: str | None = None) -> list[dict]:
+    """`lang` orders the title sorts in the reader's alphabet (`collation`); the other keys ignore it."""
     pub = {"expr": Video.published_at, "asc": False, "nullable": False, "kind": "dt"}
+    title = collated(Video.title, lang)
     table: dict[str, list[dict]] = {
         "newest": [pub],
         "oldest": [{**pub, "asc": True}],
@@ -490,10 +510,8 @@ def sort_keys(sort: str, seed: int, rank_expr=None) -> list[dict]:
         "duration_asc": [
             {"expr": Video.duration_seconds, "asc": True, "nullable": True, "kind": "num"}
         ],
-        "title": [{"expr": func.lower(Video.title), "asc": True, "nullable": True, "kind": "str"}],
+        "title": [{"expr": title, "asc": True, "nullable": True, "kind": "str"}],
-        "title_desc": [
+        "title_desc": [{"expr": title, "asc": False, "nullable": True, "kind": "str"}],
-            {"expr": func.lower(Video.title), "asc": False, "nullable": True, "kind": "str"}
-        ],
         "subscribers": [
             {"expr": Channel.subscriber_count, "asc": False, "nullable": True, "kind": "num"}
         ],
@@ -541,6 +559,54 @@ def sort_keys(sort: str, seed: int, rank_expr=None) -> list[dict]:
     return table.get(sort) or table["newest"]
 
 
+# Relevance ranks only the newest RANK_WINDOW matches (E31): ts_rank reads every match's tsvector, and
+# a broad word matches most of the library ("a" ranked 193k rows in 6.8 s on prod, against 1.1 s to just
+# find them). The rows past the window keep rank 0, so they follow the ranked ones newest first; a word
+# that broad has no meaningful relevance order anyway, and a specific one fits the window whole.
+RANK_WINDOW = 2000
+
+
+def rank_window_start(db: Session, query: Select) -> datetime | None:
+    """The published_at the ranked window starts at: the RANK_WINDOW-th newest match, None when fewer
+    match (rank them all). The newest-first walk stops after RANK_WINDOW matches (40 ms for "a")."""
+    return db.scalar(
+        query.with_only_columns(Video.published_at)
+        .order_by(Video.published_at.desc())
+        .offset(RANK_WINDOW - 1)
+        .limit(1)
+    )
+
+
+def windowed_rank(rank_expr, since: datetime | None):
+    """`rank_expr` inside the window, 0 past it. CASE evaluates ts_rank only for the rows inside."""
+    if since is None:
+        return rank_expr
+    return case((Video.published_at >= since, rank_expr), else_=0)
+
+
+def windowed_relevance_rows(
+    db: Session, query: Select, keys: list[dict], window: datetime, cursor_id: str | None, n: int
+) -> list:
+    """Up to `n` rows of the windowed relevance order, read as its two halves: the window ranked, then
+    the rest newest first (every window row sorts before every row past it). As one ORDER BY the
+    planner computes the CASE for every match and sorts them all (2.7 s for "a" on prod); split, each
+    half rides an index (80 ms + 5 ms). With a cursor, `keys` carry its values (keyset_predicate)."""
+    pub_key = keys[1]  # relevance: [rank, published_at]
+    rows: list = []
+    if cursor_id is None or pub_key["value"] >= window:
+        inside = query.where(Video.published_at >= window).order_by(*order_by(keys))
+        if cursor_id is not None:
+            inside = inside.where(keyset_predicate(keys, cursor_id))
+        rows = list(db.execute(inside.limit(n)).all())
+    if len(rows) < n:
+        newest = [pub_key]  # past the window every rank is 0: the order is newest first
+        rest = query.where(Video.published_at < window).order_by(*order_by(newest))
+        if cursor_id is not None:
+            rest = rest.where(keyset_predicate(newest, cursor_id))
+        rows += db.execute(rest.limit(n - len(rows))).all()
+    return rows
+
+
 def order_by(keys: list[dict]):
     cols = []
     for k in keys:
@@ -587,26 +653,45 @@ def keyset_predicate(keys: list[dict], cursor_id: str):
     return rec(0)
 
 
-def encode_cursor(keys: list[dict], row) -> str:
+def encode_cursor(keys: list[dict], row, window: datetime | None = None) -> str:
+    """`window` is the relevance window's start: the cursor carries it, so every page ranks against the
+    window page one did. Recomputed per page, a video ingested mid-scroll would push the window's oldest
+    row to rank 0, behind the cursor, and the reader would see it twice."""
     vals = []
     for i, k in enumerate(keys):
         v = getattr(row, f"_k{i}")
         vals.append(v.isoformat() if (v is not None and k["kind"] == "dt") else v)
-    payload = json.dumps({"k": vals, "id": row.id})
+    body: dict = {"k": vals, "id": row.id}
+    if window is not None:
+        body["w"] = window.isoformat()
+    payload = json.dumps(body)
     return base64.urlsafe_b64encode(payload.encode()).decode()
 
 
-def decode_cursor(keys: list[dict], cursor: str) -> tuple[list, str]:
+def _cursor_dt(v):
+    """A cursor's timestamp: aware or refused. The server writes only aware ones (timestamptz), and
+    windowed_relevance_rows compares them in Python, where a naive one beside an aware one raises — a
+    500 for a hand-made cursor instead of the 400 (E31 review 2)."""
+    if v is None:
+        return None
+    dt = datetime.fromisoformat(v)
+    if dt.tzinfo is None:
+        raise ValueError("naive cursor timestamp")
+    return dt
+
+
+def decode_cursor(keys: list[dict], cursor: str) -> tuple[list, str, datetime | None]:
     try:
         payload = json.loads(base64.urlsafe_b64decode(cursor.encode()))
         raw = payload["k"]
         if len(raw) != len(keys):
             raise ValueError("key arity mismatch")
-        vals = [
+        vals = [_cursor_dt(v) if k["kind"] == "dt" else v for k, v in zip(keys, raw, strict=True)]
-            datetime.fromisoformat(v) if (v is not None and k["kind"] == "dt") else v
+        for k, v in zip(keys, vals, strict=True):
-            for k, v in zip(keys, raw)
+            if v is None and not k["nullable"]:
-        ]
+                raise ValueError("null value for a non-nullable key")
-        return vals, payload["id"]
+        window = payload.get("w")
+        return vals, payload["id"], _cursor_dt(window) if window else None
     except (binascii.Error, json.JSONDecodeError, KeyError, ValueError, TypeError) as exc:
         # Framework-free: the route handler translates this into an HTTP 400 so the
         # read-model stays web-layer-agnostic (see routes/feed.get_feed).

backend/app/issues/__init__.py

@@ -0,0 +1,3 @@
+"""Bug reports built from debug sessions (E16): the files an issue carries, and where they go —
+downloaded as a zip, or (S4) sent to a Gitea issue tracker. The session itself — what is recorded
+and tagged — is the core module `app/debugsession.py`."""

backend/app/issues/bundle.py

@@ -0,0 +1,79 @@
+"""What an issue from a debug session carries, as files — one definition for the zip export and the
+Gitea upload, so the two cannot drift apart.
+
+  * `issue.md` — the title and the body the user wrote and reviewed in the composer (rendered there,
+    in their language; the server does not rewrite it);
+  * `client.log` / `server.log` — the session's evidence (`debugsession.bundle`) minus the rows the
+    user took out, one JSON object per line, browser rows and server rows apart. `.log`, not
+    `.jsonl`: Gitea refuses that extension (measured, plans/e16 S0);
+  * the screenshots, in the order they were taken.
+
+Every row is already redacted twice (at capture, and again when stored), and nothing here adds a
+field the review list did not show.
+"""
+
+import io
+import json
+import zipfile
+from dataclasses import dataclass
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+from app import debugsession, logsink
+from app.models import DebugScreenshot, DebugSession
+
+
+@dataclass(frozen=True)
+class IssueFile:
+    name: str
+    data: bytes
+    mime: str
+
+
+def title_of(session: DebugSession) -> str:
+    # Not a sentence of ours: the composer always saves a title in the user's language, and a
+    # session exported before it ever saved is named by its date alone, in no language at all.
+    title = ((session.draft or {}).get("title") or "").strip()
+    return title or f"{session.started_at:%Y-%m-%d %H:%M} · {session.id}"
+
+
+def body_of(session: DebugSession) -> str:
+    return ((session.draft or {}).get("body") or "").strip()
+
+
+def _lines(rows) -> bytes:
+    return "".join(
+        json.dumps(logsink.serialize(r), ensure_ascii=False, default=str) + "\n" for r in rows
+    ).encode()
+
+
+def files(db: Session, session: DebugSession) -> list[IssueFile]:
+    """The session's issue as files: the text, the two logs, the screenshots."""
+    rows, _ = debugsession.bundle(db, session, drop_excluded=True)
+    text = f"# {title_of(session)}\n\n{body_of(session)}\n"
+    out = [
+        IssueFile("issue.md", text.encode(), "text/markdown"),
+        IssueFile("client.log", _lines(r for r in rows if r.source == "web"), "text/plain"),
+        IssueFile("server.log", _lines(r for r in rows if r.source != "web"), "text/plain"),
+    ]
+    shots = db.scalars(
+        select(DebugScreenshot)
+        .where(DebugScreenshot.session_id == session.id)
+        .order_by(DebugScreenshot.id)
+    )
+    for n, shot in enumerate(shots, start=1):
+        ext = debugsession.IMAGE_TYPES[shot.mime]
+        name = f"{debugsession.SCREENSHOT_PREFIX}{n}.{ext}"
+        out.append(IssueFile(name, shot.data, shot.mime))
+    return out
+
+
+def zipped(items: list[IssueFile]) -> bytes:
+    buf = io.BytesIO()
+    with zipfile.ZipFile(buf, "w", zipfile.ZIP_DEFLATED) as z:
+        for item in items:
+            # The images are compressed already: deflating them again costs time and wins nothing.
+            method = zipfile.ZIP_STORED if item.mime.startswith("image/") else zipfile.ZIP_DEFLATED
+            z.writestr(item.name, item.data, compress_type=method)
+    return buf.getvalue()

backend/app/issues/gitea.py

@@ -0,0 +1,288 @@
+"""The Gitea issue tracker a debug session's issue goes to (E16 S4) — measured against Gitea 1.27.1
+before a line of this was written (plans/e16-debug-issue.md, S0).
+
+Configured by an admin (sysconfig `issues_gitea_url` / `_repo` / `_token`, the token encrypted at
+rest). The repository must be PRIVATE: an issue carries its reporter's log — page titles, filter
+values, at Teljes (TRACE) what they typed into search — and a public repository would publish it. That is
+checked on every send, not only in the admin's "test connection", because a repository can be made
+public after the fact.
+
+What an issue is made of comes from `bundle.files` (the same files the zip export carries). Sending
+is three kinds of call, measured:
+  * `POST /repos/{repo}/issues` — the title and the body the user wrote;
+  * `POST /repos/{repo}/issues/{n}/assets?name=…` — ONE file per request (field `attachment`);
+  * `PATCH /repos/{repo}/issues/{n}` — the body again, with the screenshots shown inline once their
+    links exist.
+A private repository's attachment answers 404 to a browser not signed in to Gitea, which is why the
+composer's own preview shows our copies, never these links.
+
+Delivery keeps its progress in the session's draft (`delivery`: the issue number, the files still to
+upload, the links of those uploaded, whether the body shows the screenshots yet), so a failure
+half-way keeps the issue that exists and a retry does only what is missing — never a second issue.
+Every step is checkpointed (the caller's `db.commit`), so what did happen is kept whatever fails
+next — and the first checkpoint comes before the first call, with the tracker's settings read ABOVE
+it, so neither a row lock nor a read snapshot nor a pooled connection is held across the calls. The
+caller claims the session for the send first (`debugsession.claim_send`) and passes the claim's
+token in, which is what keeps a second send, an autosave or a screenshot out of the way; the claim
+is renewed by every progress report, and a report whose claim is gone raises rather than writing
+over the send that took it. `debugsession.not_uploaded` / `images_pending` read that progress.
+"""
+
+from collections.abc import Callable, Iterator
+from contextlib import contextmanager
+from typing import Any, Literal
+
+import httpx
+from sqlalchemy.orm import Session
+
+from app import debugsession, logs, sysconfig
+from app.issues import bundle
+from app.models import DebugSession
+
+log = logs.get_logger(__name__)
+
+TIMEOUT_SECONDS = 20.0
+
+Reason = Literal[
+    "not_configured",  # url, repository or token missing
+    "unreachable",  # no answer, or a 5xx
+    "bad_token",  # 401/403: the token is wrong or lacks write:issue / read:repository
+    "repo_not_found",  # 404 on the repository
+    "public_repo",  # the repository is public — an issue would publish the reporter's log
+    "issues_disabled",  # the repository has its issue tracker switched off
+    "rejected",  # any other 4xx (a file type the instance does not allow, …), or an answer that is
+    # not the API's (a redirect — an http:// address behind https —, a page that is not JSON)
+]
+
+
+class GiteaError(Exception):
+    def __init__(self, reason: Reason) -> None:
+        super().__init__(reason)
+        self.reason: Reason = reason
+
+
+def _settings(db: Session) -> tuple[str, str, str]:
+    return (
+        sysconfig.get_str(db, "issues_gitea_url").strip().rstrip("/"),
+        sysconfig.get_str(db, "issues_gitea_repo").strip().strip("/"),
+        sysconfig.get_str(db, "issues_gitea_token").strip(),
+    )
+
+
+def target(db: Session) -> dict | None:
+    """Where a report goes, for the composer to say so — the repository — or None when no tracker is
+    configured (the report can then only be downloaded)."""
+    url, repo, token = _settings(db)
+    return {"repo": repo} if url and repo and token else None
+
+
+@contextmanager
+def _client(settings: tuple[str, str, str]) -> Iterator[tuple[httpx.Client, str]]:
+    # Takes the settings already read, not the session: reading them HERE opened a transaction one
+    # line after `deliver` had ended one, and held it across the calls below (deep-review round 4).
+    url, repo, token = settings
+    if not (url and repo and token):
+        raise GiteaError("not_configured")
+    with httpx.Client(
+        base_url=f"{url}/api/v1",
+        headers={"Authorization": f"token {token}", "Accept": "application/json"},
+        timeout=TIMEOUT_SECONDS,
+    ) as client:
+        yield client, repo
+
+
+def _call(client: httpx.Client, method: str, path: str, **kwargs) -> httpx.Response:
+    try:
+        r = client.request(method, path, **kwargs)
+    except httpx.HTTPError as exc:
+        log.warning(
+            "the issue tracker did not answer",
+            extra=logs.event("issues.gitea.unreachable", path=path, error=type(exc).__name__),
+        )
+        raise GiteaError("unreachable") from None
+    if r.status_code < 300:
+        return r
+    # The status only: Gitea's message can quote back what was sent — the user's own title.
+    log.warning(
+        "the issue tracker refused a call",
+        extra=logs.event("issues.gitea.refused", path=path, status=r.status_code),
+    )
+    if r.status_code in (401, 403):
+        raise GiteaError("bad_token")
+    if r.status_code == 404:
+        raise GiteaError("repo_not_found")
+    if r.status_code >= 500:
+        raise GiteaError("unreachable")
+    raise GiteaError("rejected")
+
+
+def _json(r: httpx.Response, *keys: str) -> dict[str, Any]:
+    """The answer as the API's JSON, holding `keys` — anything else is refused, not a crash, so a
+    step after the issue was created cannot take down the request that holds its record.
+
+    The CREATE call is the one place this cannot keep that record: its answer IS the record. An
+    answer it cannot read there leaves the session unsent, and a retry opens another issue — the
+    price of not being able to tell a refused create from an unreadable one."""
+    try:
+        data = r.json()
+    # silent-ok: not JSON is one of the answers refused (and logged) just below.
+    except ValueError:
+        data = None
+    if not isinstance(data, dict) or any(k not in data for k in keys):
+        log.warning(
+            "the issue tracker answered something that is not its API",
+            extra=logs.event("issues.gitea.unexpected", status=r.status_code),
+        )
+        raise GiteaError("rejected")
+    return data
+
+
+def _private_repo(client: httpx.Client, repo: str) -> dict:
+    data = _json(_call(client, "GET", f"/repos/{repo}"))
+    if not data.get("private"):
+        raise GiteaError("public_repo")
+    if not data.get("has_issues"):
+        raise GiteaError("issues_disabled")
+    return data
+
+
+def check(db: Session) -> dict:
+    """The admin's "test connection": the repository answers, is private and takes issues."""
+    with _client(_settings(db)) as (client, repo):
+        data = _private_repo(client, repo)
+    return {"repo": data.get("full_name") or repo, "url": data.get("html_url")}
+
+
+def preview(db: Session, text: str) -> str:
+    """The body as Gitea itself will render it (its markdown flavour, its sanitiser).
+
+    **No `Context`, deliberately — a repository context here does not fail the call, it kills the
+    tracker.** With one, Gitea's post-processor turns every hex-shaped word into a commit link, and
+    on the API's render path the commit checker it asks is nil: the whole Gitea process segfaults
+    and restarts, taking the git server (and, here, the container registry) down with it. Every
+    report carries hex — a tab id, a request id — so this was not an edge case but the only case.
+    Measured on throwaway instances (2026-09-20): `{"Mode": "gfm", "Context": "<repo>"}` with
+    `5645b1c0ae01` in the text crashes 1.26.4 and 1.27.1 and is fixed in 1.27.3 — but a self-hoster
+    runs whatever version they run, so the payload stays safe here rather than relying on theirs.
+    Dropping `Context` is not a downgrade of the preview: measured against the same body, the HTML
+    keeps its tables, its task lists and its highlighted code blocks. It only stops resolving
+    repo-relative links (a hash, a `#123`), which the issue page itself still resolves.
+    """
+    with _client(_settings(db)) as (client, _repo):
+        r = _call(client, "POST", "/markdown", json={"Text": text, "Mode": "gfm"})
+    return r.text
+
+
+def deliver(
+    db: Session,
+    session: DebugSession,
+    *,
+    token: str,
+    checkpoint: Callable[[], None],
+    before_create: Callable[[], None] | None = None,
+    on_created: Callable[[], None] | None = None,
+) -> str:
+    """Send the session's issue, or finish sending it; returns its URL.
+
+    `token` is the caller's claim on this session (`debugsession.claim_send`). Every progress report
+    renews it and checks the row still holds it, so a send that outlived its lease stops here rather
+    than writing over the one that took over (`debugsession.SendClaimLost`).
+
+    `checkpoint` persists what has happened so far and is called after every step (the caller's
+    commit — this module does not decide when a transaction ends). `before_create` runs right before
+    the tracker is asked to open a NEW issue, after the checks that can refuse it without asking,
+    and may raise to stop it; `on_created` runs right after one was opened, in time to be committed
+    by the SAME checkpoint that first writes `issue_url` — anything the caller must not be able to
+    lose separately from the link belongs there.
+    """
+    state = debugsession.delivery(session)
+    files = bundle.files(db, session)
+    attachments = [f for f in files if f.name != "issue.md"]
+    settings = _settings(db)
+    # Reading the bundle and the tracker's settings opened a transaction; end it before the first
+    # call, so nothing holds a snapshot (and an idle connection) across minutes of tracker I/O.
+    # The settings are read HERE and handed to `_client`, not read inside it: reading them there
+    # was a SELECT one line after this commit, which reopened the transaction and held it — idle,
+    # with its snapshot pinned — across the repository check and the create (deep-review round 4
+    # measured it in pg_stat_activity). A comment saying otherwise sat right here.
+    checkpoint()
+    with _client(settings) as (client, repo):
+        _private_repo(client, repo)
+        if "number" not in state:
+            if before_create is not None:
+                before_create()
+            created = _json(
+                _call(
+                    client,
+                    "POST",
+                    f"/repos/{repo}/issues",
+                    json={"title": bundle.title_of(session), "body": bundle.body_of(session)},
+                ),
+                "number",
+                "html_url",
+            )
+            # Built ON the state read at the start so nothing already in it is dropped. The claim
+            # is NOT among those things, whatever this comment used to say: `set_delivery` strips
+            # any `sending` passed to it and stamps the row's own from `token` (round 4 — removing
+            # the spread left the whole suite green, because the claim never travelled here).
+            state = {
+                **state,
+                "number": created["number"],
+                "pending": [f.name for f in attachments],
+                "links": {},
+            }
+            session.issue_url = created["html_url"]
+            try:
+                debugsession.set_delivery(db, session, state, token)
+            except debugsession.SendClaimLost:
+                # The one step where losing the claim costs something irreversible: the issue is
+                # already open on the tracker and this is what would have recorded its link. We
+                # still must not write — another send owns the row and is opening its own issue —
+                # so say which issue was left behind, because nothing else will.
+                # TRY400: the traceback is one SendClaimLost carrying the session id and
+                # nothing else; the issue number below is the whole news.
+                log.error(  # noqa: TRY400
+                    "an issue was created by a send that had already lost its claim",
+                    extra=logs.event(
+                        "issues.gitea.orphaned",
+                        number=created["number"],
+                        session_id=session.id,
+                    ),
+                )
+                raise
+            if on_created is not None:
+                on_created()
+            checkpoint()
+            log.info(
+                "issue created",
+                extra=logs.event("issues.gitea.created", number=created["number"]),
+            )
+        number = state["number"]
+        for item in attachments:
+            if item.name not in state["pending"]:
+                continue
+            asset = _json(
+                _call(
+                    client,
+                    "POST",
+                    f"/repos/{repo}/issues/{number}/assets",
+                    params={"name": item.name},
+                    files={"attachment": (item.name, item.data, item.mime)},
+                ),
+                "browser_download_url",
+            )
+            state = {
+                **state,
+                "pending": [n for n in state["pending"] if n != item.name],
+                "links": {**state["links"], item.name: asset["browser_download_url"]},
+            }
+            debugsession.set_delivery(db, session, state, token)
+            checkpoint()
+        shots = [n for n in state["links"] if n.startswith(debugsession.SCREENSHOT_PREFIX)]
+        if shots and not state.get("images_shown"):
+            images = "\n".join(f"![{n}]({state['links'][n]})" for n in sorted(shots))
+            body = f"{bundle.body_of(session)}\n\n---\n\n{images}"
+            _call(client, "PATCH", f"/repos/{repo}/issues/{number}", json={"body": body})
+            debugsession.set_delivery(db, session, {**state, "images_shown": True}, token)
+            checkpoint()
+    return session.issue_url or ""

backend/app/logs.py

@@ -0,0 +1,838 @@
+"""Structured logging for every Siftlode process — the ONE place logging is configured.
+
+Before this module a single deployment printed FOUR different log formats: uvicorn's
+`log_config.json` for the API, a hand-rolled StreamHandler in `main.py` for everything that did not
+start under uvicorn (tests, scripts), `logging.basicConfig()` in the worker and in the download
+child (no timestamp at all), and alembic's `fileConfig`. None of them carried a correlation id, and
+uvicorn's access log printed the query string of every request (`?account=1`, `?source=feed`) — so
+the logs were both unjoinable and leaky. `configure_logging()` replaces all four.
+
+Event shape — the OpenTelemetry log data model, without the SDK; one JSON object per line:
+    ts · level · sev · source · logger · event · msg · attrs
+plus whatever correlation the context carries (`request_id`, `user_id`, `debug_session_id`,
+`job_id`) and the app version. That shape is what E22.6's admin console filters on and what a Gitea
+issue attachment will carry, so it is deliberately flat and stable.
+
+Call sites stay ordinary stdlib logging:
+
+    log = logs.get_logger(__name__)  # app/downloads/pipeline.py -> siftlode.downloads.pipeline
+    log.info("job finished", extra=logs.event("download.job.done", job_id=7, bytes=n))
+
+`event()` is optional — a bare `log.info("…")` still comes out well-formed — but the dotted code is
+what makes a line queryable later, so add one wherever the line matters.
+
+REDACTION IS APPLIED AT THE HANDLER, not at the call site: every message, every attr value and
+every traceback passes `redact()` on the way out. That is the whole point — a secret cannot be
+leaked by forgetting to redact it at one of ~125 call sites, and a new call site is safe by
+default. Never logged, at any level: passwords, tokens (session, reset/verify, setup, share,
+OAuth), cookie blobs, secret config values, and email addresses.
+"""
+
+import contextlib
+import contextvars
+import json
+import logging
+import logging.config
+import re
+import sys
+import threading
+import time
+import traceback
+import uuid
+from collections.abc import Callable
+from datetime import datetime, timezone
+from typing import Any
+from urllib.parse import parse_qsl
+
+from app.config import settings
+from app.ratelimit import RateLimiter, client_key
+
+# One level below DEBUG for the "Teljes / TRACE" diagnostics level (search strings, redacted request
+# bodies). Registered on import so `logging.getLogger(...).log(logs.TRACE, …)` and a `LOG_LEVEL=TRACE`
+# both work in every process, including one that never calls configure_logging (a test).
+TRACE = 5
+logging.addLevelName(TRACE, "TRACE")
+
+# OpenTelemetry severity numbers for our level names. Kept beside the level because the number is
+# what a log backend sorts and filters on; the name is for humans reading a terminal.
+SEVERITY: dict[str, int] = {
+    "TRACE": 1,
+    "DEBUG": 5,
+    "INFO": 9,
+    "WARNING": 13,
+    "ERROR": 17,
+    "CRITICAL": 21,
+}
+
+# Which process this is. Set by configure_logging; every line carries it so one Loki/DB query can
+# separate the API's view of a job from the worker's.
+_source = "api"
+
+
+# --- correlation context -------------------------------------------------------------------
+# contextvars, like quota's actor attribution (app/quota.py): set at an entry point, read by the
+# formatter, never threaded through call signatures. A child task copies the context at creation,
+# so anything set before `await call_next(...)` is visible to the whole request.
+_request_id: contextvars.ContextVar[str | None] = contextvars.ContextVar(
+    "log_request_id", default=None
+)
+_user_id: contextvars.ContextVar[int | None] = contextvars.ContextVar("log_user_id", default=None)
+_debug_session: contextvars.ContextVar[str | None] = contextvars.ContextVar(
+    "log_debug_session", default=None
+)
+_job_id: contextvars.ContextVar[int | None] = contextvars.ContextVar("log_job_id", default=None)
+
+_CONTEXT_VARS = {
+    "request_id": _request_id,
+    "user_id": _user_id,
+    "debug_session_id": _debug_session,
+    "job_id": _job_id,
+}
+
+
+class _RequestScope:
+    """What the auth dependency learns about a request, for every line logged after it.
+
+    A MUTABLE object behind a contextvar, not plain contextvars, because of how a request actually
+    runs: FastAPI resolves a sync dependency (`auth.current_user`) in a threadpool on a COPY of the
+    context, so a `ContextVar.set()` there is gone by the time the handler runs — measured: that is
+    how the account never reached a handler's own lines. `log_requests` installs one scope per
+    request; every copied context shares the same object, so what the dependency writes is seen by
+    the handler and by the request line itself."""
+
+    __slots__ = ("user_id", "level", "debug_session_id")
+
+    def __init__(self) -> None:
+        self.user_id: int | None = None
+        # The user's consented diagnostics level as a logging number; None at the base level.
+        self.level: int | None = None
+        # The user's own open debug session this request named (E16), else None.
+        self.debug_session_id: str | None = None
+
+
+_scope: contextvars.ContextVar[_RequestScope | None] = contextvars.ContextVar(
+    "log_request_scope", default=None
+)
+
+
+def set_user(
+    user_id: int | None,
+    level: str = "INFO",
+    until: datetime | None = None,
+    debug_session: str | None = None,
+) -> None:
+    """Attribute the rest of this request's log lines to `user_id`, and log them at the user's
+    consented diagnostics `level` until `until` (E22 S5; "INFO" = no elevation). Called from the
+    auth dependency — the one place the session is actually resolved into an account — rather than
+    from the request middleware, which runs OUTSIDE the session middleware.
+
+    `debug_session` tags every line with the user's open debug session (E16). The caller has
+    already checked it is theirs and open; a session's level arrives here as `level`/`until`.
+
+    The root logger's floor is tracked per elevation: a request without the session's header (an
+    `<img>`, a `<video>` — the browser sends those itself) carries only the user's own level, and
+    must not lower the floor a concurrent request of the session still needs."""
+    _user_id.set(user_id)
+    scope = _scope.get()
+    if scope is None:  # outside log_requests (a script, a unit test): this context only
+        scope = _RequestScope()
+        _scope.set(scope)
+    elevated = level != "INFO" and until is not None and until > datetime.now(timezone.utc)
+    scope.user_id = user_id
+    scope.debug_session_id = debug_session
+    scope.level = resolve_level(level) if elevated else None
+    if user_id is not None:
+        key = user_id if debug_session is None else (user_id, debug_session)
+        _track_elevation(key, scope.level, until)
+
+
+def set_job(job_id: int | None) -> None:
+    """Attribute this process's log lines to a download job — for the download CHILD, whose whole
+    life is one job (the worker uses `bind()` per job instead, since it juggles several)."""
+    _job_id.set(job_id)
+
+
+@contextlib.contextmanager
+def bind(**fields: Any):
+    """Bind correlation fields for the duration of a block (a worker job, a scheduler run):
+
+        with logs.bind(job_id=job.id):
+            ...
+
+    Unknown field names are ignored rather than raising — a caller binding a field this module does
+    not know about is a typo, not a reason to take down a download."""
+    tokens = [(_CONTEXT_VARS[k].set(v)) for k, v in fields.items() if k in _CONTEXT_VARS]
+    keys = [k for k in fields if k in _CONTEXT_VARS]
+    try:
+        yield
+    finally:
+        for key, token in zip(keys, tokens, strict=True):
+            _CONTEXT_VARS[key].reset(token)
+
+
+def event(code: str, /, **attrs: Any) -> dict[str, Any]:
+    """`extra=` payload for a log call: a stable dotted event code plus structured attributes.
+
+    Returns a plain dict so it composes with stdlib logging and costs nothing when the level is off
+    (the dict is built either way; keep the attrs cheap — no DB reads inside the call).
+
+    `code` is positional-only so that no attribute name is reserved: `event("x", code=...)` means
+    the attribute, not this parameter. Without the `/` it was a TypeError, and `code` is exactly
+    the word a caller reaches for when the thing that failed has one of its own."""
+    return {"event": code, "attrs": attrs}
+
+
+# --- redaction -----------------------------------------------------------------------------
+# Substring matching, not exact keys: the same secret travels as `password`, `current_password`,
+# `new_password`, `smtp_password`, `plex_server_token`, `access_token`, … Matching on the PART is
+# what makes a key we have never seen safe by default.
+_SECRET_KEY_PARTS = (
+    "password",
+    "passphrase",
+    "token",
+    "secret",
+    "cookie",
+    "authorization",
+    "auth_header",
+    "ciphertext",
+    "private_key",
+    "api_key",
+    "apikey",
+    "credential",
+    "session_key",
+)
+
+# Query/fragment parameters that carry a credential in a URL we log (OAuth `code`/`state`, our own
+# `?token=` setup and share links, `#verify=` / `#reset=` fragments).
+_SECRET_PARAM_NAMES = (
+    "access_token",
+    "refresh_token",
+    "id_token",
+    "token",
+    "code",
+    "state",
+    "key",
+    "secret",
+    "password",
+    "pwd",
+    "sig",
+    "signature",
+    "verify",
+    "reset",
+    "setup",
+)
+_SECRET_PARAM_RE = re.compile(
+    r"\b((?:" + "|".join(_SECRET_PARAM_NAMES) + r")=)([^&\s\"']+)",
+    re.IGNORECASE,
+)
+
+# Deliberately loose: over-matching redacts a non-address that looks like one (harmless), while
+# under-matching puts a real address in a log we attach to an issue (not harmless).
+_EMAIL_RE = re.compile(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}")
+
+REDACTED = "[redacted]"
+# `extra={PERSIST: False}` keeps a line out of the `log_events` table while stdout still gets it —
+# how `log_requests` caps anonymous traffic per client (app/logsink.py reads it).
+PERSIST = "persist"
+REDACTED_EMAIL = "[email]"
+
+
+def redact_text(text: str) -> str:
+    """Scrub a free-text string: credential-bearing URL params and email addresses."""
+    scrubbed = _SECRET_PARAM_RE.sub(lambda m: f"{m.group(1)}{REDACTED}", text)
+    return _EMAIL_RE.sub(REDACTED_EMAIL, scrubbed)
+
+
+def redact_query(query: str) -> str:
+    """Scrub a raw query string. DECODED first: the text rules match what was typed, and a browser
+    sends it percent-encoded — `someone%40example.com` is an address `_EMAIL_RE` cannot see. Each
+    parameter also goes through the key rules, so `?new_password=` goes as `redact` would drop it."""
+    pairs = parse_qsl(query, keep_blank_values=True)
+    return "&".join(f"{key}={_redact_param(key, value)}" for key, value in pairs)
+
+
+def _redact_param(key: str, value: str) -> str:
+    """One decoded `key=value` pair. The parameter-NAME rule is applied to the key here rather than
+    left to `_SECRET_PARAM_RE` over the rebuilt pair: that pattern was written for the raw,
+    percent-encoded query, where a value holds no literal space, `&` or quote, and its value class
+    stops at the first one. On decoded text `?state=a%20b` would keep its tail and `?code=%20abc`
+    would not match at all — the whole credential straight into the line."""
+    if key.lower() in _SECRET_PARAM_NAMES:
+        return REDACTED
+    return str(redact(value, key=key))
+
+
+def redact(value: Any, *, key: str = "") -> Any:
+    """Recursively scrub a structure destined for the log.
+
+    A value whose KEY names a secret is dropped whole (we never log a prefix of a token — a prefix
+    is still a leak and still useless). Everything else is scrubbed as text. Depth is bounded by
+    the structures we actually log (attrs dicts), and non-JSON types fall through to `str`."""
+    lowered = key.lower()
+    if key and any(part in lowered for part in _SECRET_KEY_PARTS):
+        return REDACTED
+    if isinstance(value, dict):
+        return {k: redact(v, key=str(k)) for k, v in value.items()}
+    if isinstance(value, (list, tuple)):
+        return [redact(v, key=key) for v in value]
+    if isinstance(value, str):
+        return redact_text(value)
+    return value
+
+
+# --- formatters ----------------------------------------------------------------------------
+_SCRUBBED_STACK = "scrubbed_stack"
+
+
+def scrubbed_stack(exc: BaseException, scrub: Callable[[BaseException], str]) -> dict[str, str]:
+    """`extra=` fragment: `exc`'s traceback, rendered where an `exc_info` stack goes, with the
+    message of every exception in it put through `scrub`.
+
+    For a failure whose message can name someone's data (a download's URL, a file path, a video
+    id). The frames name code and are kept whole; the text at the foot of each link is the data.
+    A plain `exc_info` renders the chain too — the `__cause__`/`__context__` a DB error raised
+    inside `except DownloadError` carries — so scrubbing only the line's own message is not
+    enough. Merge it into the call's `extra`: `{**event(...), **scrubbed_stack(...)}`."""
+    links: list[BaseException] = []
+    link: BaseException | None = exc
+    while link is not None and all(link is not seen for seen in links):
+        links.append(link)
+        link = link.__cause__ or (None if link.__suppress_context__ else link.__context__)
+    parts: list[str] = []
+    for i in range(len(links) - 1, -1, -1):  # the oldest first, as Python prints a chain
+        link = links[i]
+        kind = type(link)
+        name = kind.__qualname__
+        if kind.__module__ not in ("builtins", "__main__"):
+            name = f"{kind.__module__}.{name}"
+        try:
+            message = scrub(link)
+        # silent-ok: the message is what could not be rendered, and the placeholder says so — the
+        # caller is in the middle of reporting one failure and must not get a second from here.
+        except Exception:  # noqa: BLE001 — a `__str__` that raises, of any shape
+            message = "<unprintable>"
+        frames = "".join(traceback.format_tb(link.__traceback__))
+        head = f"Traceback (most recent call last):\n{frames}" if frames else ""
+        parts.append(f"{head}{name}: {message}" if message else f"{head}{name}")
+        if i:
+            parts.append(
+                "\nThe above exception was the direct cause of the following exception:\n"
+                if links[i - 1].__cause__ is link
+                else "\nDuring handling of the above exception, another exception occurred:\n"
+            )
+    return {_SCRUBBED_STACK: "\n".join(parts)}
+
+
+# Attributes every LogRecord carries; anything else a caller passed via `extra=` is a structured
+# attribute and is merged into `attrs` (so `extra={"job_id": 7}` works as well as `event(...)`).
+_RECORD_STANDARD = frozenset(
+    vars(logging.LogRecord("", 0, "", 0, "", None, None)).keys()
+    | {"message", "asctime", "taskName", "event", "attrs", PERSIST, "elevated", _SCRUBBED_STACK}
+)
+
+
+class ContextFilter(logging.Filter):
+    """Copies the correlation contextvars onto the record. A Filter (not the formatter) so the
+    values are captured at emit time, before a QueueHandler can move the record to another
+    thread — the DB sink in S3 depends on that."""
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        for name, var in _CONTEXT_VARS.items():
+            value = var.get()
+            if value is not None and not hasattr(record, name):
+                setattr(record, name, value)
+        scope = _scope.get()
+        if scope is not None and scope.user_id is not None and not hasattr(record, "user_id"):
+            record.user_id = scope.user_id
+        if (
+            scope is not None
+            and scope.debug_session_id is not None
+            and not hasattr(record, "debug_session_id")
+        ):
+            record.debug_session_id = scope.debug_session_id
+        return True
+
+
+class LevelGate(logging.Filter):
+    """The configured level, enforced at the HANDLER, with one exception: the lines of a request
+    whose user raised their diagnostics level (E22 S5).
+
+    The root logger's level cannot do this alone: it is one number for the whole process, and the
+    point is to log ONE user's requests in detail while everyone else's stay at the base level. So
+    the root is lowered to the lowest level any active elevation needs (`_track_elevation`), and
+    this gate drops what only an elevation let through unless the line belongs to that elevated
+    request. It also marks such a line `elevated`, which is what makes the `log_events` sink keep
+    every line of that request, not just its warnings."""
+
+    def filter(self, record: logging.LogRecord) -> bool:
+        scope = _scope.get()
+        if scope is not None and scope.level is not None:
+            record.elevated = True
+            return record.levelno >= min(scope.level, _base_level)
+        return record.levelno >= _base_level
+
+
+def _record_attrs(record: logging.LogRecord) -> dict[str, Any]:
+    extra = {k: v for k, v in vars(record).items() if k not in _RECORD_STANDARD}
+    attrs = dict(getattr(record, "attrs", None) or {})
+    attrs.update({k: v for k, v in extra.items() if k not in _CONTEXT_VARS})
+    return attrs
+
+
+def _safe_message(record: logging.LogRecord) -> str:
+    """`record.getMessage()` raises if the args don't match the format string. A broken log call
+    must not take down the caller, so the failure is reported IN the line instead."""
+    try:
+        return record.getMessage()
+    # silent-ok: the failure is reported IN the returned line. Logging it would re-enter this same
+    # formatter with a record built out of the one that just failed.
+    except Exception as exc:  # noqa: BLE001 — any formatting error, reported not raised
+        return f"{record.msg!r} (log formatting failed: {exc})"
+
+
+def record_fields(record: logging.LogRecord) -> dict[str, Any]:
+    """The event as structured fields, REDACTED — the one definition both the stdout line and the
+    `log_events` row are built from, so the two sinks can never disagree about what is safe to keep.
+    Absent optional fields are left out; `ts` is an aware datetime."""
+    fields: dict[str, Any] = {
+        "ts": datetime.fromtimestamp(record.created, timezone.utc),
+        "level": record.levelname,
+        "sev": SEVERITY.get(record.levelname, 9),
+        "source": _source,
+        "logger": record.name,
+        "msg": redact_text(_safe_message(record)),
+    }
+    code = getattr(record, "event", None)
+    if code:
+        fields["event"] = code
+    for name in _CONTEXT_VARS:
+        value = getattr(record, name, None)
+        if value is not None:
+            fields[name] = value
+    attrs = _record_attrs(record)
+    if attrs:
+        fields["attrs"] = redact(attrs)
+    if record.exc_info:
+        fields["exc"] = redact_text(logging.Formatter().formatException(record.exc_info))
+    elif getattr(record, _SCRUBBED_STACK, None):
+        fields["exc"] = redact_text(getattr(record, _SCRUBBED_STACK))
+    elif record.exc_text:
+        # A QueueHandler-style hand-off formats the traceback into exc_text and drops exc_info.
+        fields["exc"] = redact_text(record.exc_text)
+    fields["version"] = settings.app_version
+    return fields
+
+
+class JsonFormatter(logging.Formatter):
+    """One JSON object per line — the shape Loki and the `log_events` table both ingest."""
+
+    def format(self, record: logging.LogRecord) -> str:
+        payload = record_fields(record)
+        payload["ts"] = payload["ts"].isoformat(timespec="milliseconds")
+        return json.dumps(payload, ensure_ascii=False, default=str)
+
+
+class TextFormatter(logging.Formatter):
+    """Human-readable single line for a terminal (`LOG_FORMAT=text`): the same content, ordered for
+    reading. Redaction applies here too — a dev log is still a log someone pastes into an issue."""
+
+    def format(self, record: logging.LogRecord) -> str:
+        stamp = datetime.fromtimestamp(record.created, timezone.utc).strftime("%H:%M:%S.%f")[:-3]
+        head = f"{stamp} {record.levelname:<5} [{record.name}]"
+        rid = getattr(record, "request_id", None)
+        if rid:
+            head += f" {rid}"
+        code = getattr(record, "event", None)
+        line = f"{head} {code + ' — ' if code else ''}{redact_text(_safe_message(record))}"
+        attrs = _record_attrs(record)
+        if attrs:
+            line += f" {json.dumps(redact(attrs), ensure_ascii=False, default=str)}"
+        if record.exc_info:
+            line += "\n" + redact_text(self.formatException(record.exc_info))
+        elif getattr(record, _SCRUBBED_STACK, None):
+            line += "\n" + redact_text(getattr(record, _SCRUBBED_STACK))
+        return line
+
+
+# --- configuration -------------------------------------------------------------------------
+# Third-party loggers that are noisy at our own level. Pinned explicitly so the diagnostics knob can
+# lower the ROOT level to DEBUG/TRACE without drowning the output in SQL statements and HTTP chatter
+# — raising THESE is a deliberate, separate act (E22.6's console).
+_THIRD_PARTY_LEVELS = {
+    "sqlalchemy": "WARNING",
+    "sqlalchemy.engine": "WARNING",
+    "apscheduler": "WARNING",
+    "urllib3": "WARNING",
+    "httpx": "WARNING",
+    "httpcore": "WARNING",
+    "asyncio": "WARNING",
+    "PIL": "WARNING",
+    "alembic": "INFO",
+    # uvicorn's own startup/shutdown/WebSocket lines are worth keeping; its ACCESS log is not —
+    # it is one line per request with the query string in it, replaced by `log_requests` below.
+    "uvicorn": "INFO",
+    "uvicorn.error": "INFO",
+    "uvicorn.access": "CRITICAL",
+}
+
+
+def get_logger(name: str) -> logging.Logger:
+    """The logger for one module: `logs.get_logger(__name__)` in `app/plex/sync.py` is
+    `siftlode.plex.sync`.
+
+    The names were literals before, and 25 of them covered 34 modules — six different Plex modules
+    all called themselves `siftlode.plex`, so neither a level pinned per logger nor E22.6's console
+    filter could tell a sync line from a stream line. Derived from `__name__` there is exactly one
+    name per module, it cannot drift from the file it sits in, and `test_logger_names.py` fails if a
+    module hand-rolls one again.
+
+    `__main__` is resolved to the module it really is. The worker runs as `python -m app.worker`
+    and spawns `python -m app.downloads.runner`, so in the two long-running processes `__name__`
+    is `__main__` — taken at its word, both would log as the bare root. `-m` records the real name
+    on the module's spec; a plain script has none and stays the root."""
+    if name == "__main__":
+        spec = getattr(sys.modules.get("__main__"), "__spec__", None)
+        name = getattr(spec, "name", None) or name
+    stem = name.removeprefix("app.") if name.startswith("app.") else ""
+    # logger-name-ok: this call IS the derivation
+    return logging.getLogger(f"siftlode.{stem}" if stem else "siftlode")
+
+
+def resolve_level(name: str) -> int:
+    """Level name → number. Both ways in (the `LOG_LEVEL` env and the `log_level` admin knob) are
+    already held to the `LogLevel` choices, so an unknown name here is a programming error."""
+    value = logging.getLevelName(name)
+    if not isinstance(value, int):
+        raise ValueError(f"unknown log level {name!r}")
+    return value
+
+
+def configure_logging(source: str) -> None:
+    """Configure logging for THIS process. Call once, first thing, in every entry point:
+    `main.py` (api), `worker.main`, `downloads/runner.main`, `alembic/env.py`.
+
+    `source` names the process in every line. The level starts at `LOG_LEVEL` (the bootstrap default);
+    the runtime admin knob goes through `set_level()` instead, so changing it needs no restart.
+
+    Our loggers deliberately keep `propagate=True` and own no handler: everything reaches the single
+    root StreamHandler. That is what keeps `caplog` working in tests, and it means a new logger name
+    is configured the moment it is created."""
+    global _source
+    _source = source
+    formatter = settings.log_format
+    logging.config.dictConfig(
+        {
+            "version": 1,
+            # Never disable loggers created at import time (module-level `getLogger` calls run
+            # before this function does).
+            "disable_existing_loggers": False,
+            "formatters": {
+                "json": {"()": f"{__name__}.JsonFormatter"},
+                "text": {"()": f"{__name__}.TextFormatter"},
+            },
+            "filters": {
+                "context": {"()": f"{__name__}.ContextFilter"},
+                "gate": {"()": f"{__name__}.LevelGate"},
+            },
+            "handlers": {
+                "stdout": {
+                    "class": "logging.StreamHandler",
+                    "stream": sys.stdout,
+                    "formatter": formatter,
+                    "filters": ["context", "gate"],
+                }
+            },
+            "root": {"level": resolve_level(settings.log_level), "handlers": ["stdout"]},
+            "loggers": {name: {"level": lvl} for name, lvl in _THIRD_PARTY_LEVELS.items()},
+        }
+    )
+    set_level(settings.log_level)
+
+
+def set_level(name: str) -> None:
+    """Change the running process's level (the `log_level` admin knob). Third-party loggers keep
+    their pinned levels, so turning on DEBUG shows OUR detail, not SQLAlchemy's."""
+    global _base_level
+    _base_level = resolve_level(name)
+    _apply_root_level(force=True)
+
+
+def current_level() -> str:
+    """The configured level — not the root logger's, which an active elevation may have lowered."""
+    return logging.getLevelName(_base_level or logging.getLogger().getEffectiveLevel())
+
+
+# --- per-user elevation (E22 S5) -----------------------------------------------------------
+# The base level (`LOG_LEVEL`, then the admin knob). NOTSET until configure_logging runs — a unit
+# test, a script — so the gate lets everything through there, as it did before the gate existed.
+_base_level = logging.NOTSET
+# → (level number, expiry) for every elevation this process has seen and not yet expired: a user's
+# own, by user id, and a debug session's (E16), by (user id, session id) — its requests alone carry
+# its level, so it is tracked apart from the user's own requests.
+_ElevationKey = int | tuple[int, str]
+_elevations: dict[_ElevationKey, tuple[int, datetime]] = {}
+_elevations_lock = threading.Lock()
+_root_applied: int | None = None
+
+
+def _track_elevation(key: _ElevationKey, level: int | None, until: datetime | None) -> None:
+    with _elevations_lock:
+        if level is None or until is None:
+            _elevations.pop(key, None)
+        else:
+            _elevations[key] = (level, until)
+        active = bool(_elevations)
+    if active or _root_applied != _base_level:
+        _apply_root_level()
+
+
+def _apply_root_level(force: bool = False) -> None:
+    """Root level = the lowest of the base and every unexpired elevation (expired ones are pruned
+    here). `setLevel` clears every logger's level cache, so it runs only when the number changes."""
+    global _root_applied
+    if _base_level == logging.NOTSET:
+        return  # not configured: the root keeps whatever the process set up
+    now = datetime.now(timezone.utc)
+    # The level is SET under the lock too, not just computed: two threads that each computed a
+    # different answer would otherwise apply them in either order, and the older one could win —
+    # the root back at the base while an elevation is still active.
+    with _elevations_lock:
+        for key in [k for k, (_, until) in _elevations.items() if until <= now]:
+            del _elevations[key]
+        lowest = min([_base_level, *(lvl for lvl, _ in _elevations.values())])
+        if force or lowest != _root_applied:
+            logging.getLogger().setLevel(lowest)
+            _root_applied = lowest
+
+
+# --- erased accounts ------------------------------------------------------------------------
+# `auth.purge_user` deletes an account's `log_events` rows, but a line about that account can still
+# be on its way: the erasing request's own access line is logged AFTER the handler committed, a queued
+# row waits for the sink's writer, a background task logs later. Those rows are stored without the
+# account (`logsink` asks here). Remembered for an hour in this process — a line later than that has
+# no request left to come from, and the retention job is the backstop for any other process.
+ERASED_MEMORY_SECONDS = 3600.0
+_erased: dict[int, float] = {}
+_erased_lock = threading.Lock()
+# Held by whoever checks rows against `is_erased` AND writes them (`erasure_guard`). `forget_user`
+# takes it too, so an erasure waits for a write that already passed the check to commit — and the
+# purge's DELETE, which runs after `forget_user`, then removes those rows as well. Without it a row
+# checked a moment before the erasure could be inserted a moment after the DELETE.
+_erasure_guard = threading.Lock()
+
+
+def erasure_guard() -> threading.Lock:
+    """Hold this around "drop erased accounts from these rows, insert them, commit"."""
+    return _erasure_guard
+
+
+def forget_user(user_id: int) -> None:
+    """Mark `user_id` erased: no line logged about it from now on is stored under its id. Call BEFORE
+    deleting its rows (see `_erasure_guard`)."""
+    now = time.monotonic()
+    with _erasure_guard, _erased_lock:
+        for stale in [u for u, at in _erased.items() if now - at > ERASED_MEMORY_SECONDS]:
+            del _erased[stale]
+        _erased[user_id] = now
+    with _elevations_lock:
+        for key in [
+            k for k in _elevations if k == user_id or (isinstance(k, tuple) and k[0] == user_id)
+        ]:
+            del _elevations[key]
+    # That pop can have removed the lowest elevation, and the root level is the lowest of the base
+    # and every unexpired one. Recomputed here rather than left to whichever later request happens
+    # to call `_track_elevation` — the invariant `_apply_root_level` documents holds at every pop.
+    _apply_root_level()
+
+
+def end_session_elevation(session_id: str) -> None:
+    """A debug session was stopped: its elevation ends now, not at its cap."""
+    with _elevations_lock:
+        for key in [k for k in _elevations if isinstance(k, tuple) and k[1] == session_id]:
+            del _elevations[key]
+    _apply_root_level()
+
+
+def is_erased(user_id: int | None) -> bool:
+    if user_id is None:
+        return False
+    with _erased_lock:
+        at = _erased.get(user_id)
+    return at is not None and time.monotonic() - at <= ERASED_MEMORY_SECONDS
+
+
+def current_source() -> str:
+    """The process name `configure_logging` was given (api, worker, runner, alembic)."""
+    return _source
+
+
+# --- request logging -----------------------------------------------------------------------
+# A client may supply its own request id so a browser log line and a server line join up (E22 S4).
+# Accepted only in this shape: it ends up in a log field, and an unbounded header would be a way to
+# write arbitrary text into the log.
+_REQUEST_ID_RE = re.compile(r"^[A-Za-z0-9._-]{8,64}$")
+
+# Paths whose success is logged at DEBUG rather than INFO: the container healthcheck runs every 15s
+# (5.7k lines/day saying nothing), hashed static assets are served by the hundred per page load, and
+# the admin log console reads the stored log — its live tail every few seconds, so at INFO the log
+# would fill with the console watching itself. A FAILURE on any of them still logs at WARNING/ERROR
+# — this lowers the floor, it does not hide.
+_QUIET_PREFIXES = ("/healthz", "/assets/", "/static/", "/api/admin/logs")
+
+# The reads the frontend POLLS (a `useLiveQuery`, the `me` heartbeat, the sync status, the subtitle
+# progress), by route template, GET only — logged at DEBUG like the prefixes above. One open tab asked
+# for the two unread counts, `/api/me` and the download index every 30 s, and at INFO those were ~95%
+# of the stored log (measured on prod 2026-09-19): 11k lines a day per tab burying every line someone
+# caused. The same route answered with a failure still logs at WARNING/ERROR.
+_POLLED_ROUTES = frozenset(
+    {
+        "/api/me",
+        "/api/me/notifications/unread_count",
+        "/api/messages/unread_count",
+        "/api/downloads",
+        "/api/downloads/index",
+        "/api/downloads/usage",
+        "/api/sync/my-status",
+        "/api/admin/scheduler",
+        "/api/admin/downloads",
+        "/api/admin/downloads/storage",
+        "/api/plex/subtitle/{rating_key}/{sub_ord}/status",
+    }
+)
+
+# Not `siftlode.logs`: this is the replacement for uvicorn's ACCESS log, and `siftlode.http` is
+# what an operator pins a level on when they want request lines and nothing else. The module it
+# happens to live in is not what the name is about.
+_request_log = logging.getLogger("siftlode.http")  # logger-name-ok: named for the traffic
+
+
+async def log_requests(request, call_next):
+    """The single access-log line, replacing uvicorn's.
+
+    What changes versus `uvicorn.access`: the ROUTE TEMPLATE instead of the concrete path (so
+    `/api/videos/{video_id}` groups), never the query string (it carried `?account=`, and at TRACE
+    the redacted query belongs in attrs, not in the URL), a request id echoed back in
+    `X-Request-Id` (with the template in `X-Route`), the duration, and the account — read from `request.state`, which the auth
+    dependency fills, because this middleware sits outside both the session middleware and the
+    downstream task's context."""
+    incoming = request.headers.get("x-request-id", "")
+    rid = incoming if _REQUEST_ID_RE.match(incoming) else uuid.uuid4().hex[:16]
+    token = _request_id.set(rid)
+    scope = _RequestScope()
+    scope_token = _scope.set(scope)
+    request.state.request_id = rid
+    started = time.perf_counter()
+    try:
+        response = await call_next(request)
+    except Exception:
+        # The unhandled-exception path: FastAPI's ServerErrorMiddleware re-raises after its own
+        # 500, so without this the ONE request that most needs a log line is the one with no
+        # access-log entry at all.
+        _request_log.exception(
+            "request failed",
+            extra={
+                **event(
+                    "http.request",
+                    method=request.method,
+                    path=_route_template(request),
+                    duration_ms=round((time.perf_counter() - started) * 1000, 1),
+                ),
+                **_request_owner(request, logging.ERROR),
+            },
+        )
+        _scope.reset(scope_token)
+        _request_id.reset(token)
+        raise
+    duration_ms = round((time.perf_counter() - started) * 1000, 1)
+    status = response.status_code
+    quiet = request.url.path.startswith(_QUIET_PREFIXES) or (
+        request.method == "GET" and _route_template(request) in _POLLED_ROUTES
+    )
+    if status >= 500:
+        level = logging.ERROR
+    elif status >= 400:
+        level = logging.WARNING
+    else:
+        level = logging.DEBUG if quiet else logging.INFO
+    attrs: dict[str, Any] = {
+        "method": request.method,
+        "path": _route_template(request),
+        "status": status,
+        "duration_ms": duration_ms,
+    }
+    # Teljes (TRACE) the user consented to: the query string too — redacted, and in attrs only, never
+    # in the message.
+    if request.url.query and _traced(scope):
+        attrs["query"] = redact_query(request.url.query)
+    _request_log.log(
+        level,
+        "%s %s %s",
+        request.method,
+        _route_template(request),
+        status,
+        extra={**event("http.request", **attrs), **_request_owner(request, level)},
+    )
+    response.headers["X-Request-Id"] = rid
+    # The browser logs its own copy of this line (frontend lib/http.ts) and must pick the same level:
+    # it learns "this one is quiet" here instead of keeping a second copy of the polled-route list.
+    if level == logging.DEBUG:
+        response.headers["X-Log-Quiet"] = "1"
+    # The template, so the browser's copy of this line groups the same way (frontend lib/http.ts).
+    # Only a MATCHED route's: a template never carries a path parameter's value (a share token stays
+    # `{token}`), while the raw-path fallback of an unmatched request could — the browser then
+    # templates the path itself.
+    matched = _matched_route(request)
+    if matched:
+        response.headers["X-Route"] = matched
+    _scope.reset(scope_token)
+    _request_id.reset(token)
+    return response
+
+
+def consented_trace() -> bool:
+    """Whether the current request's OWN user turned Teljes (TRACE) on — the gate for a line that
+    carries what they typed or what was quoted back to them. The level alone is not enough: an
+    admin running the whole process at TRACE would pass it for every user, and a scheduler thread,
+    which has no request, for none of them in particular. Guard such a line with this, then log it
+    at TRACE."""
+    scope = _scope.get()
+    return scope is not None and _traced(scope)
+
+
+def _traced(scope: _RequestScope) -> bool:
+    """Whether this request's query string is logged: only under its OWN user's Teljes consent. An
+    admin setting the whole process to TRACE gets more detail, not every user's search terms — the
+    privacy policy promises those are kept only for a user who agreed to it."""
+    return scope.level is not None and scope.level <= TRACE
+
+
+# Anonymous request lines kept per client per minute. Signed-in traffic is not capped: it is a known
+# account and exactly what an admin needs to see. Anonymous traffic is where the volume can come from
+# nobody in particular — a scanner walking `/wp-login.php`, a bot on a share link — and each of its
+# 404s is a WARNING that would otherwise land in `log_events` without limit. The client key is an
+# in-memory counter key only; no address is ever written to a log line or a row.
+ANONYMOUS_LINES_PER_MINUTE = 60
+_anonymous_lines = RateLimiter(ANONYMOUS_LINES_PER_MINUTE, 60.0)
+
+
+def _request_owner(request, level: int) -> dict[str, Any]:
+    """`extra=` fields naming who the request belongs to, and whether its line may be stored. The
+    account comes from `request.state` (filled by the auth dependency), so it lands in the line's
+    top-level `user_id` — the correlation field — rather than in `attrs`."""
+    user_id = getattr(request.state, "user_id", None)
+    if user_id is not None:
+        return {"user_id": user_id}
+    if level < logging.INFO:
+        return {}  # a quiet line is never stored anyway — don't spend the client's allowance on it
+    return {} if _anonymous_lines.allow(client_key(request)) else {PERSIST: False}
+
+
+def _matched_route(request) -> str | None:
+    """The matched route's template (`/api/videos/{video_id}`), or None when nothing matched."""
+    return getattr(request.scope.get("route"), "path", None)
+
+
+def _route_template(request) -> str:
+    """The matched route's template, else the raw path for a 404 or a static file. Never the query
+    string."""
+    return _matched_route(request) or request.url.path

backend/app/logsink.py

@@ -0,0 +1,367 @@
+"""The `log_events` sink (E22 S3): the stored copy of the log the diagnostics console reads.
+
+stdout stays the complete log (docker keeps 3×50 MB of it); this table keeps the part worth querying
+for `log_retention_days`:
+
+  * every WARNING and above,
+  * the `http.request` line of every request at INFO or above (not the DEBUG-level health checks and
+    static files),
+  * every line of a request whose user raised their diagnostics level (E22 S5 — `logs.LevelGate`
+    marks them), which is the detail they consented to have kept,
+  * nothing a caller marked `extra={logs.PERSIST: False}` — how the request log caps anonymous
+    traffic per client.
+
+What the table must never do is slow down or break the code that logs. So a line is redacted and
+turned into a row on the CALLER's thread (the same `logs.record_fields` the stdout line is built
+from), put on a bounded queue without blocking, and written in batches by one background thread
+on its own connection — never inside a request's transaction. A full queue drops the row and
+counts it; a failed write drops the batch and says so on stderr. Both are reported rather than
+retried: a log that retries into a sick database is how a small outage becomes a big one.
+
+Flood control is the second half. On 2026-09-16 a YouTube RSS outage made prod log 11k near-identical
+warnings in four hours. stdout keeps them all; the table keeps `REPEATS_PER_MINUTE` of each message
+template per minute and one `log.repeats_suppressed` row saying how many more there were.
+
+Core layer: imports `logs`, `db`, `models`, `config` — nothing above them.
+"""
+
+import atexit
+import json
+import logging
+import queue
+import re
+import sys
+import threading
+import time
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+from sqlalchemy import delete, insert
+from sqlalchemy.engine import Engine
+from sqlalchemy.orm import Session
+
+from app import logs
+from app.config import settings
+from app.db import engine
+from app.models import LogEvent
+from app.utils import iso
+
+QUEUE_MAX = 10_000
+BATCH_MAX = 500
+FLUSH_SECONDS = 1.0
+REPEATS_PER_MINUTE = 30
+REPEAT_WINDOW_SECONDS = 60.0
+# Distinct templates tracked per window. Past this many, a new template is stored without counting
+# — the limiter must not become the unbounded thing it exists to prevent.
+_REPEAT_KEYS_MAX = 2_000
+
+# Column widths from models.LogEvent. A value past its column would fail the WHOLE batch it rides in.
+_LIMITS = {
+    "level": 10,
+    "source": 16,
+    "logger": 128,
+    "event": 64,
+    "request_id": 64,
+    "debug_session_id": 64,
+    "version": 64,
+    "msg": 8_000,
+    "exc": 32_000,
+}
+
+_log = logs.get_logger(__name__)
+
+
+def should_store(record: logging.LogRecord) -> bool:
+    """Whether this record belongs in `log_events` (see the module docstring)."""
+    if getattr(record, logs.PERSIST, True) is False:
+        return False
+    if record.levelno >= logging.WARNING or getattr(record, "elevated", False):
+        return True
+    return getattr(record, "event", None) == "http.request" and record.levelno >= logging.INFO
+
+
+def _without_nul(value: Any) -> Any:
+    """`value` with every NUL character replaced, in keys and nested values alike. Postgres refuses
+    NUL in text and `\\u0000` in JSONB, and a refused row fails the WHOLE batch it rides in — every
+    user's rows with it. Anything logged can carry one (`GET /assets/%00` logs the decoded path)."""
+    if isinstance(value, str):
+        return value.replace("\x00", "�")
+    if isinstance(value, dict):
+        return {_without_nul(k): _without_nul(v) for k, v in value.items()}
+    if isinstance(value, list):
+        return [_without_nul(v) for v in value]
+    return value
+
+
+def to_row(fields: dict[str, Any]) -> dict[str, Any]:
+    """`logs.record_fields` output → a `log_events` row: known columns only, text clipped to its
+    column and free of NUL, attrs made JSON-safe (anything exotic becomes its `str`, as on stdout)."""
+    row = {
+        column: fields.get(column) for column in LogEvent.__table__.columns.keys() if column != "id"
+    }
+    for column, limit in _LIMITS.items():
+        value = row.get(column)
+        if isinstance(value, str):
+            value = _without_nul(value)
+            row[column] = value[: limit - 1] + "…" if len(value) > limit else value
+    if row["attrs"] is not None:
+        row["attrs"] = _without_nul(
+            json.loads(json.dumps(row["attrs"], ensure_ascii=False, default=str))
+        )
+    return row
+
+
+def without_erased(rows: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """The rows, with the account taken off any whose account was erased meanwhile
+    (`logs.forget_user`). Checked right before the INSERT, under `logs.erasure_guard()` held until
+    the commit: a row built before the erasure can still be waiting for its batch."""
+    for row in rows:
+        if logs.is_erased(row.get("user_id")):
+            row["user_id"] = None
+    return rows
+
+
+def _notice(code: str, msg: str, **attrs: Any) -> dict[str, Any]:
+    """A row the sink writes about ITSELF. Built directly, not logged: logging it would route it back
+    through this sink (and the writer thread's own records are deliberately ignored)."""
+    return to_row(
+        {
+            "ts": datetime.now(timezone.utc),
+            "level": "WARNING",
+            "sev": logs.SEVERITY["WARNING"],
+            "source": logs.current_source(),
+            "logger": _log.name,
+            "event": code,
+            "msg": msg,
+            "attrs": attrs,
+            "version": settings.app_version,
+        }
+    )
+
+
+_DIGITS = re.compile(r"\d+")
+
+
+def _template(record: logging.LogRecord) -> str:
+    """The flood limiter's grouping key: the message template. Except when the template is a bare
+    `%s` — a pass-through for someone else's text (yt-dlp's lines, `formats._YdlLog`), where the
+    template says nothing and every distinct line would share one bucket. Then the text itself
+    groups, with its numbers folded, so `Retrying fragment 5 (1/10)` and `(2/10)` are still one
+    repeat and a different warning is not crowded out by them."""
+    msg = str(record.msg)
+    if msg != "%s" or not record.args:
+        return msg
+    return _DIGITS.sub("#", record.getMessage())[:200]
+
+
+class _Repeats:
+    """Per-template admission for one minute at a time. Used by the writer thread only (unlocked)."""
+
+    def __init__(self, per_window: int, window_seconds: float):
+        self.per_window = per_window
+        self.window = window_seconds
+        self.started = time.monotonic()
+        self.seen: dict[tuple[str, str, int], int] = {}
+
+    def admit(self, key: tuple[str, str, int] | None) -> bool:
+        if key is None:
+            return True
+        count = self.seen.get(key)
+        if count is None and len(self.seen) >= _REPEAT_KEYS_MAX:
+            return True
+        self.seen[key] = (count or 0) + 1
+        return self.seen[key] <= self.per_window
+
+    def roll(self, now: float, *, final: bool = False) -> list[dict[str, Any]]:
+        """Close the window if it is over — or, `final`, because the process is: one notice per
+        template that went past the limit. Without `final` a process that lives under a minute
+        (every download child) would drop its repeats and never say so."""
+        if not final and now - self.started < self.window:
+            return []
+        notices = [
+            _notice(
+                "log.repeats_suppressed",
+                f"{count - self.per_window} more like this were not stored",
+                logger=logger,
+                template=logs.redact_text(template),
+                suppressed=count - self.per_window,
+                window_seconds=int(self.window),
+            )
+            for (logger, template, _level), count in self.seen.items()
+            if count > self.per_window
+        ]
+        self.seen.clear()
+        self.started = now
+        return notices
+
+
+class LogSink(logging.Handler):
+    """The handler plus its writer. `start()` runs the writer thread; a test drives `write_pending()`
+    directly instead, on its own thread, with no timing involved."""
+
+    def __init__(self, bind: Engine = engine):
+        super().__init__(level=logging.NOTSET)
+        # Correlation fields are contextvars: capture them now, on the thread that logged.
+        self.addFilter(logs.ContextFilter())
+        self._bind = bind
+        self._queue: queue.Queue[tuple[tuple[str, str, int] | None, dict[str, Any]]] = queue.Queue(
+            QUEUE_MAX
+        )
+        self._dropped = 0
+        self._dropped_lock = threading.Lock()
+        self._repeats = _Repeats(REPEATS_PER_MINUTE, REPEAT_WINDOW_SECONDS)
+        self._stopping = threading.Event()
+        self._writer: threading.Thread | None = None
+
+    # --- caller side ---------------------------------------------------------------------------
+    def emit(self, record: logging.LogRecord) -> None:
+        if threading.current_thread() is self._writer:
+            return  # the writer's own trouble (a driver warning mid-insert) must not feed itself
+        if not should_store(record):
+            return
+        try:
+            row = to_row(logs.record_fields(record))
+        # silent-ok: `handleError` is the report — logging's own out-of-band channel to stderr.
+        # A log call here would be handed to this handler, which is the one that just failed.
+        except Exception:  # noqa: BLE001 — a record we cannot turn into a row is reported, not raised
+            self.handleError(record)
+            return
+        # Request lines are one per request by nature; only the repeated SAME message is a flood.
+        key = (
+            None
+            if row["event"] == "http.request"
+            else (record.name, _template(record), record.levelno)
+        )
+        try:
+            self._queue.put_nowait((key, row))
+        # silent-ok: counted, not logged — the count goes out as the `log.sink.dropped` notice row
+        # the writer emits, which is the one report that does not feed this queue again.
+        except queue.Full:
+            with self._dropped_lock:
+                self._dropped += 1
+
+    # --- writer side ---------------------------------------------------------------------------
+    def write_pending(self, wait: float = 0.0, *, final: bool = False) -> int:
+        """Write one batch: whatever is queued (up to BATCH_MAX), waiting up to `wait` seconds for the
+        first item. Returns the number of rows written, notices included. `final` closes the
+        repeat window early — the last batch before the process exits."""
+        rows: list[dict[str, Any]] = []
+        try:
+            item = self._queue.get(timeout=wait) if wait > 0 else self._queue.get_nowait()
+        # silent-ok: an idle sink. Nothing failed; there was simply nothing waiting.
+        except queue.Empty:
+            item = None
+        while item is not None:
+            key, row = item
+            if self._repeats.admit(key):
+                rows.append(row)
+            if len(rows) >= BATCH_MAX:
+                break
+            try:
+                item = self._queue.get_nowait()
+            # silent-ok: the batch ends here because the queue ran dry, which is the normal end.
+            except queue.Empty:
+                item = None
+        rows.extend(self._repeats.roll(time.monotonic(), final=final))
+        with self._dropped_lock:
+            dropped, self._dropped = self._dropped, 0
+        if dropped:
+            rows.append(
+                _notice("log.sink.dropped", f"{dropped} log events were dropped", dropped=dropped)
+            )
+        if not rows:
+            return 0
+        try:
+            # The connection is checked out BEFORE the guard, and every request path takes them in
+            # that order too (its session already holds a connection when it reaches the guard). The
+            # other way round, this thread would hold a process-global lock while waiting up to
+            # `pool_timeout` for a connection, and the requests that could return one are the ones
+            # blocked on the lock. The guard still spans the check, the insert and the commit.
+            with self._bind.connect() as conn, logs.erasure_guard(), conn.begin():
+                conn.execute(insert(LogEvent), without_erased(rows))
+        except Exception as exc:  # noqa: BLE001 — the database is down or refused the batch
+            # silent-ok / noqa T201: the sink's OWN failure. A log call here would be handed to
+            # this sink, which is the thing that just failed — straight into a loop. stderr is
+            # the one way out that does not re-enter.
+            print(  # noqa: T201
+                f"log sink: could not store {len(rows)} log events: "
+                f"{logs.redact_text(str(exc))[:300]}",
+                file=sys.stderr,
+            )
+            return 0
+        return len(rows)
+
+    def _run(self) -> None:
+        while not self._stopping.is_set():
+            self.write_pending(wait=FLUSH_SECONDS)
+        while self.write_pending():  # drain what arrived before the stop
+            pass
+        self.write_pending(final=True)  # and the repeat notices of a window cut short
+
+    def start(self) -> None:
+        self._writer = threading.Thread(target=self._run, name="log-sink", daemon=True)
+        self._writer.start()
+
+    def stop(self, timeout: float = 5.0) -> None:
+        self._stopping.set()
+        if self._writer is not None:
+            self._writer.join(timeout)
+
+
+_sink: LogSink | None = None
+
+
+def start() -> None:
+    """Attach the sink to the root logger and start its writer. Call after `logs.configure_logging`
+    (which replaces the root handlers) and once the schema exists. A no-op when `LOG_DB_SINK` is off
+    or it is already running."""
+    global _sink
+    if _sink is not None or not settings.log_db_sink:
+        return
+    _sink = LogSink()
+    # The same level gate as stdout: the root may be lowered for one user's elevated requests, and
+    # everyone else's detail must not reach the table through that gap either.
+    _sink.addFilter(logs.LevelGate())
+    logging.getLogger().addHandler(_sink)
+    _sink.start()
+    # A short-lived process (the download child) exits without a lifespan: flush on the way out.
+    atexit.register(stop)
+
+
+def stop() -> None:
+    """Detach the sink and write what is still queued."""
+    global _sink
+    if _sink is None:
+        return
+    logging.getLogger().removeHandler(_sink)
+    _sink.stop()
+    _sink = None
+
+
+def gc(db: Session, retention_days: int) -> int:
+    """Stage the deletion of rows older than `retention_days` (at least one day — stored logs always
+    expire). The caller commits. Returns the number of rows removed."""
+    cutoff = datetime.now(timezone.utc) - timedelta(days=max(1, retention_days))
+    result = db.execute(delete(LogEvent).where(LogEvent.ts < cutoff))
+    return result.rowcount or 0
+
+
+def serialize(row: LogEvent) -> dict:
+    """A stored row as the API sends it — one shape for every reader: the admin console (E22.6) and
+    a debug session's evidence list (E16)."""
+    return {
+        "id": row.id,
+        "ts": iso(row.ts),
+        "level": row.level,
+        "source": row.source,
+        "logger": row.logger,
+        "event": row.event,
+        "msg": row.msg,
+        "attrs": row.attrs,
+        "exc": row.exc,
+        "request_id": row.request_id,
+        "user_id": row.user_id,
+        "job_id": row.job_id,
+        "debug_session_id": row.debug_session_id,
+        "version": row.version,
+    }

backend/app/main.py

@@ -1,43 +1,44 @@
 import asyncio
-import logging
 import math
-import sys
+import mimetypes
 from contextlib import asynccontextmanager, suppress
 from pathlib import Path
 
 from fastapi import FastAPI, HTTPException, Request
 from fastapi.encoders import jsonable_encoder
 from fastapi.exceptions import RequestValidationError
-
-# When started via uvicorn --log-config the "siftlode" logger is already configured
-# (see log_config.json). This block is a timestamped fallback for other entrypoints
-# (tests, scripts) so our logs are never silently dropped.
-_siftlode_logger = logging.getLogger("siftlode")
-if not _siftlode_logger.handlers:
-    _handler = logging.StreamHandler(sys.stdout)
-    _handler.setFormatter(logging.Formatter("%(asctime)s %(levelname)-5s [%(name)s] %(message)s"))
-    _siftlode_logger.addHandler(_handler)
-    _siftlode_logger.setLevel(logging.INFO)
-    _siftlode_logger.propagate = False
-
-log = logging.getLogger("siftlode.app")
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import FileResponse, HTMLResponse, JSONResponse
 from fastapi.staticfiles import StaticFiles
 from sqlalchemy.exc import DataError
 from starlette.middleware.sessions import SessionMiddleware
 
-from app import state
+# Logging is configured HERE, before the app's own modules are imported, so a line one of them logs
-from app.config import settings
+# WHILE importing already has the format (a logger is configured whenever it is created —
-from app.db import SessionLocal
+# app/logs.py — but a line emitted before `configure_logging` goes out raw). ONE format for the api,
-from app.plex import stream as plex_stream
+# the worker, the download child and alembic alike — this replaces the old uvicorn --log-config +
-from app.plex import subtitles as plex_subtitles
+# the hand-rolled fallback that used to live here, and uvicorn's access log is replaced by the
-from app.routes import (
+# `http.request` middleware below. The imports after it are E402 on purpose, each one marked.
+from app import logs
+from app.logs import configure_logging, current_level, log_requests, set_level
+
+configure_logging("api")
+
+log = logs.get_logger(__name__)
+from app import bodylimit, collation, logsink, state, sysconfig  # noqa: E402 — logging is configured first
+from app.config import settings  # noqa: E402 — logging is configured first
+from app.db import SessionLocal, engine  # noqa: E402 — logging is configured first
+from app.plex import stream as plex_stream  # noqa: E402 — logging is configured first
+from app.plex import subtitles as plex_subtitles  # noqa: E402 — logging is configured first
+from app.routes import (  # noqa: E402 — logging is configured first
     admin,
+    admin_logs as admin_logs_routes,
     audit as audit_routes,
     auth as auth_routes,
     channels,
     config as config_routes,
+    debug_sessions as debug_sessions_routes,
+    diagnostics as diagnostics_routes,
     discover as discover_routes,
     downloads,
     feed,
@@ -57,22 +58,34 @@ from app.routes import (
     tags,
     version,
 )
-from app.scheduler import shutdown_scheduler, start_scheduler
+from app.scheduler import shutdown_scheduler, start_scheduler  # noqa: E402 — logging is configured first
 
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     log.info("Siftlode starting up")
+    # The stored copy of the log (log_events). The entrypoint has already migrated the schema.
+    logsink.start()
     # First-run: if the instance isn't configured yet, mint a fresh one-time setup token and print
     # the wizard URL so the operator can open it (the app runs in setup mode until they finish).
     with SessionLocal() as db:
+        # The admin's runtime level (sysconfig `log_level`) wins over the LOG_LEVEL bootstrap
+        # default, so a level chosen in the UI survives a restart without touching the environment.
+        # Best-effort on purpose: a LOGGING knob must never be the thing that stops the app from
+        # booting — an unreachable or not-yet-migrated database has to start into the env level.
+        try:
+            set_level(sysconfig.get_str(db, "log_level"))
+        except Exception:  # any DB/config failure keeps the bootstrap level
+            # With the traceback: this catch spans a connection, a query and a coercion at boot,
+            # and `%s` of the exception does not say which of them raised.
+            log.exception("could not read the stored log level; staying at %s", current_level())
         if not state.is_configured(db):
             token = state.rotate_setup_token(db)
             url = f"{settings.app_base}/setup?token={token}"
             log.warning(
                 "FIRST-RUN SETUP REQUIRED — this instance isn't configured yet.\n"
                 "    Open the install wizard (internal access only):\n    %s",
-                url,
+                url,  # log-pii-ok: this instance's OWN setup URL, and `?token=` is redacted
             )
     # No HLS session survives a restart (they live in this process), so anything still on disk is
     # a crashed run's segments — sweep them or they accumulate forever. FIRE AND FORGET, not
@@ -83,6 +96,9 @@ async def lifespan(app: FastAPI):
     # owns, so racing it against early playback is safe.
     sweep = asyncio.create_task(_sweep_hls_orphans())
     reaper = asyncio.create_task(_reap_hls_sessions())
+    # Fire and forget as well: a rebuild after an ICU upgrade takes seconds, and the title sorts serve
+    # meanwhile (CONCURRENTLY) in the old index's order, as they did before the restart.
+    heal = asyncio.create_task(_heal_collation_drift())
     start_scheduler()
     try:
         yield
@@ -92,7 +108,7 @@ async def lifespan(app: FastAPI):
         # pending ("Task was destroyed but it is pending!" — noise that reads like a bug during
         # triage). Cancelling does NOT interrupt the rmtree thread itself; this only makes our own
         # bookkeeping honest.
-        for task in (sweep, reaper):
+        for task in (sweep, reaper, heal):
             task.cancel()
             with suppress(asyncio.CancelledError):
                 await task
@@ -100,6 +116,7 @@ async def lifespan(app: FastAPI):
         # Kills a running subtitle extraction: its worker would otherwise hold interpreter exit for
         # the rest of the read (see plex_subtitles.shutdown).
         plex_subtitles.shutdown()
+        logsink.stop()  # last: write what the shutdown itself just logged
 
 
 # How often idle HLS sessions are reaped (see plex/stream.py `reap_idle`).
@@ -124,6 +141,18 @@ async def _reap_hls_sessions() -> None:
             log.exception("HLS session reap failed")
 
 
+async def _heal_collation_drift() -> None:
+    """Background startup check of the ICU collation the title-sort indexes are built in (see
+    `collation.heal_drift`). A failure is logged, never fatal: the sorts still answer, only possibly in
+    the old order."""
+    try:
+        await asyncio.to_thread(collation.heal_drift, engine)
+    except asyncio.CancelledError:
+        raise
+    except Exception:
+        log.exception("ICU collation drift check failed")
+
+
 async def _sweep_hls_orphans() -> None:
     """Background half of the startup HLS sweep (see lifespan). Owns its own errors: a failure here
     must never take the app down, and a cancelled shutdown is not one."""
@@ -175,15 +204,6 @@ app.add_middleware(
     https_only=settings.session_https_only,  # Secure flag when served over HTTPS (prod)
 )
 
-if settings.frontend_origin:
-    app.add_middleware(
-        CORSMiddleware,
-        allow_origins=[settings.frontend_origin],
-        allow_credentials=True,
-        allow_methods=["*"],
-        allow_headers=["*"],
-    )
-
 # Paths reachable while the instance is unconfigured (setup mode). Everything else under /api or
 # /auth is locked until the wizard finishes; static/SPA always loads so the wizard page can render.
 _SETUP_OPEN_PREFIXES = ("/api/setup", "/api/version")
@@ -203,6 +223,34 @@ async def setup_gate(request, call_next):
     return await call_next(request)
 
 
+# SEC-20 — the global request-body cap (app/bodylimit.py). FastAPI reads a declared body IN FULL
+# before the first dependency, so without this the anonymous login route is a free memory/time sink.
+#
+# WHERE it sits is the whole design, and it is pinned by `test_body_limit.py` rather than by this
+# comment, because `add_middleware` inverts the reading order (last registered = OUTERMOST) and a
+# stack described in prose drifts. Registered here it lands directly INSIDE CORS and directly
+# OUTSIDE the setup gate, which is the narrowest position that satisfies both constraints:
+#
+#   * inside CORS and the security headers, so the 413 is decorated like every other response. A
+#     refusal the browser drops for a missing Access-Control-Allow-Origin is no better than a hang,
+#     and `capped_body`'s 413 one layer down IS decorated — two shapes for one fact is exactly what
+#     bodylimit.py exists to avoid.
+#   * outside the setup gate, the session and every auth dependency under the router — none of which
+#     run for a body that is already too large. The setup gate matters specifically: on an instance
+#     that is not yet configured it opens a DB session per request, which is precisely the anonymous
+#     pre-auth work SEC-20 exists to remove.
+app.add_middleware(bodylimit.BodyLimitMiddleware)
+
+if settings.frontend_origin:
+    app.add_middleware(
+        CORSMiddleware,
+        allow_origins=[settings.frontend_origin],
+        allow_credentials=True,
+        allow_methods=["*"],
+        allow_headers=["*"],
+    )
+
+
 @app.middleware("http")
 async def static_cache_headers(request, call_next):
     """Long-cache the content-hashed SPA bundles. Vite hashes their filename, so a given
@@ -243,6 +291,15 @@ async def security_headers(request, call_next):
     return response
 
 
+# The access log (app/logs.py): request id, route template, status, duration — one line, never a
+# query string. Registered LAST so it is the OUTERMOST http middleware and therefore times the whole
+# request, including what every layer above does — the body cap included, so a 413 is still logged
+# with its request id and duration. Being outermost also puts it OUTSIDE SessionMiddleware,
+# which is why the account comes from `request.state.user_id` (set by the auth dependency, the one
+# place a session is actually resolved into an account) rather than from `request.session`.
+app.middleware("http")(log_requests)
+
+
 app.include_router(health.router)
 app.include_router(auth_routes.router)
 app.include_router(sync.router)
@@ -251,6 +308,8 @@ app.include_router(feed.router)
 app.include_router(discover_routes.router)
 app.include_router(search_routes.router)
 app.include_router(me.router)
+app.include_router(diagnostics_routes.router)
+app.include_router(debug_sessions_routes.router)
 app.include_router(notifications.router)
 app.include_router(messages.router)
 app.include_router(channels.router)
@@ -264,14 +323,13 @@ app.include_router(admin.router)
 app.include_router(config_routes.router)
 app.include_router(scheduler_routes.router)
 app.include_router(audit_routes.router)
+app.include_router(admin_logs_routes.router)
 app.include_router(quota.router)
 app.include_router(version.router)
 app.include_router(setup_routes.router)
 
 # Ensure modern image types resolve to the right Content-Type when FileResponse guesses from the
 # filename (the runtime's mimetypes db doesn't always know .webp → it'd fall back to octet-stream).
-import mimetypes
-
 mimetypes.add_type("image/webp", ".webp")
 mimetypes.add_type("image/avif", ".avif")
 

backend/app/models.py

@@ -12,6 +12,8 @@ from sqlalchemy import (
     Index,
     Integer,
     JSON,
+    LargeBinary,
+    SmallInteger,
     String,
     Text,
     UniqueConstraint,
@@ -23,6 +25,32 @@ from sqlalchemy.orm import Mapped, mapped_column, relationship, validates
 from app.db import Base
 from app.utils import canonical_email
 
+# Every user-supplied display string in this schema shares one width, and it lives HERE so the
+# columns below are declared with it rather than with a literal: a name a person typed goes into
+# `User.display_name` (their Google profile name), `DownloadJob.display_name` and
+# `DownloadJob.display_uploader`. A value wider than the column is a psycopg DataError, which
+# `main.py` re-raises for everything but the NUL case — i.e. a bare 500 for a name somebody pasted.
+DISPLAY_TEXT_MAX = 255
+
+
+def display_text(raw: str | None) -> str | None:
+    """A user-supplied display string on its way into one of those columns: stripped, truncated,
+    blank -> None.
+
+    One definition because there are SIX writers and they had drifted: `POST /api/downloads`,
+    `POST /api/downloads/edit`, `POST /api/downloads/upload`, the metadata PATCH's two fields, and
+    the Google sign-in callback. Only the PATCH truncated. Truncating is the right answer rather
+    than refusing: the name is a label on work the user has already done — a file already
+    transferred, an account already authenticated — and losing its tail is not worth failing the
+    request over.
+
+    `test_upload.py` holds both halves: that every route assigning one of these fields goes through
+    this helper (parsed, not grepped), and that each of those routes truncates rather than 500s. A
+    sweep is exactly the kind of thing that gets done five-sixths of the way, which is how this
+    ended up being written twice.
+    """
+    return (raw or "").strip()[:DISPLAY_TEXT_MAX] or None
+
 
 # The resume-position columns (VideoState / PlexState.position_seconds) are int4. The checkpoint
 # routes bound their bodies by this, so an out-of-range position is a 422 rather than a 500.
@@ -102,7 +130,17 @@ class User(Base, TimestampMixin, CanonicalEmailMixin):
     # current_user rejects any cookie whose epoch is stale, so a stolen/copied cookie dies on those
     # events (client-side signed cookies otherwise stay valid until they expire).
     session_epoch: Mapped[int] = mapped_column(Integer, default=0, server_default="0")
-    display_name: Mapped[str | None] = mapped_column(String(255))
+    # When this account last signed in (Google or password), stamped by the one sign-in path,
+    # routes/auth._establish_session. NULL = not since the column existed (0085; no backfill is
+    # possible). The shared demo account is never stamped — its "last login" would mean nobody.
+    last_login_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    # The diagnostics level this user consented to (E22 S5) and when it lapses back to the base
+    # level. Both NULL = base. Written only by `PUT /api/me/diagnostics`, which computes the expiry
+    # itself; read through `app.diagnostics.effective_level`, never directly (an expired level is
+    # not cleared eagerly — it simply stops counting).
+    diagnostics_level: Mapped[str | None] = mapped_column(String(8))
+    diagnostics_until: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    display_name: Mapped[str | None] = mapped_column(String(DISPLAY_TEXT_MAX))
     avatar_url: Mapped[str | None] = mapped_column(String(1024))
     role: Mapped[str] = mapped_column(String(16), default="user", server_default="user")
     # The user's own YouTube channel id, resolved lazily (channels.list?mine=true) and
@@ -230,6 +268,9 @@ class Channel(Base, TimestampMixin):
     details_synced_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
     recent_synced_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
     last_rss_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    # When the newest uploads were last read through the Data API because the RSS feed would not
+    # answer (`sync/runner.py` `run_feed_fallback`): the once-per-`rss_fallback_hours` clock.
+    feed_fallback_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
     backfill_cursor: Mapped[str | None] = mapped_column(String(128))
     backfill_done: Mapped[bool] = mapped_column(Boolean, default=False, server_default="false")
     # True when this channel only entered the catalog because a live YouTube search
@@ -287,10 +328,12 @@ class Video(Base, TimestampMixin):
     __tablename__ = "videos"
 
     id: Mapped[str] = mapped_column(String(16), primary_key=True)  # video id
-    # No standalone channel_id index: the composite `ix_videos_channel_pub_desc`
+    # No standalone channel_id index: the composite `ix_videos_channel_pub_cover`
-    # (channel_id, published_at DESC NULLS LAST, id) from migration 0067 leads with channel_id, so it
+    # (channel_id, published_at DESC NULLS LAST, id) from migration 0095 leads with channel_id, so it
-    # serves every channel_id lookup (FK cascade, sync upsert, channel-detail) AND the per-channel feed
+    # serves every channel_id lookup (FK cascade, sync upsert, channel-detail) and the per-channel feed
-    # sort. Like the search_vector GIN, this composite is migration-defined, not an inline `index=True`.
+    # sort. Its INCLUDEd columns make the channel manager's sums and the feed sidebar's counts Index Only
+    # Scans (no heap): a column those queries start reading belongs in it (test_channel_list_index_only).
+    # Like the search_vector GIN, this composite is migration-defined, not an inline `index=True`.
     channel_id: Mapped[str] = mapped_column(ForeignKey("channels.id", ondelete="CASCADE"))
     title: Mapped[str | None] = mapped_column(Text)  # normalized for display (see app.titles)
     original_title: Mapped[str | None] = mapped_column(Text)  # raw YouTube title, preserved
@@ -304,6 +347,9 @@ class Video(Base, TimestampMixin):
         DateTime(timezone=True), index=True, nullable=False, server_default=func.now()
     )
     thumbnail_url: Mapped[str | None] = mapped_column(String(1024))
+    # The feed's views / duration / title sorts each ride a migration-defined index in their exact
+    # ORDER BY, one per direction (and per ICU collation for title): migration 0096, held to
+    # `feedquery.order_by` by test_feed_sort_indexes.
     duration_seconds: Mapped[int | None] = mapped_column(Integer)
     view_count: Mapped[int | None] = mapped_column(BigInteger)
     like_count: Mapped[int | None] = mapped_column(BigInteger)
@@ -539,6 +585,91 @@ class AuditLog(Base):
     )
 
 
+class LogEvent(Base):
+    """One stored log event (E22 S3) — the `log_events` table the diagnostics console reads. Rows are
+    written by app/logsink.py's background writer, never inside a request's transaction, already
+    redacted (the same `logs.record_fields` the stdout line is built from).
+
+    `user_id` has NO foreign key on purpose: the writer runs behind the event, so a line about an
+    account can arrive after that account was purged, and a constraint would make it fail the whole
+    batch it travels in. `auth.purge_user` deletes an account's rows explicitly instead, after
+    `logs.forget_user` makes every later write in this process store such a line without the account;
+    the retention job (`log_retention_days`) bounds whatever another process writes late."""
+
+    __tablename__ = "log_events"
+    __table_args__ = (
+        Index("ix_log_events_user_ts", "user_id", "ts"),
+        # E16: an issue's bundle is `WHERE debug_session_id = …`; nearly every row has none.
+        Index(
+            "ix_log_events_debug_session",
+            "debug_session_id",
+            postgresql_where="debug_session_id IS NOT NULL",
+        ),
+    )
+
+    id: Mapped[int] = mapped_column(BigInteger, primary_key=True)
+    ts: Mapped[datetime] = mapped_column(DateTime(timezone=True), index=True)
+    level: Mapped[str] = mapped_column(String(10))
+    sev: Mapped[int] = mapped_column(SmallInteger)
+    source: Mapped[str] = mapped_column(String(16))
+    logger: Mapped[str] = mapped_column(String(128))
+    event: Mapped[str | None] = mapped_column(String(64))
+    msg: Mapped[str] = mapped_column(Text)
+    attrs: Mapped[dict | None] = mapped_column(JSONB)
+    exc: Mapped[str | None] = mapped_column(Text)
+    request_id: Mapped[str | None] = mapped_column(String(64), index=True)
+    user_id: Mapped[int | None] = mapped_column(Integer)
+    debug_session_id: Mapped[str | None] = mapped_column(String(64))
+    job_id: Mapped[int | None] = mapped_column(Integer)
+    version: Mapped[str | None] = mapped_column(String(64))
+
+
+class DebugSession(Base):
+    """One recording: Start → reproduce → Stop (E16, app/debugsession.py). While it is open, every
+    request of its user that names it (`X-Debug-Session`) is logged at `level` and tagged with `id`,
+    and the issue it ends in is built from the rows carrying that tag. `until` is the hard cap the
+    server set at Start; an open session past it is over whether or not the sweep has closed it yet.
+    One OPEN session per user (`uq_debug_sessions_open_per_user`, partial on `ended_at IS NULL`)."""
+
+    __tablename__ = "debug_sessions"
+    __table_args__ = (
+        Index(
+            "uq_debug_sessions_open_per_user",
+            "user_id",
+            unique=True,
+            postgresql_where="ended_at IS NULL",
+        ),
+    )
+
+    id: Mapped[str] = mapped_column(String(16), primary_key=True)
+    user_id: Mapped[int] = mapped_column(ForeignKey("users.id", ondelete="CASCADE"), index=True)
+    level: Mapped[str] = mapped_column(String(8))
+    started_at: Mapped[datetime] = mapped_column(DateTime(timezone=True))
+    until: Mapped[datetime] = mapped_column(DateTime(timezone=True))
+    ended_at: Mapped[datetime | None] = mapped_column(DateTime(timezone=True))
+    # app/debugsession.py `EndReason`: stopped by the user, or past `until`.
+    ended_reason: Mapped[str | None] = mapped_column(String(16))
+    issue_url: Mapped[str | None] = mapped_column(String(512))
+    # The issue being written from it (E16 S3; app/debugsession.py `save_draft` holds its shape).
+    draft: Mapped[dict | None] = mapped_column(JSONB)
+
+
+class DebugScreenshot(Base):
+    """A screenshot attached to a debug session (E16) — pasted, dropped, or captured from the pill.
+    Stored in the row: bounded per session (app/debugsession.py), and gone with it by cascade."""
+
+    __tablename__ = "debug_screenshots"
+
+    id: Mapped[int] = mapped_column(primary_key=True)
+    session_id: Mapped[str] = mapped_column(
+        ForeignKey("debug_sessions.id", ondelete="CASCADE"), index=True
+    )
+    created_at: Mapped[datetime] = mapped_column(DateTime(timezone=True))
+    mime: Mapped[str] = mapped_column(String(16))
+    size: Mapped[int] = mapped_column(Integer)
+    data: Mapped[bytes] = mapped_column(LargeBinary)
+
+
 class AppState(Base):
     """Single-row global app state (admin-controlled)."""
 
@@ -853,11 +984,11 @@ class DownloadJob(Base, TimestampMixin, UpdatedAtMixin):
         ForeignKey("download_profiles.id", ondelete="SET NULL")
     )
     profile_snapshot: Mapped[dict] = mapped_column(JSON)
-    display_name: Mapped[str | None] = mapped_column(String(255))
+    display_name: Mapped[str | None] = mapped_column(String(DISPLAY_TEXT_MAX))
     # Per-user display overrides + extras (like display_name — metadata only, never the file).
     # `display_uploader`/`display_uploader_url` override the asset's auto channel name/link;
     # `extra_links` is an optional list of extra reference URLs shown on the card + watch page.
-    display_uploader: Mapped[str | None] = mapped_column(String(255))
+    display_uploader: Mapped[str | None] = mapped_column(String(DISPLAY_TEXT_MAX))
     display_uploader_url: Mapped[str | None] = mapped_column(Text)
     extra_links: Mapped[list | None] = mapped_column(JSON)
     status: Mapped[str] = mapped_column(
@@ -1096,6 +1227,7 @@ class PlexShow(Base, TimestampMixin, UpdatedAtMixin):
     # Hungarian-metadata twin fields, written by the overlay sync (E5 S5). NULL = no HU twin matched,
     # so serving falls back to the EN columns. `people_text` is language-neutral, reused for both vectors.
     title_hu: Mapped[str | None] = mapped_column(String(512))
+    tagline_hu: Mapped[str | None] = mapped_column(Text)  # the HU overlay's tagline (E30)
     summary_hu: Mapped[str | None] = mapped_column(Text)
     genres_hu: Mapped[list | None] = mapped_column(JSONB(none_as_null=True))
     search_vector: Mapped[object | None] = mapped_column(
@@ -1257,6 +1389,7 @@ class PlexItem(Base, TimestampMixin, UpdatedAtMixin):
     # Hungarian-metadata twin fields, written by the overlay sync (E5 S5). NULL = no HU twin matched
     # (serving falls back to the EN columns). `people_text` is language-neutral, reused for both vectors.
     title_hu: Mapped[str | None] = mapped_column(String(512))
+    tagline_hu: Mapped[str | None] = mapped_column(Text)  # the HU overlay's tagline (E30)
     summary_hu: Mapped[str | None] = mapped_column(Text)
     genres_hu: Mapped[list | None] = mapped_column(JSONB(none_as_null=True))
     search_vector: Mapped[object | None] = mapped_column(
@@ -1287,7 +1420,8 @@ class PlexPerson(Base, TimestampMixin, UpdatedAtMixin):
     top few cast. Rows whose name no longer appears in ANY item's cast_names/directors (a library
     removal orphaned them) are pruned by the maintenance sweep, mirroring how Plex drops orphaned
     metadata. The filterable cast lives in plex_items/plex_shows.cast_names (self-cleaning with the
-    row); this table is ONLY the headshot cache."""
+    row); this table is the people's own facts: the headshot link, the TMDB link and the TMDB bio
+    (the headshot FILE lives in the image cache, filled by the people_enrich job)."""
 
     __tablename__ = "plex_people"
 
@@ -1297,11 +1431,15 @@ class PlexPerson(Base, TimestampMixin, UpdatedAtMixin):
     # TMDB person id, linked by name from a title's TMDB credits (E5 S2 lays the foundation; the /person
     # biography fetch + person-page render land in S6). NULL until a TMDB-enriched title names this person.
     tmdb_person_id: Mapped[int | None] = mapped_column(Integer)
-    # --- TMDB person enrichment (E5 S6): fetched from /person/{tmdb_person_id} the first time the person
+    # --- TMDB person enrichment (E5 S6): fetched from /person/{tmdb_person_id} by the people_enrich job
-    #     page is opened, then SERVED FROM THE ROW. `tmdb_enriched_at` is the watermark (set even on a
+    #     (E30 S6), or by the first person-page open the job hasn't reached, then SERVED FROM THE ROW. `tmdb_enriched_at` is the watermark (set even on a
     #     no-op fetch so it isn't retried every open). `known_for` = top combined-credits [{tmdb_id,
     #     media_type, title}], resolved to library rows by guid at serve time (like recommendations). ---
     biography: Mapped[str | None] = mapped_column(Text)
+    # TMDB's Hungarian biography (E30 S4), from the same request's `translations`; a Hungarian reader is
+    # served it, and the instance-language `biography` when TMDB has none. Birthplace, aliases and
+    # keywords have no Hungarian on TMDB, so they stay in the instance language.
+    biography_hu: Mapped[str | None] = mapped_column(Text)
     birthday: Mapped[str | None] = mapped_column(String(10))  # "YYYY-MM-DD", served as-is
     deathday: Mapped[str | None] = mapped_column(String(10))
     place_of_birth: Mapped[str | None] = mapped_column(Text)

backend/app/plex/cards.py

@@ -15,20 +15,35 @@ def _art_url(rating_key: str, art_key: str | None) -> str | None:
 
 
 # Card hover-overview: a short synopsis for the reveal panel, trimmed so a strip/related row of cards
-# stays a few hundred bytes each rather than carrying full synopses. (These non-grid card paths serve
+# stays a few hundred bytes each rather than carrying full synopses.
-# the EN summary/rating; the main grid coalesces the HU twin in browse.unified_library.)
 def _card_summary(summary: str | None) -> str | None:
     return (summary or "")[:300] or None
 
 
-def movie_card(it: PlexItem, st: PlexState | None) -> dict:
+def norm_lang(lang: str | None) -> str:
+    """The overlay has an EN base + an optional HU twin; anything other than `hu` serves EN."""
+    return "hu" if lang == "hu" else "en"
+
+
+# Every card speaks the reader's language (E30): `lang="hu"` serves the HU overlay twin's title and
+# summary, falling back to EN where a row has none — the same rule as browse.unified_library's
+# `title_col`/`summary_col`. These rails once served English to a Hungarian reader.
+def display_title(row: PlexItem | PlexShow, lang: str) -> str:
+    return (row.title_hu or row.title) if lang == "hu" else row.title
+
+
+def display_summary(row: PlexItem | PlexShow, lang: str) -> str | None:
+    return (row.summary_hu or row.summary) if lang == "hu" else row.summary
+
+
+def movie_card(it: PlexItem, st: PlexState | None, *, lang: str) -> dict:
     return {
         "id": it.rating_key,
         "type": "movie",
-        "title": it.title,
+        "title": display_title(it, lang),
         "year": it.year,
         "rating": it.rating,
-        "summary": _card_summary(it.summary),
+        "summary": _card_summary(display_summary(it, lang)),
         "duration_seconds": it.duration_s,
         "thumb": f"/api/plex/image/{it.rating_key}",
         "art": _art_url(it.rating_key, it.art_key),
@@ -39,14 +54,14 @@ def movie_card(it: PlexItem, st: PlexState | None) -> dict:
     }
 
 
-def show_card(sh: PlexShow, status: str = "new") -> dict:
+def show_card(sh: PlexShow, status: str = "new", *, lang: str) -> dict:
     return {
         "id": sh.rating_key,
         "type": "show",
-        "title": sh.title,
+        "title": display_title(sh, lang),
         "year": sh.year,
         "rating": sh.rating,
-        "summary": _card_summary(sh.summary),
+        "summary": _card_summary(display_summary(sh, lang)),
         "thumb": f"/api/plex/image/{sh.rating_key}",
         "art": _art_url(sh.rating_key, sh.art_key),
         "season_count": sh.child_count,
@@ -55,12 +70,12 @@ def show_card(sh: PlexShow, status: str = "new") -> dict:
     }
 
 
-def episode_card(e: PlexItem, st: PlexState | None) -> dict:
+def episode_card(e: PlexItem, st: PlexState | None, *, lang: str) -> dict:
     return {
         "id": e.rating_key,
         "type": "episode",
-        "title": e.title,
+        "title": display_title(e, lang),
-        "summary": e.summary,
+        "summary": display_summary(e, lang),
         "season_number": e.season_number,
         "episode_number": e.episode_number,
         "duration_seconds": e.duration_s,
@@ -77,12 +92,15 @@ def leaf_card(
     st: PlexState | None,
     show_title: str | None = None,
     show_rk: str | None = None,
+    *,
+    lang: str,
 ) -> dict:
     """A playable-leaf card (movie or episode). Used by playlists, which can mix both kinds.
-    `show_rk` (the show's rating_key) lets the client group a playlist's episodes by show/season."""
+    `show_rk` (the show's rating_key) lets the client group a playlist's episodes by show/season.
+    `show_title` is the caller's, already in `lang` (see `display_title`)."""
     if it.kind == "episode":
-        card = episode_card(it, st)
+        card = episode_card(it, st, lang=lang)
         card["show_title"] = show_title
         card["show_id"] = show_rk
         return card
-    return movie_card(it, st)
+    return movie_card(it, st, lang=lang)

backend/app/plex/client.py

@@ -6,16 +6,15 @@ metadata, image proxying, and (P5) optional watch-state write-back. Requests JSO
 ``Accept: application/json`` header (Plex serves XML by default).
 """
 
-import logging
 from urllib.parse import quote
 
 import httpx
 from sqlalchemy.orm import Session
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.config import settings
 
-log = logging.getLogger("siftlode.plex")
+log = logs.get_logger(__name__)
 
 
 class PlexError(Exception):
@@ -79,9 +78,10 @@ class PlexClient:
     # --- Catalog mirror (P1) ---
     def section_items(
         self, key: str, item_type: int, start: int = 0, size: int = 200
-    ) -> tuple[list[dict], int]:
+    ) -> tuple[list[dict], int | None]:
         """A page of a section's top-level items. type 1 = movie, 2 = show. Returns
-        (items, total_size)."""
+        (items, total_size); the total is None when the answer carries none, so a container that
+        says "0" (a genuinely empty section) stays distinguishable from one that says nothing."""
         mc = self._get(
             f"/library/sections/{key}/all",
             params={
@@ -94,7 +94,10 @@ class PlexClient:
                 "X-Plex-Container-Size": size,
             },
         )
-        return mc.get("Metadata", []) or [], int(mc.get("totalSize", mc.get("size", 0)) or 0)
+        # `totalSize` only: `size` is THIS page's count, and read as the total it made a single
+        # page look like the whole section — "complete" to a prune or a reconcile (E32 review).
+        total = mc.get("totalSize")
+        return mc.get("Metadata", []) or [], None if total is None else int(total)
 
     def children(self, rating_key: str) -> list[dict]:
         """Direct children (show → seasons, season → episodes)."""
@@ -106,7 +109,18 @@ class PlexClient:
         Best-effort (needs the library's related data) — returns [] if Plex has none."""
         try:
             mc = self._get(f"/library/metadata/{rating_key}/related")
-        except PlexError:
+        except PlexError as exc:
+            # An empty list is also the answer for an item that genuinely has no related titles,
+            # so the caller cannot tell the two apart — and it deliberately declines to CACHE an
+            # empty one for exactly that reason. This line is what says which it was.
+            log.debug(
+                "the related items for a Plex item could not be read",
+                extra=logs.event(
+                    "plex.client.related_unavailable",
+                    rating_key=rating_key,
+                    reason=type(exc).__name__,
+                ),
+            )
             return []
         out: list[dict] = []
         for hub in mc.get("Hub", []) or []:
@@ -126,9 +140,14 @@ class PlexClient:
     # --- Collection editing (P2 — write-back to Plex; all shared/global, admin-gated in the routes) ---
     @property
     def machine_id(self) -> str:
-        """Server machineIdentifier — needed in the `uri` of collection add/create calls. Cached."""
+        """Server machineIdentifier — needed in the `uri` of collection add/create calls. Cached.
+        Without one the uri would read `server:///…`, a malformed write Plex answers in its own way;
+        it is refused here, by name (E32 S2)."""
         if not getattr(self, "_machine_id", None):
-            self._machine_id = self._get("/identity").get("machineIdentifier") or ""
+            mid = self._get("/identity").get("machineIdentifier")
+            if not mid:
+                raise PlexError("Plex did not report its machine identifier")
+            self._machine_id = mid
         return self._machine_id
 
     def _write(self, method: str, path: str) -> dict:
@@ -139,6 +158,8 @@ class PlexClient:
             raise PlexError(str(e)) from e
         try:
             return (r.json() or {}).get("MediaContainer", {}) or {} if r.content else {}
+        # silent-ok: the HTTP call already succeeded — this is a write endpoint, and Plex answers
+        # those with an empty or non-JSON body as a matter of course.
         except ValueError:
             return {}
 

backend/app/plex/detail.py

@@ -22,6 +22,8 @@ def rich_detail(meta: dict) -> dict:
         if str(r.get("image") or "").startswith("imdb://"):
             try:
                 imdb_rating = round(float(r.get("value")), 1)
+            # silent-ok: no IMDb score, which is the common and legitimate case — the UI has a
+            # separate audience score to fall back on and shows nothing if neither exists.
             except (TypeError, ValueError):
                 pass
             break

backend/app/plex/genres.py

@@ -0,0 +1,100 @@
+"""One Hungarian word per Plex genre (E30).
+
+Plex's Hungarian agent returns a MIXED vocabulary: some genres come back translated ("Dráma",
+"Vígjáték"), others stay English ("Action", "Adventure", "Biography"), and a title with no HU twin has
+only its English genres. The HU genre list therefore offered "Drama" beside "Dráma" and a dozen English
+words. This map sends every genre word we have met — English (Plex EN agent, TMDB, IMDb) or Plex's
+Hungarian — to the ONE word a Hungarian reader sees. The overlay sync writes it into `genres_hu`, and a
+genre filter is read through it too, so a saved filter still holding "Action" keeps matching "Akció".
+
+A word not in the map passes through unchanged: an unknown genre stays visible (in whatever language
+Plex gave it) rather than disappearing.
+"""
+
+from collections.abc import Iterable
+
+# Keys are matched case-insensitively (see `_BY_KEY`). The Hungarian values double as keys mapping to
+# themselves, so running a list through twice changes nothing.
+_HU: dict[str, str] = {
+    "Action": "Akció",
+    "Action & Adventure": "Akció és kaland",
+    "Adventure": "Kaland",
+    "Animation": "Animációs",
+    "Anime": "Anime",
+    "Biography": "Életrajzi",
+    "Children": "Gyerek",
+    "Comedy": "Vígjáték",
+    "Crime": "Bűnügyi",
+    "Documentary": "Dokumentum",
+    "Drama": "Dráma",
+    "Family": "Családi",
+    "Fantasy": "Fantasy",
+    "Film-Noir": "Film noir",
+    "Food": "Gasztronómia",
+    "Game Show": "Vetélkedő",
+    "Game-Show": "Vetélkedő",
+    "History": "Történelmi",
+    "Home and Garden": "Otthon és kert",
+    "Horror": "Horror",
+    "Indie": "Független",
+    "Kids": "Gyerek",
+    "Martial Arts": "Harcművészet",
+    "Mini-Series": "Minisorozat",
+    "Music": "Zenei",
+    "Musical": "Musical",
+    "Mystery": "Rejtély",
+    "News": "Hírek",
+    "Reality": "Valóságshow",
+    "Reality-TV": "Valóságshow",
+    "Romance": "Romantikus",
+    "Sci-Fi": "Sci-Fi",
+    "Sci-Fi & Fantasy": "Sci-Fi és fantasy",
+    "Science Fiction": "Sci-Fi",
+    "Short": "Rövidfilm",
+    "Soap": "Szappanopera",
+    "Sport": "Sport",
+    "Suspense": "Feszültség",
+    "Talk": "Talkshow",
+    "Talk Show": "Talkshow",
+    "Thriller": "Thriller",
+    "Travel": "Utazás",
+    "TV Movie": "TV film",
+    "War": "Háborús",
+    "War & Politics": "Háború és politika",
+    "Western": "Western",
+    # Plex's own Hungarian words, pinned to themselves (and to the spelling above where they differ).
+    "Akció": "Akció",
+    "Animációs": "Animációs",
+    "Bűnügyi": "Bűnügyi",
+    "Családi": "Családi",
+    "Dokumentum": "Dokumentum",
+    "Dokumentumfilm": "Dokumentum",
+    "Dráma": "Dráma",
+    "Életrajzi": "Életrajzi",
+    "Háborús": "Háborús",
+    "Kaland": "Kaland",
+    "Rejtély": "Rejtély",
+    "Romantikus": "Romantikus",
+    "Történelmi": "Történelmi",
+    "TV film": "TV film",
+    "Vígjáték": "Vígjáték",
+    "Zene": "Zenei",
+    "Zenei": "Zenei",
+}
+_BY_KEY = {k.casefold(): v for k, v in _HU.items()}
+
+
+def hu_genre(word: str) -> str:
+    """The Hungarian word for one genre; an unknown word comes back unchanged."""
+    return _BY_KEY.get(word.strip().casefold(), word)
+
+
+def hu_genres(words: Iterable[str] | None) -> list[str] | None:
+    """A genre list in Hungarian, in its first-seen order, each word once ("Drama" + "Dráma" → one
+    "Dráma"). None or empty stays None, like the listing's own `_tags`."""
+    out: list[str] = []
+    for w in words or ():
+        hu = hu_genre(w)
+        if hu not in out:
+            out.append(hu)
+    return out or None

backend/app/plex/imgcache.py

@@ -8,18 +8,38 @@ few hundred MB at most for a big library — acceptable, matching the download d
 The cache directory (``cache.CACHE_DIR``) is shared with the subtitle cache (see ``subtitles.py``).
 """
 
+import time
+from datetime import timedelta
+
 from fastapi import Response
 from sqlalchemy.orm import Session
 
+from app import logs
 from app.models import PlexCollection, PlexItem, PlexSeason, PlexShow
 from app.plex.cache import CACHE_DIR
 from app.plex.lookups import by_rating_key
 
+log = logs.get_logger(__name__)
+
 _EXT_CT = [("jpg", "image/jpeg"), ("png", "image/png"), ("webp", "image/webp")]
 _CT_EXT = {ct: ext for ext, ct in _EXT_CT}
-# Target box per variant — Plex resizes to these server-side. Posters fill ≤176px cells (grid/info/
+# Target box per variant — Plex resizes to these server-side. The full poster (400px) covers a detail
-# strips), so 400px covers 2× DPR; the faint backdrop art only needs ~1280px. Keeps images ~tens of KB.
+# page's ≤224px poster at 2× DPR, the full art (1280px) a hero backdrop. A card is much smaller: a grid
-SIZES = {"thumb": (400, 600), "art": (1280, 720)}
+# poster is ~144px and a shelf's landscape art ~212px, so the `_sm` pair serves those at 1.33× DPR
+# (E31 S4e: cards fetched the 400 poster, 42 KB, and the 1280 art for a 212px picture). The browser
+# picks between a pair itself (`srcset`, built by the frontend's `lib/plexImage` cardImage, whose
+# widths plexImage.contract.test holds to these).
+SIZES = {
+    "thumb": (400, 600),
+    "thumb_sm": (240, 360),
+    "art": (1280, 720),
+    "art_sm": (480, 270),
+}
+
+
+def picture(variant: str) -> str:
+    """Which stored picture a size variant is cut from: the poster (`thumb`) or the backdrop (`art`)."""
+    return "art" if variant.startswith("art") else "thumb"
 
 
 def key_for(db: Session, rating_key: str, variant: str) -> str | None:
@@ -46,11 +66,56 @@ def read(key: str) -> tuple[bytes, str] | None:
         try:
             if p.exists() and p.stat().st_size > 0:
                 return p.read_bytes(), ct
+        # silent-ok: a miss. The caller fetches from Plex and writes the file again, and the write
+        # path below is the one that reports a cache that cannot be used.
         except OSError:
             pass
     return None
 
 
+def has(key: str) -> bool:
+    """A cached, non-empty file exists for `key` (a prefetch's cheap check — no read)."""
+    for ext, _ in _EXT_CT:
+        try:
+            if (CACHE_DIR / f"{key}.{ext}").stat().st_size > 0:
+                return True
+        # silent-ok: a miss, as in read().
+        except OSError:
+            pass
+    return False
+
+
+def drop(key: str) -> None:
+    """Remove `key`'s file, whatever its extension (a superseded variant). Best-effort."""
+    for ext, _ in _EXT_CT:
+        try:
+            (CACHE_DIR / f"{key}.{ext}").unlink(missing_ok=True)
+        # silent-ok: a leftover original only costs disk; the small twin is already written.
+        except OSError:
+            pass
+
+
+def mark_miss(key: str) -> None:
+    """Remember that `key`'s source could not be fetched (an empty `.miss` file; its mtime is when), so
+    a prefetch can leave a dead source alone for a while. Best-effort."""
+    try:
+        CACHE_DIR.mkdir(parents=True, exist_ok=True)
+        (CACHE_DIR / f"{key}.miss").touch()
+    # silent-ok: without the mark the dead source is only asked for again on the next run.
+    except OSError:
+        pass
+
+
+def missed_within(key: str, age: timedelta) -> bool:
+    """`key` was marked as a miss less than `age` ago."""
+    try:
+        mtime = (CACHE_DIR / f"{key}.miss").stat().st_mtime
+    # silent-ok: no mark (the usual case).
+    except OSError:
+        return False
+    return time.time() - mtime < age.total_seconds()
+
+
 def write(key: str, data: bytes, ct: str) -> None:
     ext = _CT_EXT.get((ct or "").split(";")[0].strip(), "jpg")
     try:
@@ -58,8 +123,14 @@ def write(key: str, data: bytes, ct: str) -> None:
         tmp = CACHE_DIR / f"{key}.{ext}.tmp"
         tmp.write_bytes(data)
         tmp.replace(CACHE_DIR / f"{key}.{ext}")  # atomic publish
-    except OSError:
+    except OSError as exc:
-        pass  # cache is best-effort
+        # Best-effort, and silently so: with an unwritable cache every poster on every grid goes
+        # to Plex live, for as long as nobody notices. Debug, because a grid is ~100 images and
+        # the point is the condition, not the count.
+        log.debug(
+            "the Plex image cache could not be written",
+            extra=logs.event("plex.imgcache.write_failed", reason=type(exc).__name__),
+        )
 
 
 def response(data: bytes, ct: str) -> Response:

backend/app/plex/overlay.py

@@ -0,0 +1,150 @@
+"""The HU overlay: a Hungarian Plex section written onto its English twin's rows (E30), split out of
+`plex/sync.py` (E32 S2), which runs it once per paired section during the full sync.
+
+The pairing is admin-configured (`plex_overlay_map`, ``huKey:enKey,...``). An overlay section creates
+no browsable rows of its own; it only fills the `*_hu` columns of the EN rows it matches by guid.
+`plex/sync.py` and this module import each other as MODULES (attribute access at call time), which is
+what lets either be imported first.
+"""
+
+from collections.abc import Iterable
+
+from sqlalchemy import select
+from sqlalchemy.orm import Session
+
+from app import logs, sysconfig
+from app.models import PlexItem, PlexLibrary, PlexShow
+from app.plex import sync as plex_sync
+from app.plex.client import PlexClient
+from app.plex.genres import hu_genres
+
+log = logs.get_logger(__name__)
+
+
+def overlay_map(db: Session) -> dict[str, str]:
+    """The configured HU-overlay pairing: {huSectionKey: enSectionKey} from `plex_overlay_map`
+    ('huKey:enKey,...'). Each HU section is synced as a non-browsable overlay onto its EN twin.
+    Malformed pairs are skipped."""
+    raw = sysconfig.get_str(db, "plex_overlay_map")
+    out: dict[str, str] = {}
+    for pair in raw.split(","):
+        hu, sep, en = pair.partition(":")
+        hu, en = hu.strip(), en.strip()
+        if sep and hu and en:
+            out[hu] = en
+    return out
+
+
+def sync_overlay(
+    db: Session, plex: PlexClient, lib: PlexLibrary, en_lib_id: int, stats: dict
+) -> None:
+    """Write an overlay section's Hungarian metadata onto its EN twin's rows, matched by guid. Creates
+    NO browsable rows — an overlay library is never listed on its own. Only title/summary/tagline/genres
+    differ by language; the language-neutral people_text feeds both search vectors. A show section's
+    episodes are overlaid too (E30): the HU section lists every episode with its Hungarian title and
+    summary, and the episode rows live in the EN show library.
+
+    This is a full RECONCILE: after a COMPLETE HU listing every guid-bearing EN row's HU fields are
+    cleared, then the matches are repopulated — so an EN row whose HU twin vanished (or was re-guid'd)
+    falls back to EN instead of keeping stale Hungarian metadata. A listing cut short only adds. HU
+    items with no EN match are skipped."""
+    if lib.kind == "movie":
+        passes = [
+            (
+                plex_sync._TYPE_MOVIE,
+                select(PlexItem).where(
+                    PlexItem.library_id == en_lib_id,
+                    PlexItem.kind == "movie",
+                    PlexItem.guid.is_not(None),
+                ),
+            )
+        ]
+    else:
+        passes = [
+            (
+                plex_sync._TYPE_SHOW,
+                select(PlexShow).where(
+                    PlexShow.library_id == en_lib_id, PlexShow.guid.is_not(None)
+                ),
+            ),
+            (
+                plex_sync._TYPE_EPISODE,
+                select(PlexItem).where(
+                    PlexItem.library_id == en_lib_id,
+                    PlexItem.kind == "episode",
+                    PlexItem.guid.is_not(None),
+                ),
+            ),
+        ]
+    for item_type, rows_q in passes:
+        _overlay_pass(db, plex, lib, item_type, db.execute(rows_q).scalars(), stats)
+    db.flush()
+
+
+def _overlay_pass(
+    db: Session,
+    plex: PlexClient,
+    lib: PlexLibrary,
+    item_type: int,
+    rows: Iterable[PlexItem] | Iterable[PlexShow],
+    stats: dict,
+) -> None:
+    """One kind of the overlay (movies, shows or episodes) — see `sync_overlay`."""
+    # guid -> the EN rows carrying it (a list: two editions of one film can share a guid; both overlay).
+    by_guid: dict[str, list[PlexItem | PlexShow]] = {}
+    for r in rows:
+        if r.guid:  # the passes select guid-bearing rows only
+            by_guid.setdefault(r.guid, []).append(r)
+    # The HU listing is read first and applied after, so the clear can wait for its verdict.
+    hu_by_guid, complete = _read_hu_listing(plex, lib, item_type, by_guid)
+    episodes = item_type == plex_sync._TYPE_EPISODE
+    if complete:
+        _clear_hu(by_guid, episodes)
+    else:
+        stats["overlay_incomplete"] = stats.get("overlay_incomplete", 0) + 1
+        log.warning(
+            "the HU listing ended before the server's total; existing HU text is kept",
+            extra=logs.event("plex.sync.overlay_incomplete", library=lib.plex_key),
+        )
+    for g, meta in hu_by_guid.items():
+        targets = by_guid[g]
+        title_hu = (meta.get("title") or "").strip() or None
+        summary_hu = meta.get("summary")
+        tagline_hu = (meta.get("tagline") or "").strip() or None
+        genres_hu = hu_genres(plex_sync._tags(meta, "Genre"))
+        for row in targets:
+            row.title_hu = title_hu
+            row.summary_hu = summary_hu
+            if not episodes:
+                row.tagline_hu = tagline_hu
+                row.genres_hu = genres_hu or row.genres_hu
+        stats["overlaid"] = stats.get("overlaid", 0) + 1
+
+
+def _read_hu_listing(
+    plex: PlexClient, lib: PlexLibrary, item_type: int, by_guid: dict
+) -> tuple[dict[str, dict], bool]:
+    """The HU section's entries for the guids the EN side carries, and whether the walk reached the
+    server's total. The first HU entry for a guid wins; a dup HU item can't overwrite or double-count."""
+    progress: dict = {}
+    hu_by_guid: dict[str, dict] = {}
+    for meta in plex_sync._paginate(plex, lib.plex_key, item_type, progress):
+        g = plex_sync._canonical_guid(meta)
+        if g and g in by_guid and g not in hu_by_guid:
+            hu_by_guid[g] = meta
+    return hu_by_guid, bool(progress.get("complete"))
+
+
+def _clear_hu(by_guid: dict, episodes: bool) -> None:
+    """Reconcile step: clear HU fields on all guid-bearing rows; the matches are re-set after. Only
+    rows that actually held HU data become dirty (SQLAlchemy skips a no-net-change UPDATE), so this is
+    cheap. The genres are the exception: every row gets Hungarian ones, from its HU twin or, lacking
+    one, from its own English list — otherwise the HU genre list mixed "Drama" (no twin) with "Dráma".
+    Only after a COMPLETE listing: one a Plex hiccup cut short would clear the HU text of every title
+    it failed to reach, until the next full sync (E32 S2, the watch reconcile's twin)."""
+    for group in by_guid.values():
+        for row in group:
+            row.title_hu = row.summary_hu = None
+            if not episodes:
+                row.tagline_hu = None
+                row.genres_hu = hu_genres(row.genres)

backend/app/plex/paths.py

@@ -12,7 +12,9 @@ from pathlib import Path
 
 from sqlalchemy.orm import Session
 
-from app import sysconfig
+from app import logs, sysconfig
+
+log = logs.get_logger(__name__)
 
 
 def parse_map(raw: str) -> list[tuple[str, str]]:
@@ -47,6 +49,13 @@ def local_media_path(db: Session, plex_path: str | None) -> Path | None:
         return None
     try:
         p = Path(map_path(db, plex_path)).resolve()
-    except OSError:
+    except OSError as exc:
+        # None is answered as a 404 "no local file", which is also the answer for a title the
+        # server does not hold — so a dead mount or a wrong path map reads as missing media for
+        # the whole library at once. The path itself stays out: it carries the title.
+        log.warning(
+            "a Plex media path could not be resolved locally",
+            extra=logs.event("plex.paths.resolve_failed", reason=type(exc).__name__),
+        )
         return None
     return p if p.is_file() else None

backend/app/plex/people.py

@@ -3,13 +3,23 @@ or show (cast_names + the people_text search blob + the plex_people headshot cac
 orphaned people. Used by BOTH the info-page lazy enrichment (routes.plex) and the background rich
 re-sync (plex.sync), so the two paths never diverge."""
 
-from urllib.parse import quote
+import hashlib
+import re
+from datetime import timedelta
+from typing import Literal
+from urllib.parse import quote, unquote
 
+import httpx
 from sqlalchemy import select, text
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.orm import Session
 
+from app import logs, sysconfig
 from app.models import PlexPerson
+from app.plex import imgcache
+from app.plex.client import PlexClient, PlexError, PlexNotConfigured
+
+log = logs.get_logger(__name__)
 
 # Cast/crew photos come from Plex's PUBLIC metadata CDN (no token). We proxy them (host-whitelisted)
 # so the user's browser makes no third-party request (privacy + self-host CSP friendliness).
@@ -19,9 +29,7 @@ PERSON_IMG_HOST = "https://metadata-static.plex.tv/"
 def person_thumb_url(thumb: str | None) -> str | None:
     """Proxied /person-image URL for a Plex CDN headshot, or None when it's missing / off-CDN."""
     t = str(thumb or "")
-    return (
+    return f"/api/plex/person-image?u={quote(t, safe='')}" if is_portrait_url(t) else None
-        f"/api/plex/person-image?u={quote(t, safe='')}" if t.startswith(PERSON_IMG_HOST) else None
-    )
 
 
 def cast_from_meta(meta: dict, limit: int = 24) -> list[dict]:
@@ -89,14 +97,23 @@ def enrich_row_people(db: Session, row, cast: list[dict]) -> bool:
     return dirty
 
 
+# Every credit reference: one person name per (movie/show, credit column). The ONE list of credit
+# columns, so the orphan prune and the enrichment order never disagree about who is credited.
+_CREDIT_COLUMNS = [
+    (table, col) for table in ("plex_items", "plex_shows") for col in ("cast_names", "directors")
+]
+
+
+def _credit_refs_sql(union: Literal["UNION", "UNION ALL"]) -> str:
+    return f" {union} ".join(
+        f"SELECT jsonb_array_elements_text({col}) FROM {table} WHERE jsonb_typeof({col}) = 'array'"
+        for table, col in _CREDIT_COLUMNS
+    )
+
+
 # The person names still referenced by ANY movie/show (cast OR directors). plex_people rows outside
 # this set are orphans left by a library removal.
-_REFERENCED_PEOPLE_SQL = """
+_REFERENCED_PEOPLE_SQL = _credit_refs_sql("UNION")
-    SELECT jsonb_array_elements_text(cast_names) FROM plex_items WHERE jsonb_typeof(cast_names) = 'array'
-    UNION SELECT jsonb_array_elements_text(directors) FROM plex_items WHERE jsonb_typeof(directors) = 'array'
-    UNION SELECT jsonb_array_elements_text(cast_names) FROM plex_shows WHERE jsonb_typeof(cast_names) = 'array'
-    UNION SELECT jsonb_array_elements_text(directors) FROM plex_shows WHERE jsonb_typeof(directors) = 'array'
-"""
 
 
 def prune_orphan_people(db: Session) -> int:
@@ -113,3 +130,152 @@ def prune_orphan_people(db: Session) -> int:
     )
     db.commit()
     return res.rowcount or 0
+
+
+# How often each person is credited across movies, shows and episodes (cast OR director) — the order the
+# people_enrich job works in, so the faces met most often are local first.
+_CREDIT_COUNTS_SQL = (
+    f"SELECT ref, count(*) AS n FROM ({_credit_refs_sql('UNION ALL')}) AS refs(ref) GROUP BY ref"
+)
+_NEEDS = {
+    # linked to TMDB, bio never fetched
+    "bio": "p.tmdb_person_id IS NOT NULL AND p.tmdb_enriched_at IS NULL",
+    # has a Plex headshot (whether it is cached is the disk's answer, not the row's)
+    "portrait": "p.photo_thumb IS NOT NULL",
+}
+
+
+def people_by_credits(
+    db: Session, need: Literal["bio", "portrait"], limit: int | None
+) -> list[int]:
+    """plex_people ids that `need` something, the most-credited first (ties by id)."""
+    rows = db.execute(
+        text(
+            f"SELECT p.id FROM plex_people p LEFT JOIN ({_CREDIT_COUNTS_SQL}) c ON c.ref = p.name "
+            f"WHERE {_NEEDS[need]} ORDER BY coalesce(c.n, 0) DESC, p.id"
+            + (" LIMIT :limit" if limit is not None else "")
+        ),
+        {"limit": limit} if limit is not None else {},
+    )
+    return [r[0] for r in rows]
+
+
+# --- Portraits (E30 S6) ---------------------------------------------------------------------------
+# A Plex CDN headshot is a full-size original (~170 KB on average, measured). Plex's photo transcoder
+# fits it into this box (~13 KB), which covers the largest place a portrait is drawn (the person page's
+# 6rem avatar) at 2x density and a raised text size. The box is part of the cache key, like a poster's:
+# a new box re-caches instead of serving a stale size. Only a resized image is ever written under it —
+# an original fetched without Plex keeps its own key, and goes once its small twin lands.
+PORTRAIT_BOX = 320
+# A headshot Plex answered "not found" for (a dead CDN file) is not asked for again this long, so a
+# few dead faces never hold up the rest of the prefetch.
+PORTRAIT_RETRY = timedelta(days=7)
+# The one shape a Plex CDN headshot has (…/7/people/<hash>.jpg): plain path segments, no query, no
+# fragment, no dot segment or empty segment (an HTTP client folds `a/../b` and the CDN `a//b` onto one
+# file). The person-image proxy takes nothing else, so it cannot be made to cache endless spellings of
+# one file (the image cache has no eviction). Used where a stored headshot is made, served and warmed.
+_SEGMENT = r"[A-Za-z0-9_-]+(?:\.[A-Za-z0-9_-]+)*"  # a name, dots only between its parts
+_PORTRAIT_URL = re.compile(re.escape(PERSON_IMG_HOST) + _SEGMENT + rf"(?:/{_SEGMENT})*")
+
+
+def is_portrait_url(url: str) -> bool:
+    return _PORTRAIT_URL.fullmatch(url) is not None
+
+
+def _portrait_keys(url: str) -> tuple[str, str]:
+    """(the resized portrait's key, the full-size original's key)."""
+    base = "person_" + hashlib.sha1(url.encode()).hexdigest()
+    return f"{base}_{PORTRAIT_BOX}", base
+
+
+def portrait_cached(url: str) -> tuple[bytes, str] | None:
+    """The cached portrait: the resized one, else an original kept from a time without Plex."""
+    small, original = _portrait_keys(url)
+    return imgcache.read(small) or imgcache.read(original)
+
+
+def cache_portrait(plex: PlexClient | None, url: str) -> tuple[bytes, str]:
+    """Fetch one CDN headshot and cache it. With a Plex client it is resized (the box key; the original
+    is removed); without one the original comes straight from the public CDN and is cached under its own
+    key, so the job still resizes it once Plex is there. Raises PlexError / httpx.HTTPError."""
+    small, original = _portrait_keys(url)
+    if plex is None:
+        with httpx.Client(timeout=10.0, follow_redirects=True) as c:
+            r = c.get(url)
+            r.raise_for_status()
+        data, ct = r.content, r.headers.get("content-type", "image/jpeg")
+        imgcache.write(original, data, ct)
+        return data, ct
+    data, ct = plex.image_bytes(url, PORTRAIT_BOX, PORTRAIT_BOX)
+    imgcache.write(small, data, ct)
+    imgcache.drop(original)
+    return data, ct
+
+
+def cdn_url(photo_thumb: str) -> str | None:
+    """The CDN url inside a stored proxied `/person-image?u=` value (see person_thumb_url)."""
+    _, _, q = photo_thumb.partition("?u=")
+    url = unquote(q)
+    return url if is_portrait_url(url) else None
+
+
+# Plex failing this many headshots in a row ends a run: it is Plex that is failing, not the files.
+_PLEX_DOWN_AFTER = 3
+# The statuses that mean the one file is gone.
+_DEAD = frozenset({404, 410})
+
+
+def warm_portraits(db: Session, limit: int) -> dict:
+    """Cache up to `limit` not-yet-cached headshots, the most-credited people first (E30 S6), so a
+    person's face comes from local disk. Needs Plex (the resize) and the Plex module on; without them
+    nothing is prefetched — full-size originals for every person would cost gigabytes. A headshot Plex
+    answers "not found" for is marked and left for PORTRAIT_RETRY; Plex failing in a row ends the run."""
+    if not sysconfig.get_bool(db, "plex_enabled"):
+        return {"cached": 0, "skipped": "disabled"}
+    try:
+        plex = PlexClient(db)
+    # silent-ok: no Plex server set up is a setting, not a failure; the run's result says "disabled".
+    except PlexNotConfigured:
+        return {"cached": 0, "skipped": "disabled"}
+    photos: dict[int, str] = {
+        pid: thumb
+        for pid, thumb in db.execute(
+            select(PlexPerson.id, PlexPerson.photo_thumb).where(PlexPerson.photo_thumb.is_not(None))
+        ).tuples()
+        if thumb
+    }
+    # `down` counts Plex failing in a row (any answer from Plex breaks it), `plex_failed` in all
+    done = dead = plex_failed = down = 0
+    with plex:
+        for pid in people_by_credits(db, "portrait", None):
+            if done + dead >= limit or down >= _PLEX_DOWN_AFTER:
+                break
+            url = cdn_url(photos.get(pid) or "")
+            if url is None:
+                continue
+            key = _portrait_keys(url)[0]
+            if imgcache.has(key) or imgcache.missed_within(key, PORTRAIT_RETRY):
+                continue
+            try:
+                cache_portrait(plex, url)
+                done, down = done + 1, 0
+            except PlexError as exc:
+                # "Not found" is Plex's answer about this one file (a dead CDN headshot: measured 404,
+                # the CDN itself says 403). Anything else — no answer, a refused token, a 5xx — is
+                # Plex failing, which must neither mark the file nor keep the run going.
+                cause = exc.__cause__
+                if isinstance(cause, httpx.HTTPStatusError) and cause.response.status_code in _DEAD:
+                    dead, down = dead + 1, 0  # Plex answered: it is up, the streak is broken
+                    imgcache.mark_miss(key)
+                else:
+                    plex_failed, down = plex_failed + 1, down + 1
+                log.warning(
+                    "a portrait could not be cached",
+                    extra=logs.event(
+                        "plex.people.portrait_failed",
+                        person_id=pid,
+                        reason=type(exc.__cause__).__name__,
+                    ),
+                )
+    failed = dead + plex_failed
+    return {"cached": done, **({"failed": failed} if failed else {})}

backend/app/plex/query.py

@@ -11,6 +11,7 @@ from sqlalchemy.dialects.postgresql import JSONB
 from sqlalchemy.orm import Session
 
 from app.models import PlexItem, PlexLibrary
+from app.plex.genres import hu_genre
 
 TS_CONFIG = "public.unaccent_simple"
 
@@ -183,7 +184,13 @@ def meta_filter_conds(model, p: dict, lang: str = "en") -> list:
     gsel = split_csv(p.get("genres"))
     if gsel:
         gcol = genres_col(model, lang)
-        gc = [gcol.contains([g]) for g in gsel]
+        if lang == "hu":
+            # The HU column holds one Hungarian word per genre (plex/genres.py), so a filter saved while
+            # it offered "Action" is read as "Akció" too. The word as picked still matches as well: a row
+            # the sync has not reached since keeps its English (or Plex's mixed) list until it does.
+            gc = [or_(*(gcol.contains([w]) for w in dict.fromkeys((hu_genre(g), g)))) for g in gsel]
+        else:
+            gc = [gcol.contains([g]) for g in gsel]
         conds.append(and_(*gc) if p.get("genre_mode") == "all" else or_(*gc))
     crs = split_csv(p.get("content_ratings"))
     if crs:
@@ -298,10 +305,6 @@ class LibraryFilters:
         }
 
 
-def like_escape(s: str) -> str:
-    return s.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
-
-
 # --- Watch-state scope (shared by the /library grid AND the /facets aggregation) ---------------
 #
 # The `show` scope (all | watched | in_progress | unwatched) filters the browse result by per-user

backend/app/plex/rollup.py

@@ -43,7 +43,9 @@ def rollup(cards: list[dict]) -> dict:
     ]
     resume = None
     if inprog:
-        resume = inprog[-1]  # continue the latest-started episode
+        # The LAST in-progress episode in broadcast order — no timestamp is consulted. Home's
+        # `discover.queries.next_episode` picks the same one, so both Resume buttons agree.
+        resume = inprog[-1]
     else:
         resume = next((c for c in cards if c.get("status") != "watched"), None)
     return {

backend/app/plex/stream.py

@@ -18,7 +18,6 @@ segments. The remux is CPU-light (video copy + a tiny audio transcode), so it ru
 CPU-only prod host; only P3's full transcode is CPU-heavy.
 """
 
-import logging
 import os
 import re
 import secrets
@@ -31,15 +30,14 @@ from pathlib import Path
 
 from sqlalchemy.orm import Session as DbSession
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.config import settings
 from app.downloads.edit import encode_thread_cap
 from app.models import PlexItem
 from app.plex import paths
 from app.plex.sync import HW_DECODABLE_CODECS
 
-log = logging.getLogger("siftlode.plex")
+log = logs.get_logger(__name__)
-
 _HLS_ROOT = (
     Path(settings.plex_hls_dir)
     if settings.plex_hls_dir
@@ -62,8 +60,10 @@ _WARMUP_SEG_TIMEOUT_S = 8.0
 # its index for that one. See `segment_name` for why this is a constant and not a comment.
 _VIDEO_VARIANT = "0"
 _SESSION_IDLE_S = 600  # reap a session with no access for this long
-# A session directory is named `{rating_key}_{int(start_s)}_{tag}_{aoff:+.2f}` (see start_session);
+# A session directory is named `{rating_key}_u{user_id}_{int(start_s)}_{tag}_{aoff:+.2f}` (see
-# before multi-audio it was just `{rating_key}_{int(start_s)}`. `sweep_orphans` deletes ONLY names
+# start_session); before per-viewer sessions it was `{rating_key}_{int(start_s)}_{tag}_{aoff:+.2f}`,
+# and before multi-audio just `{rating_key}_{int(start_s)}` — the older shapes stay sweepable, since a
+# restart leaves them behind like any other. `sweep_orphans` deletes ONLY names
 # matching one of those two shapes: `_HLS_ROOT` is admin-configurable (PLEX_HLS_DIR), so it may
 # legitimately point at a shared scratch path we do not own, and a blanket "delete every directory
 # under the root" would take the neighbours with it.
@@ -74,7 +74,9 @@ _SESSION_IDLE_S = 600  # reap a session with no access for this long
 # non-numeric rating_key would merely leave its directory unswept, which is the safe direction to
 # fail. Covered by test_plex_stream.py — this regex is the only thing standing between the sweep
 # and someone else's data.
-_SESSION_DIR_RE = re.compile(r"^\d+_\d+(_[^_]+_[+-]\d+\.\d{2})?$")
+_SESSION_DIR_RE = re.compile(
+    r"^\d+_\d+(_[^_]+_[+-]\d+\.\d{2})?$|^\d+_u\d+_\d+_[^_]+_[+-]\d+\.\d{2}$"
+)
 # When this process started. `sweep_orphans` only deletes directories older than this: everything
 # newer belongs to a session THIS run created, and since `start_session` reuses a directory path
 # for the same (rating_key, offset, tag), "stale name" is not the same as "stale directory".
@@ -90,7 +92,16 @@ _PROCESS_START = time.time()
 _MTIME_GRACE_S = 2.0
 
 _lock = threading.Lock()
-_sessions: dict[str, "HlsSession"] = {}
+# Keyed by VIEWER and item, (user_id, rating_key). Keyed by the item alone, a second account opening
+# the same film `_kill`ed the first one's session: the displaced viewer got a 404, reclaimed at its
+# playhead, and so evicted the other in turn — two people could not watch one film at once. (One
+# account on two devices still shares a session; that case ends in the player's bounded reclaim.)
+_sessions: dict[tuple[int, str], "HlsSession"] = {}
+# How many live sessions one account may hold at once (two devices, or a next item starting while the
+# last one winds down). A new one past it replaces that account's least recently used — whose waiting
+# session request then returns at once, since `wait_for` watches the killed ffmpeg — so one account
+# can neither pile up film-sized scratch copies nor park more than this many request workers.
+_MAX_SESSIONS_PER_USER = 2
 
 
 class HlsSession:
@@ -104,16 +115,17 @@ class HlsSession:
         transcode: bool = False,
         hw: bool = False,
         tonemap: bool = False,
+        user_id: int = 0,
     ):
-        # Unguessable per-session id, carried in the URLs the client is handed. Sessions are keyed
+        # Unguessable per-session id, carried in the URLs the client is handed. A seek-restart
-        # by rating_key alone, so a seek-restart or a SECOND viewer of the same item replaces the
+        # replaces the viewer's registered session for the item — and the old player's next request
-        # registered one — and the old viewer's next request for `seg_12.ts`, resolved by rating_key,
+        # for `seg_12.ts`, resolved by (viewer, rating_key) alone, was answered 200 with the NEW session's twelfth segment, produced from a different offset
-        # was answered 200 with the NEW session's twelfth segment, produced from a different offset
         # under `-copyts`. Wrong bytes, not an error, so the client's reclaim never fired on it
         # (deep-review round 5). A request naming a session that is no longer registered is a 404,
         # which the player recovers from by restarting at the playhead.
         self.id = secrets.token_urlsafe(9)
-        self.key = key
+        self.key = key  # the item's rating_key (what the logs name)
+        self.user_id = user_id  # whose session it is: the registry key is (user_id, key)
         self.dir = directory
         self.proc = proc
         self.transcode = transcode  # video re-encoded (non-h264 source), not stream-copied
@@ -135,6 +147,8 @@ class HlsSession:
         # single index. Derived, not passed, so it cannot disagree with the argv.
         self.entry = _MASTER_PL if self.variants else _INDEX_PL
         self.last_access = time.time()
+        # `artifact_may_arrive` reports a refusing session directory once, not per segment.
+        self.probe_failure_reported = False
 
 
 def _seek_opts(start_s: float) -> list[str]:
@@ -231,8 +245,19 @@ def _probe_video_kind(src: Path) -> str:
             capture_output=True,
             text=True,
             timeout=10,
+            # A non-zero exit is a failed probe (CalledProcessError, a SubprocessError → the branch
+            # below), not an answer: whatever it printed before refusing is not the file's truth.
+            check=True,
+        )
+    except (subprocess.SubprocessError, OSError) as exc:
+        # The safe kind is a real answer, but it is the answer for an 8-bit SDR source — a blind
+        # ffprobe sends every HDR file down it, so playback quietly degrades to a software
+        # transcode with nothing anywhere saying why. The path stays out of the line: it is the
+        # title of whatever is playing.
+        log.info(
+            "video-kind probe failed; assuming the HW-safe pipeline",
+            extra=logs.event("plex.stream.probe_kind_failed", reason=type(exc).__name__),
         )
-    except (subprocess.SubprocessError, OSError):
         return _VK_HWSAFE
     pix_fmt = bits = transfer = ""
     for line in out.stdout.splitlines():
@@ -261,7 +286,14 @@ def _opencl_present() -> bool:
     software libx264. Its own function so it's monkeypatchable in tests, mirroring _render_node_present."""
     try:
         return any(_OPENCL_ICD_DIR.glob("*.icd"))
-    except OSError:
+    except OSError as exc:
+        # A MISSING directory does not come through here — `glob` simply yields nothing for one.
+        # This is a directory we cannot read, and the answer it forces (no OpenCL) costs every HDR
+        # source the hardware tonemap path.
+        log.info(
+            "the OpenCL ICD directory could not be read; treating OpenCL as absent",
+            extra=logs.event("plex.stream.opencl_probe_failed", reason=type(exc).__name__),
+        )
         return False
 
 
@@ -404,7 +436,7 @@ def _select_encode(db: DbSession, item: PlexItem, transcode: bool, src: Path) ->
     source that has an OpenCL runtime, while a 10-bit SDR source drops to software."""
     if not transcode:
         return (False, False)
-    mode = (sysconfig.get_str(db, "plex_transcode_hw") or "auto").lower()
+    mode = sysconfig.get_str(db, "plex_transcode_hw")
     if mode == "off" or not _render_node_present():
         return (False, False)
     if mode == "on":
@@ -447,6 +479,8 @@ def _sw_child_preexec() -> None:
     playback."""
     try:
         os.nice(_SW_TRANSCODE_NICE)
+    # silent-ok: this runs in the forked child, between fork and exec. Logging there can deadlock
+    # on a lock some other thread held at fork time, and the child is about to become ffmpeg.
     except OSError:
         pass
 
@@ -471,9 +505,19 @@ def _probe_audio_langs(src: Path) -> list[str]:
             capture_output=True,
             text=True,
             timeout=10,
+            # A non-zero exit is a failed probe (CalledProcessError, a SubprocessError → the branch
+            # below), not an answer: whatever it printed before refusing is not the file's truth.
+            check=True,
         )
         return [ln.strip() for ln in out.stdout.splitlines()]
-    except (subprocess.SubprocessError, OSError):
+    except (subprocess.SubprocessError, OSError) as exc:
+        # An empty list IS the single-audio path, so a failed probe silently collapses a file with
+        # several audio tracks down to one — and the viewer gets a player with no language picker
+        # rather than an error.
+        log.warning(
+            "audio-track probe failed; falling back to a single-audio session",
+            extra=logs.event("plex.stream.probe_audio_failed", reason=type(exc).__name__),
+        )
         return []
 
 
@@ -577,13 +621,22 @@ def _probe_first_pts(seg: Path) -> float | None:
             capture_output=True,
             text=True,
             timeout=10,
+            # A non-zero exit is a failed probe (CalledProcessError, a SubprocessError → the branch
+            # below), not an answer: whatever it printed before refusing is not the file's truth.
+            check=True,
         )
         for line in out.stdout.splitlines():
             tok = line.strip().strip(",")
             if tok:
                 return float(tok)
-    except (subprocess.SubprocessError, ValueError, OSError):
+    except (subprocess.SubprocessError, ValueError, OSError) as exc:
-        pass
+        # None means "keep the offset the client asked for", and the whole point of this probe is
+        # that the two differ: without it the subtitle track is shifted by whatever the container
+        # started its timestamps at. Seconds of drift, and nothing on screen calls it a failure.
+        log.warning(
+            "first-PTS probe failed; the requested offset is used as-is",
+            extra=logs.event("plex.stream.first_pts_probe_failed", reason=type(exc).__name__),
+        )
     return None
 
 
@@ -592,29 +645,62 @@ def _kill(s: HlsSession) -> None:
         s.proc.terminate()
         try:
             s.proc.wait(timeout=3)
+        # silent-ok: the escalation IS the handling. A transcode that ignores TERM for three
+        # seconds gets KILL, which is exactly what this branch does.
         except subprocess.TimeoutExpired:
             s.proc.kill()
-    except Exception:
+    except Exception as exc:  # noqa: BLE001 — teardown finishes whatever the process does
-        pass
+        # Not nothing: what survives this is an ffmpeg nobody owns any more, holding a VAAPI
+        # handle and a core, and the transcode cap that is supposed to bound them counts rows in
+        # `_sessions` rather than processes. One line is the only way that ever surfaces.
+        log.info(
+            "could not stop the transcode process",
+            extra=logs.event("plex.stream.kill_failed", key=s.key, reason=type(exc).__name__),
+        )
     shutil.rmtree(s.dir, ignore_errors=True)
 
 
-def _enforce_cap(cap: int) -> None:
+def _enforce_caps(user_id: int, max_transcodes: int, max_sessions: int) -> None:
-    # Caller holds _lock. Only TRANSCODE sessions count toward plex_max_transcodes — remux
+    """Caller holds _lock. Three bounds, applied after a session is registered, each dropping the
-    # (stream-copy) and direct-serve are cheap and deliberately uncapped (see config.py:
+    least recently accessed sessions over it:
-    # plex_max_transcodes, and the user-facing config hint). Drop the least-recently-accessed
+
-    # transcode sessions over the cap, leaving any concurrent remux streams untouched.
+    * this account's own sessions (`_MAX_SESSIONS_PER_USER`) — first, so a single account filling the
-    tx = [(key, s) for key, s in _sessions.items() if s.transcode]
+      box evicts itself before anyone else;
-    if len(tx) <= cap:
+    * TRANSCODE sessions (plex_max_transcodes): the CPU/GPU-heavy ones;
-        return
+    * EVERY session (plex_max_sessions). A remux is cheap on the CPU but not on the disk: each one is a
-    for key, s in sorted(tx, key=lambda kv: kv[1].last_access)[: len(tx) - cap]:
+      whole segmented copy of the film under `_HLS_ROOT`, written faster than realtime. Counted only
-        _kill(s)
+      by the transcode cap, N distinct items were N ffmpegs and N film-sized scratch directories.
-        _sessions.pop(key, None)
+
+    Somebody's playback stops at each drop. The route turns the missing session into a retryable 404
+    and the player restarts at the playhead, so from the server nothing failed — and "my film keeps
+    rebuffering" has no other explanation anywhere, hence the line."""
+
+    def cull(pool: list[tuple[tuple[int, str], HlsSession]], cap: int, bound: str) -> None:
+        if len(pool) <= cap:
+            return
+        for reg, sess in sorted(pool, key=lambda kv: kv[1].last_access)[: len(pool) - cap]:
+            log.info(
+                "a session was dropped to stay under a cap",
+                extra=logs.event(
+                    "plex.stream.session_culled",
+                    key=sess.key,
+                    bound=bound,
+                    cap=cap,
+                    sessions=len(pool),
+                ),
+            )
+            _kill(sess)
+            _sessions.pop(reg, None)
+
+    cull([kv for kv in _sessions.items() if kv[0][0] == user_id], _MAX_SESSIONS_PER_USER, "user")
+    cull([kv for kv in _sessions.items() if kv[1].transcode], max_transcodes, "transcode")
+    cull(list(_sessions.items()), max_sessions, "total")
 
 
 def start_session(
     db: DbSession,
     item: PlexItem,
+    user_id: int,
     start_s: float,
     audio_ord: int | None = None,
     aoff: float = 0.0,
@@ -634,6 +720,7 @@ def start_session(
     # session respawn (the HW->SW fallback) must not pick up a mid-playback config change.
     loudnorm = sysconfig.get_bool(db, "plex_audio_loudnorm")
     key = item.rating_key
+    reg = (user_id, key)  # the registry key: one session per viewer per item
     start_s = max(0.0, float(start_s))
     audio_langs = _probe_audio_langs(src) if multi else []
     is_multi = len(audio_langs) > 1  # only worth a master playlist when there really are ≥2 tracks
@@ -646,7 +733,7 @@ def start_session(
     # the tag with a DASH — `_SESSION_DIR_RE` treats the tag as one underscore-free field, and that
     # regex is the only thing keeping the orphan sweep off a shared scratch path (see its comment).
     tag = ("multi" if is_multi else str(audio_ord)) + ("-ln" if loudnorm else "")
-    directory = _HLS_ROOT / f"{key}_{int(start_s)}_{tag}_{aoff:+.2f}"
+    directory = _HLS_ROOT / f"{key}_u{user_id}_{int(start_s)}_{tag}_{aoff:+.2f}"
 
     def _seg(i: int) -> Path:
         return directory / segment_name(i, _VIDEO_VARIANT if variants else None)
@@ -658,13 +745,12 @@ def start_session(
         # check-and-replace happens under ONE _lock, so a newer same-key session (a concurrent
         # seek-restart) installed between the fallback's decision and here is never clobbered. When the
         # registered session is no longer `replace_only`, we abort and return None (the caller yields to
-        # whoever is registered now). The INITIAL spawn passes replace_only=None and always supersedes —
+        # whoever is registered now). The INITIAL spawn never gets None — see its call site.
-        # that is the seek-restart's intended last-writer-wins.
         # Software-fallback encode tuning — read only on the SW transcode path (the VAAPI path and
         # remux ignore preset/crf/max_height, and remux must not touch the DB for them).
         preset, crf, max_height = "veryfast", 20, 0
         if transcode and not hw:
-            preset = sysconfig.get_str(db, "plex_transcode_preset") or "veryfast"
+            preset = sysconfig.get_str(db, "plex_transcode_preset")
             crf = sysconfig.get_int(db, "plex_transcode_crf")
             max_height = sysconfig.get_int(db, "plex_transcode_max_height")
         # Nice the SOFTWARE transcode child (not the HW path, and not remux) so a CPU-bound re-encode
@@ -672,9 +758,9 @@ def start_session(
         # platform without fork (the API always runs in a Linux container in prod) it's simply omitted.
         preexec = _sw_child_preexec if (transcode and not hw and os.name == "posix") else None
         with _lock:
-            if replace_only is not None and _sessions.get(key) is not replace_only:
+            if replace_only is not None and _sessions.get(reg) is not replace_only:
                 return None  # superseded/culled while we waited out of lock — don't clobber the newer one
-            old = _sessions.pop(key, None)
+            old = _sessions.pop(reg, None)
             if old is not None:
                 _kill(old)
             shutil.rmtree(directory, ignore_errors=True)
@@ -708,14 +794,35 @@ def start_session(
                 transcode=transcode,
                 hw=hw,
                 tonemap=tonemap,
+                user_id=user_id,
+            )
+            _sessions[reg] = s
+            # The one line that says which pipeline a session actually got. Everything downstream
+            # — the cap, the HW→SW fallback, the subtitle offset — is about this choice, and until
+            # now it was only inferable from the ffmpeg command nobody keeps.
+            log.info(
+                "an HLS session started",
+                extra=logs.event(
+                    "plex.stream.session_started",
+                    key=key,
+                    transcode=transcode,
+                    hw=hw,
+                    tonemap=tonemap,
+                    variants=len(variants or ()),
+                    start_s=round(start_s, 1),
+                ),
+            )
+            # Admin-configurable caps (Configuration → Plex); at least 1 each so playback can't be
+            # capped to zero.
+            _enforce_caps(
+                user_id,
+                max(1, sysconfig.get_int(db, "plex_max_transcodes")),
+                max(1, sysconfig.get_int(db, "plex_max_sessions")),
             )
-            _sessions[key] = s
-            # Admin-configurable concurrency cap (Configuration → Plex → Max concurrent transcodes);
-            # at least 1 so playback can't be capped to zero.
-            _enforce_cap(max(1, sysconfig.get_int(db, "plex_max_transcodes")))
         return s
 
     s = _spawn(*_select_encode(db, item, transcode, src))
+    assert s is not None  # replace_only=None always supersedes: a seek-restart's last-writer-wins
     # Wait for the first video segment outside the lock (this blocks on ffmpeg, and we must not hold up
     # other sessions). Pass the proc so a HW session that dies at VAAPI init fails fast (not after 25s).
     ready = wait_for(_seg(0), timeout=25.0, proc=s.proc)
@@ -728,7 +835,8 @@ def start_session(
         # ONE lock in _spawn, so we can never tear down the newer session. s.hw is now False, so the route
         # honestly reports "CPU" and the player's Preparing-stream state stays correct.
         log.warning(
-            "plex transcode: HW session produced no segments — falling back to software key=%s", key
+            "the hardware transcode produced no segments; falling back to software",
+            extra=logs.event("plex.stream.hw_fallback", key=key),
         )
         respawned = _spawn(False, False, replace_only=s)
         if respawned is None:
@@ -737,7 +845,7 @@ def start_session(
             # turns a None return into a "Media file not found on disk" 404 — dishonest for a file that
             # exists. A (dead) session instead degrades to a retryable 404 ("No active playback session" /
             # "Not available yet") the client re-requests.
-            return current_session(key) or s
+            return current_session(user_id, key) or s
         s = respawned
         ready = wait_for(_seg(0), timeout=25.0, proc=s.proc)
     if ready:
@@ -768,9 +876,9 @@ def start_session(
     return s
 
 
-def current_session(key: str) -> HlsSession | None:
+def current_session(user_id: int, key: str) -> HlsSession | None:
     with _lock:
-        s = _sessions.get(key)
+        s = _sessions.get((user_id, key))
         if s is not None:
             s.last_access = time.time()
         return s
@@ -867,7 +975,23 @@ def artifact_may_arrive(s: "HlsSession", filename: str) -> bool:
         return True
     try:
         return (s.dir / segment_name(seq - 1, variant)).exists()
-    except OSError:
+    except OSError as exc:
+        # `exists()` answers False for a missing file without raising, so this is the session
+        # directory itself refusing — a full or unmounted disk. Every segment then 404s, which the
+        # player shows as a stall. Once per session: this runs for every segment request and
+        # every hls.js retry of one, and the sink's limiter still stores 30 a minute of a
+        # template — stdout would get them all.
+        if not s.probe_failure_reported:
+            s.probe_failure_reported = True
+            log.warning(
+                "a session segment could not be checked",
+                extra=logs.event(
+                    "plex.stream.segment_probe_failed",
+                    key=s.key,
+                    seq=seq,
+                    reason=type(exc).__name__,
+                ),
+            )
         return False
 
 
@@ -884,12 +1008,16 @@ def wait_for(path: Path, timeout: float = 20.0, proc: subprocess.Popen | None =
         try:
             if path.exists() and path.stat().st_size > 0:
                 return True
+        # silent-ok: this is a 150 ms poll — a line here is a line every 150 ms for the whole
+        # timeout. The final check below decides, and the 404 the caller then returns is honest.
         except OSError:
             pass
         if proc is not None and proc.poll() is not None:
             # the producer exited — no more segments are coming; one last check then give up.
             try:
                 return path.exists() and path.stat().st_size > 0
+            # silent-ok: one last look after the producer exited. False here becomes a 404 the
+            # player recovers from, which is the honest answer either way.
             except OSError:
                 return False
         time.sleep(0.15)
@@ -922,7 +1050,16 @@ def sweep_orphans(older_than: float | None = None) -> int:
     dropped = 0
     try:
         entries = list(_HLS_ROOT.iterdir())
-    except OSError:  # the root doesn't exist yet — nothing to sweep
+    # silent-ok: nothing has been transcoded yet, so there is nothing to sweep.
+    except FileNotFoundError:
+        return 0
+    except OSError as exc:
+        # Any OTHER reason the root will not list is not "nothing to sweep": the sweep reports 0
+        # like a clean run while every crashed session's segments stay on disk forever.
+        log.warning(
+            "the HLS root could not be listed; no orphans were swept",
+            extra=logs.event("plex.stream.sweep_root_unreadable", reason=type(exc).__name__),
+        )
         return 0
     for p in entries:
         if not _SESSION_DIR_RE.match(p.name):
@@ -932,7 +1069,9 @@ def sweep_orphans(older_than: float | None = None) -> int:
         try:
             if p.stat().st_mtime >= cutoff:
                 continue  # this run made (or reused) it — never ours to delete
-        except OSError:  # vanished under us (a concurrent restart of that session)
+        # silent-ok: vanished under us — a concurrent restart of that session removed it, which
+        # is the outcome this sweep wanted anyway.
+        except OSError:
             continue
         with _lock:
             if any(s.dir == p for s in _sessions.values()):
@@ -968,21 +1107,21 @@ def reap_idle(finished_idle_s: float = _FINISHED_IDLE_S) -> int:
     now = time.time()
     dropped = 0
     with _lock:
-        for key, s in list(_sessions.items()):
+        for reg, s in list(_sessions.items()):
             done = s.proc.poll() is not None
             idle = now - s.last_access
             if idle > _SESSION_IDLE_S or (done and idle > finished_idle_s):
                 _kill(s)
-                _sessions.pop(key, None)
+                _sessions.pop(reg, None)
                 dropped += 1
     return dropped
 
 
-def touch_session(key: str) -> None:
+def touch_session(user_id: int, key: str) -> None:
-    """Say a viewer is still there for `key`, without fetching anything. `last_access` used to move
+    """Say this viewer is still there for `key`, without fetching anything. `last_access` used to move
     only on an HLS file fetch, which a paused player does not do — so the only way to keep a paused
     viewer's session alive was to keep EVERY abandoned one just as long."""
     with _lock:
-        s = _sessions.get(key)
+        s = _sessions.get((user_id, key))
         if s is not None:
             s.last_access = time.time()

backend/app/plex/subtitles.py

@@ -32,18 +32,20 @@ window still took the full 145 s and emitted the whole track), so windows are ex
 zero-based clock and shifted back to absolute time here with `shift_vtt`.
 """
 
+import codecs
 import json
-import logging
 import re
 import subprocess
 import tempfile
 import threading
 import time
+import logging
 from concurrent.futures import ThreadPoolExecutor
 from pathlib import Path
 
 from sqlalchemy.orm import Session
 
+from app import logs
 from app.db import SessionLocal
 from app.models import PlexItem
 from app.plex import paths as plex_paths
@@ -51,8 +53,7 @@ from app.plex.cache import CACHE_DIR
 from app.plex.client import PlexClient, PlexError
 from app.plex.lookups import by_rating_key
 
-log = logging.getLogger("siftlode.plex")
+log = logs.get_logger(__name__)
-
 # Image-based subtitle codecs (Blu-ray/DVD) — can't be turned into text WebVTT (would need burn-in).
 BITMAP_SUBS = {
     "pgs",
@@ -104,6 +105,7 @@ def shift_vtt(vtt: str, offset: float) -> str:
     straddling 0 is clamped to start at 0."""
     if not offset:
         return vtt
+    unshifted = 0
     out_blocks: list[str] = []
     for block in re.split(r"\n[ \t]*\n", vtt):
         if "-->" not in block:
@@ -121,7 +123,12 @@ def shift_vtt(vtt: str, offset: float) -> str:
                 st = _vtt_ts_to_s(left.strip()) - offset
                 en = _vtt_ts_to_s(rparts[0]) - offset
             except (ValueError, IndexError):
-                break  # unparseable timing → leave the block as-is
+                # silent-ok: left as-is means left UNSHIFTED — the viewer sees that one cue at
+                # the wrong moment. Counted rather than logged here: this is per cue, and the file
+                # being parsed is one ffmpeg wrote for us a moment ago, so a broken timestamp is
+                # a fact about the extraction, reported once below.
+                unshifted += 1
+                break
             if en <= 0:
                 keep = False  # whole cue is before the session start → drop the block
             else:
@@ -131,6 +138,11 @@ def shift_vtt(vtt: str, offset: float) -> str:
             break
         if keep:
             out_blocks.append("\n".join(lines))
+    if unshifted:
+        log.info(
+            "some cues could not be shifted and stay at their original time",
+            extra=logs.event("plex.subtitles.cues_unshifted", cues=unshifted, offset_s=offset),
+        )
     return "\n\n".join(out_blocks)
 
 
@@ -150,7 +162,8 @@ def ffmpeg_to_vtt(src: str, sub_ord: int | None) -> str:
         "webvtt",
         "pipe:1",
     ]
-    r = subprocess.run(cmd, capture_output=True, timeout=EXTRACT_TIMEOUT_S)
+    # check=False: the exit status is read on the next line, with ffmpeg's own stderr as the reason.
+    r = subprocess.run(cmd, capture_output=True, timeout=EXTRACT_TIMEOUT_S, check=False)
     if r.returncode != 0 or not r.stdout:
         raise PlexError(
             (r.stderr or b"").decode("utf-8", "replace")[:200] or "ffmpeg subtitle convert failed"
@@ -162,7 +175,15 @@ def cache_read(key: str) -> str | None:
     try:
         p = CACHE_DIR / f"{key}.vtt"
         return p.read_text(encoding="utf-8") if p.exists() else None
-    except OSError:
+    except OSError as exc:
+        # None is the same answer as "never cached", so a cache we cannot read looks exactly like
+        # a cold one — and every request re-extracts instead of saying the disk is the problem.
+        log.warning(
+            "the cached subtitle could not be read",
+            extra=logs.event(
+                "plex.subtitles.cache_read_failed", key=key, reason=type(exc).__name__
+            ),
+        )
         return None
 
 
@@ -175,7 +196,16 @@ def cache_write(key: str, vtt: str) -> bool:
         tmp.write_text(vtt, encoding="utf-8")
         tmp.replace(CACHE_DIR / f"{key}.vtt")  # atomic publish
         return True
-    except OSError:
+    except OSError as exc:
+        # `publish` turns False into "the region was not claimed" and moves on, so the track just
+        # never appears. Which errno it was — full disk, read-only mount, permissions — is the
+        # whole diagnosis, and it exists only here.
+        log.warning(
+            "the subtitle cache could not be written",
+            extra=logs.event(
+                "plex.subtitles.cache_write_failed", key=key, reason=type(exc).__name__
+            ),
+        )
         return False
 
 
@@ -189,7 +219,11 @@ _FULL_END_S = 1_000_000_000.0  # stand-in for "to the end of the file" (JSON can
 
 
 def cache_key(rating_key: str, sub_ord: int) -> str:
-    return f"sub_{rating_key}_{sub_ord}"
+    # Versioned whenever `decode_subtitle` changes what a track decodes to — `sub2_` when legacy code
+    # pages were cached as U+FFFD, `sub3_` when a stray byte, an untagged file or Serbian still went
+    # to the wrong code page. The cache has no eviction, so every track is read again once rather
+    # than a broken one being served for good; the older files are just left behind.
+    return f"sub3_{rating_key}_{sub_ord}"
 
 
 def _vtt_path(key: str) -> Path:
@@ -206,7 +240,17 @@ def read_meta(key: str) -> tuple[list[list[float]], bool]:
     try:
         raw = json.loads(_meta_path(key).read_text(encoding="utf-8"))
         return [[float(a), float(b)] for a, b in raw.get("ranges", [])], bool(raw.get("complete"))
-    except (OSError, ValueError, TypeError):
+    # silent-ok: no sidecar at all is the documented older-build case the fallback exists for.
+    except FileNotFoundError:
+        return ([[0.0, _FULL_END_S]], True) if _exists(_vtt_path(key)) else ([], False)
+    except (OSError, ValueError, TypeError) as exc:
+        # A sidecar that EXISTS and will not parse is different, and the fallback is the dangerous
+        # answer to it: beside a cached VTT it reports the track as complete, so a partial track is
+        # never extended again and the viewer loses every cue past the window that was cached.
+        log.warning(
+            "the subtitle sidecar could not be read; the cached track is taken as complete",
+            extra=logs.event("plex.subtitles.meta_unreadable", key=key, reason=type(exc).__name__),
+        )
         return ([[0.0, _FULL_END_S]], True) if _exists(_vtt_path(key)) else ([], False)
 
 
@@ -216,6 +260,8 @@ def _exists(p: Path) -> bool:
     escaped as a 500 from the status route."""
     try:
         return p.exists()
+    # silent-ok: this function's whole job is to answer False for a path the OS refuses. The
+    # refusal IS the return value; a caller that needs to report one reports it itself.
     except (OSError, ValueError):
         return False
 
@@ -226,8 +272,18 @@ def _write_meta(key: str, ranges: list[list[float]], complete: bool) -> None:
         tmp = CACHE_DIR / f"{key}.json.tmp"
         tmp.write_text(json.dumps({"ranges": ranges, "complete": complete}), encoding="utf-8")
         tmp.replace(_meta_path(key))
-    except OSError:
+    except OSError as exc:
-        pass
+        # Without the sidecar beside it, a cached window reads back as a COMPLETE track (see
+        # `read_meta`) — so one failed write here permanently costs the rest of that subtitle.
+        log.warning(
+            "the subtitle sidecar could not be written",
+            extra=logs.event(
+                "plex.subtitles.meta_write_failed",
+                key=key,
+                complete=complete,
+                reason=type(exc).__name__,
+            ),
+        )
 
 
 def generation(key: str) -> int:
@@ -235,7 +291,9 @@ def generation(key: str) -> int:
     window is superseded by a wider one the browser refetches instead of keeping the short VTT."""
     try:
         return _vtt_path(key).stat().st_mtime_ns
-    except (OSError, ValueError):  # ValueError: a NUL in the key — `stat` refuses the path outright
+    # silent-ok: 0 for a track that is not cached yet, which is the ordinary case on every first
+    # request. (ValueError: a NUL in the key — `stat` refuses the path outright.)
+    except (OSError, ValueError):
         return 0
 
 
@@ -267,16 +325,27 @@ def merge_vtt(*sources: str) -> str:
     happen to share a time (two speakers, a sign over dialogue) are both kept."""
     seen: set[str] = set()
     cues: list[tuple[float, str]] = []
+    dropped = 0
     for src in sources:
         for block in re.split(r"\n[ \t]*\n", src or ""):
-            canon = _canonical_cue(block.strip())
+            stripped = block.strip()
+            canon = _canonical_cue(stripped)
             if canon is None:
+                # A block WITH a timing line that still would not parse is a cue we are throwing
+                # away, not a header — and the merged track would simply be missing those lines.
+                if "-->" in stripped:
+                    dropped += 1
                 continue  # header / NOTE / STYLE — the merged file gets its own single header
             start, key, text = canon
             if key in seen:
                 continue
             seen.add(key)
             cues.append((start, text))
+    if dropped:
+        log.info(
+            "some cues were dropped from the merged subtitle: their timing would not parse",
+            extra=logs.event("plex.subtitles.cues_dropped", cues=dropped, kept=len(cues)),
+        )
     cues.sort(key=lambda c: c[0])
     return "WEBVTT\n\n" + "\n\n".join(b for _, b in cues) + "\n"
 
@@ -295,6 +364,8 @@ def _canonical_cue(block: str) -> tuple[float, str, str] | None:
         try:
             st = _vtt_ts_to_s(left.strip())
             en = _vtt_ts_to_s(rparts[0])
+        # silent-ok: None here means "not a cue", and the caller tells a header apart from a
+        # broken timing line by whether the block had a `-->` at all — see `merge_vtt`.
         except (ValueError, IndexError):
             return None
         lines[i] = f"{_vtt_s_to_ts(st)} --> {_vtt_s_to_ts(en)}{settings}"
@@ -310,6 +381,8 @@ def _last_cue_end(vtt: str) -> float:
             continue
         try:
             end = max(end, _vtt_ts_to_s(ln.partition("-->")[2].strip().split(None, 1)[0]))
+        # silent-ok: this is a floor — "how far the pass provably got". Skipping a line it cannot
+        # read makes the answer smaller, never larger, so the error is always on the safe side.
         except (ValueError, IndexError):
             continue
     return end
@@ -394,11 +467,18 @@ _FULL_POOL = ThreadPoolExecutor(max_workers=2, thread_name_prefix="plex-subs-ful
 # which time something has finished.
 _MAX_WINDOWS_INFLIGHT = 8
 _MAX_FULL_INFLIGHT = 6
+# …of which one account may hold this many. The full-pass slots are minutes each, and a hostile (or
+# merely curious) account opening subtitles on film after film could otherwise hold all of them while
+# everyone else's tracks never completed. A viewer over their share still gets windows — the part
+# that makes a track watchable now — just not another whole-file read until one of theirs ends.
+_MAX_FULL_PER_USER = 3
 _MAX_TRACK_WINDOWS = 2
 
 _LOCK = threading.Lock()
 # Queued or running jobs → the absolute span a window job reads (None for a full pass).
 _INFLIGHT: dict[tuple, tuple[float, float] | None] = {}
+# rating_key → the account whose poll queued that item's full pass (see _MAX_FULL_PER_USER).
+_FULL_OWNERS: dict[str, int] = {}
 # Track-level failures: a track that cannot be served at all (no item, a bitmap track, Plex down…).
 _FAILURES: dict[str, tuple[str, float]] = {}
 # rating_key → when its full pass last failed (see FULL_RETRY_S).
@@ -448,6 +528,7 @@ def reset_state() -> None:
     global _STOPPING
     with _LOCK:
         _INFLIGHT.clear()
+        _FULL_OWNERS.clear()
         _FAILURES.clear()
         _FULL_FAILED.clear()
         _WIN_FAILED.clear()
@@ -531,7 +612,7 @@ def _window_failed(key: str, at: float) -> bool:
         return any(a <= at <= b for a, b, _ in _live_window_failures(key))
 
 
-def ensure(rating_key: str, sub_ord: int, at: float = 0.0) -> dict:
+def ensure(rating_key: str, sub_ord: int, at: float = 0.0, viewer: int = 0) -> dict:
     """State of one subtitle track, scheduling whatever extraction is still missing. NEVER blocks.
 
     ``ready`` = fully extracted · ``partial`` = usable, covers ``at``, still growing · ``preparing``
@@ -574,8 +655,23 @@ def ensure(rating_key: str, sub_ord: int, at: float = 0.0) -> dict:
             if edge - at <= LOOKAHEAD_S and edge < _FULL_END_S:
                 probe = edge + 1.0
                 want_window = not _window_failed(key, probe)
-        _schedule(rating_key, sub_ord, probe, want_window=want_window, want_full=not full_failed)
+        _schedule(
-    return {"state": "partial" if covered else "preparing", "code": None, "gen": gen}
+            rating_key,
+            sub_ord,
+            probe,
+            want_window=want_window,
+            want_full=not full_failed,
+            viewer=viewer,
+        )
+    # `growing`: whether the track will fill in on its own. False once the full pass has failed and
+    # is backed off — a partial track then stays partial until a window is asked for, so the player
+    # may poll slowly (only to read ahead of an edge) instead of every two seconds for the whole film.
+    return {
+        "state": "partial" if covered else "preparing",
+        "code": None,
+        "gen": gen,
+        "growing": not full_failed,
+    }
 
 
 def cached_track(rating_key: str, sub_ord: int) -> tuple[str | None, str | None]:
@@ -592,7 +688,12 @@ def cached_track(rating_key: str, sub_ord: int) -> tuple[str | None, str | None]
 
 
 def _schedule(
-    rating_key: str, sub_ord: int, at: float, want_window: bool, want_full: bool = True
+    rating_key: str,
+    sub_ord: int,
+    at: float,
+    want_window: bool,
+    want_full: bool = True,
+    viewer: int = 0,
 ) -> None:
     """Kick off the jobs this track still needs: the window that makes it watchable now, and the one
     full pass that makes it complete. A 1 s status poll (or a second viewer) joins a job already
@@ -625,8 +726,15 @@ def _schedule(
                 ident = ("win", rating_key, sub_ord, w_start)
                 _INFLIGHT[ident] = (w_start, w_start + WINDOW_PRE_S + WINDOW_LEN_S)
                 jobs.append((ident, w_start))
-        if want_full and n_full < _MAX_FULL_INFLIGHT and ("full", rating_key) not in _INFLIGHT:
+        mine = sum(1 for owner in _FULL_OWNERS.values() if owner == viewer)
+        if (
+            want_full
+            and n_full < _MAX_FULL_INFLIGHT
+            and ("full", rating_key) not in _INFLIGHT
+            and mine < _MAX_FULL_PER_USER
+        ):
             _INFLIGHT[("full", rating_key)] = None
+            _FULL_OWNERS[rating_key] = viewer
             jobs.append((("full", rating_key), None))
     for ident, window_start in jobs:
         pool = _FULL_POOL if window_start is None else _WIN_POOL
@@ -652,12 +760,35 @@ def _run(
             return
         _extract_and_publish(rating_key, sub_ord, window_start)
     except SubtitleUnavailable as e:
+        # The player's own state was the only place this ever showed. `_pass_failed` backs the item
+        # off for an hour and `_fail` records a code, but neither leaves a line, so "my subtitles
+        # never appear" had nothing on the server to read.
+        # A final code is a fact about the file (no such track, a bitmap subtitle); anything else
+        # is this instance failing at something it could do.
+        log.log(
+            logging.INFO if e.code in FINAL_CODES else logging.WARNING,
+            "subtitle extraction gave up",
+            extra=logs.event(
+                "plex.subtitles.pass_failed",
+                rating_key=rating_key,
+                sub_ord=sub_ord,
+                code=e.code,
+                windowed=window_start is not None,
+            ),
+        )
         if e.code == "extract_failed":
             _pass_failed(rating_key, [key], window_start)
         else:
             _fail(key, e.code)
     except PlexError as e:
-        log.warning("subtitle extraction failed for %s/%s: %s", rating_key, sub_ord, e)
+        # No traceback on purpose: a PlexError carries httpx's own message, which names the method,
+        # the URL and the status — the stack under it is this `except` and the client call, and
+        # says nothing the message does not. (The URL holds no token; it travels as a header.)
+        log.warning(
+            "subtitle extraction failed talking to Plex: %s",
+            e,
+            extra=logs.event("plex.subtitles.plex_error", rating_key=rating_key, sub_ord=sub_ord),
+        )
         _fail(key, "plex_error")
     except (
         Exception
@@ -667,6 +798,8 @@ def _run(
     finally:
         with _LOCK:
             _INFLIGHT.pop(ident, None)
+            if ident[0] == "full":
+                _FULL_OWNERS.pop(rating_key, None)
 
 
 def _pass_failed(
@@ -732,10 +865,13 @@ def _extract_and_publish(rating_key: str, sub_ord: int, window_start: float | No
             for i, s in enumerate(subs)
             if s.get("index") is not None and not _is_bitmap(s)
         }
-    start = window_start or 0.0
+    # A full pass starts where every track's leading coverage ends: a pass that ran out of time
+    # (`EXTRACT_TIMEOUT_S`) kept what it reached, and restarting from 0 each `FULL_RETRY_S` meant a file
+    # too big to read in one go never completed. Windows handle an offset start already; so does this.
+    start = window_start if window_start is not None else _full_pass_start(rating_key, targets)
     length = (WINDOW_LEN_S + WINDOW_PRE_S) if window_start is not None else None
     end = _FULL_END_S if length is None else start + length
-    texts, clean = _ffmpeg_extract(str(src), sorted(targets), window_start, length)
+    texts, clean = _ffmpeg_extract(str(src), sorted(targets), start or None, length)
     for emb_ord, vtt in texts.items():
         key = cache_key(rating_key, targets[emb_ord])
         # A pass that exited non-zero holds only what ffmpeg reached before it died. Claim no more
@@ -756,15 +892,108 @@ def _extract_and_publish(rating_key: str, sub_ord: int, window_start: float | No
             _pass_failed(rating_key, [key], window_start, (covered_end, end))
 
 
+def _full_pass_start(rating_key: str, targets: dict[int, int]) -> float:
+    """Where the full pass may begin: the end of the span from 0 that EVERY target track already has
+    (one read feeds them all, so the least-covered track decides), backed off by `WINDOW_PRE_S` so a
+    cue running across the join is read whole — `merge_vtt` folds the overlap. 0 when any track has
+    nothing from the start."""
+    ends = []
+    for ord_ in targets.values():
+        key = cache_key(rating_key, ord_)
+        ranges, _ = read_meta(key) if _exists(_vtt_path(key)) else ([], False)
+        lead = [b for a, b in ranges if a <= 0.0]
+        if not lead:
+            return 0.0
+        ends.append(max(lead))
+    if not ends:
+        return 0.0
+    return max(0.0, float(f"{min(ends) - WINDOW_PRE_S:.3f}"))
+
+
+# The legacy code page a subtitle in this language (Plex's ISO 639-2 `languageCode`) was most likely
+# saved in. Serbian sits here, not with Cyrillic: its releases are overwhelmingly Latin script in
+# Windows-1250, like Croatian and Bosnian.
+_CP1250 = {
+    "hun",
+    "ces",
+    "cze",
+    "pol",
+    "slk",
+    "slo",
+    "slv",
+    "hrv",
+    "bos",
+    "srp",
+    "ron",
+    "rum",
+    "sqi",
+    "alb",
+}
+_CP1251 = {"rus", "ukr", "bel", "bul", "mkd", "mac"}
+# Western European languages keep Windows-1252 (cp1250 has no à/è/ñ). An UNTAGGED file falls to
+# Windows-1250: its ASCII and á/é/í/ó/ö/ú/ü match Windows-1252 byte for byte, and it also gets ő/ű
+# right, which Windows-1252 turns into õ/û — the library this reads is mostly Hungarian.
+_CP1252 = {
+    "eng",
+    "fre",
+    "fra",
+    "ger",
+    "deu",
+    "spa",
+    "ita",
+    "por",
+    "dut",
+    "nld",
+    "swe",
+    "dan",
+    "nor",
+    "fin",
+    "ice",
+    "isl",
+    "cat",
+}
+# A file that is almost all valid UTF-8 is UTF-8 with a few stray bytes (a pasted Windows-1252 quote),
+# not a legacy file: accents in a legacy code page are nearly never valid UTF-8 sequences, so a real
+# legacy file has hardly any. At least this many valid multi-byte characters per invalid byte.
+_UTF8_VALID_PER_STRAY = 10
+
+
+def decode_subtitle(raw: bytes, language_code: str | None) -> str:
+    """A sidecar subtitle file as text. SubRip carries no charset, and a Hungarian release is
+    typically Windows-1250: decoded as UTF-8 with `replace`, every ő/ű/é came out as U+FFFD (found
+    live, 1391 of them in one episode). So: a BOM decides; else valid UTF-8 stays UTF-8 (a legacy
+    code page with accents is almost never valid UTF-8), and so does NEARLY valid UTF-8 — one stray
+    byte must not send a whole UTF-8 file to a legacy code page; else the language's legacy code page."""
+    if raw.startswith(codecs.BOM_UTF8):
+        return raw[len(codecs.BOM_UTF8) :].decode("utf-8", "replace")
+    if raw.startswith((codecs.BOM_UTF16_LE, codecs.BOM_UTF16_BE)):
+        return raw.decode("utf-16", "replace")
+    try:
+        return raw.decode("utf-8")
+    except UnicodeDecodeError:
+        # silent-ok: not strictly UTF-8 — nearly-UTF-8 or a legacy code page, told apart below.
+        pass
+    lenient = raw.decode("utf-8", "replace")
+    stray = lenient.count("�")
+    valid = sum(1 for ch in lenient if ch > "" and ch != "�")
+    if valid >= stray * _UTF8_VALID_PER_STRAY:
+        return lenient
+    lang = (language_code or "").lower()
+    legacy = "cp1251" if lang in _CP1251 else "cp1252" if lang in _CP1252 else "cp1250"
+    return raw.decode(legacy, "replace")
+
+
 def _sidecar_vtt(db: Session, stream: dict) -> str:
-    """An external subtitle file, fetched through Plex and converted to WebVTT."""
+    """An external subtitle file, fetched through Plex and converted to WebVTT. Its bytes are made
+    UTF-8 first (`decode_subtitle`), on the ffmpeg path too: ffmpeg reads a text subtitle as UTF-8."""
     with PlexClient(db) as plex:
         raw = plex.raw_get(stream["key"])
+    text = decode_subtitle(raw, stream.get("languageCode"))
     codec = (stream.get("codec") or stream.get("format") or "").lower()
     if codec in ("srt", "subrip", ""):
-        return srt_to_vtt(raw.decode("utf-8", "replace"))
+        return srt_to_vtt(text)
     with tempfile.NamedTemporaryFile(suffix="." + (codec or "sub"), delete=False) as fh:
-        fh.write(raw)
+        fh.write(text.encode("utf-8"))
         tmp = fh.name
     try:
         return ffmpeg_to_vtt(tmp, None)
@@ -794,9 +1023,10 @@ def _ffmpeg_extract(
             cmd += ["-map", f"0:s:{emb_ord}"]
             if length is not None:
                 cmd += ["-t", f"{length:.3f}"]
+            if start:
                 # The input seek lands on the keyframe at or BEFORE `start`, so a cue already running
                 # at `start` comes out at a negative time — and by default the muxer then shifts the
-                # WHOLE output so that cue sits at 0: every cue of the window late by the distance to
+                # WHOLE output so that cue sits at 0: every cue of the read late by the distance to
                 # that keyframe (1–5 s measured on a real library file, a whole GOP on a synthetic
                 # one), and doubled for good once the full pass merged in beside it. Kept negative, the
                 # times are exact on the zero-based clock (`_vtt_ts_to_s` reads the sign).
@@ -829,6 +1059,16 @@ def _ffmpeg_extract(
                 try:
                     _, err = proc.communicate(timeout=_TERM_GRACE_S)
                 except subprocess.TimeoutExpired:
+                    # TERM is what makes a timed-out pass worth keeping: ffmpeg closes its outputs
+                    # on it. An ffmpeg that ignores TERM leaves only whole blocks already flushed,
+                    # so the partial track is short for a reason the later line cannot tell apart.
+                    log.info(
+                        "the subtitle extraction ignored TERM and was killed",
+                        extra=logs.event(
+                            "plex.subtitles.extract_killed_after_term",
+                            windowed=start is not None,
+                        ),
+                    )
                     proc.kill()
                     _, err = proc.communicate()
         finally:

backend/app/plex/sync.py

@@ -15,20 +15,19 @@ dead `enabled` column) is a separate follow-up. Cast + intro/credit markers are
 item-detail/player path (P2), not here.
 """
 
-import logging
 from datetime import datetime, timezone
 
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.models import PlexCollection, PlexItem, PlexLibrary, PlexSeason, PlexShow
-from app.plex import detail, people
+from app.plex import detail, overlay, people
 from app.plex.client import PlexClient, PlexError, PlexNotConfigured
+from app.plex.genres import hu_genres
 from app.plex.lookups import by_rating_key
 
-log = logging.getLogger("siftlode.plex")
+log = logs.get_logger(__name__)
-
 # Codecs/containers a browser can play in a native <video> without any transcoding.
 _BROWSER_VIDEO = {"h264"}
 _BROWSER_AUDIO = {"aac", "mp3", "opus", "vorbis"}
@@ -102,6 +101,9 @@ def _video_stream_props(meta: dict) -> tuple[int | None, bool]:
                 bd = s.get("bitDepth")
                 try:
                     depth = int(bd) if bd else None
+                # silent-ok: unknown bit depth, which is also what most listings report. The
+                # playback class then falls back to the codec alone, and per-stream noise over a
+                # whole library would drown the sync's own lines.
                 except (TypeError, ValueError):
                     depth = None
                 hdr = str(s.get("colorTrc") or "").lower() in _HDR_TRANSFERS
@@ -112,6 +114,8 @@ def _video_stream_props(meta: dict) -> tuple[int | None, bool]:
 def _epoch(v) -> datetime | None:
     try:
         return datetime.fromtimestamp(int(v), tz=timezone.utc) if v else None
+    # silent-ok: a field Plex leaves out of most listing rows. None is a legitimate stored value,
+    # and every caller treats it as "not known" rather than as an error.
     except (TypeError, ValueError, OSError):
         return None
 
@@ -120,6 +124,7 @@ def _date(v):
     """Parse a Plex 'YYYY-MM-DD' string (originallyAvailableAt) → date, or None."""
     try:
         return datetime.strptime(v, "%Y-%m-%d").date() if v else None
+    # silent-ok: as `_epoch` above — an absent or unparsable air date is stored as no air date.
     except (TypeError, ValueError):
         return None
 
@@ -135,6 +140,7 @@ def _tags(meta: dict, key: str, limit: int | None = None) -> list | None:
 def _rating(meta: dict) -> float | None:
     try:
         return float(meta.get("audienceRating") or meta.get("rating"))
+    # silent-ok: TypeError on `float(None)` is the ordinary path — most items carry no rating.
     except (TypeError, ValueError):
         return None
 
@@ -187,20 +193,6 @@ def _enabled_section_keys(db: Session) -> set[str] | None:
     return keys or None
 
 
-def _overlay_map(db: Session) -> dict[str, str]:
-    """The configured HU-overlay pairing: {huSectionKey: enSectionKey} from `plex_overlay_map`
-    ('huKey:enKey,...'). Each HU section is synced as a non-browsable overlay onto its EN twin.
-    Malformed pairs are skipped."""
-    raw = sysconfig.get_str(db, "plex_overlay_map")
-    out: dict[str, str] = {}
-    for pair in raw.split(","):
-        hu, sep, en = pair.partition(":")
-        hu, en = hu.strip(), en.strip()
-        if sep and hu and en:
-            out[hu] = en
-    return out
-
-
 def sync(db: Session) -> dict:
     """Full reconcile of the enabled movie/show sections. Returns a small stats dict."""
     if not sysconfig.get_bool(db, "plex_enabled"):
@@ -216,7 +208,7 @@ def sync(db: Session) -> dict:
         "pruned": 0,
     }
     wanted = _enabled_section_keys(db)
-    overlay_map = _overlay_map(db)  # {huKey: enKey}
+    overlay_map = overlay.overlay_map(db)  # {huKey: enKey}
     overlay_keys = set(overlay_map)
     try:
         with PlexClient(db) as plex:
@@ -260,21 +252,25 @@ def sync(db: Session) -> dict:
                 lib.metadata_lang = "hu"
                 lib.overlay_of = en_lib.id
                 db.flush()
-                _sync_overlay(db, plex, lib, en_lib.id, stats)
+                overlay.sync_overlay(db, plex, lib, en_lib.id, stats)
                 db.commit()
             # Rich-fetch a bounded batch of not-yet-enriched titles' FULL cast, so the whole library
             # becomes filterable over successive runs (the cheap listing above only stored the top ~3).
             try:
                 stats["people_enriched"] = _enrich_people_batch(db, plex)
-            except Exception as e:  # best-effort — never fail the sync over the backfill
+            except Exception:  # best-effort; never fail the sync over a backfill
-                log.warning("Plex people enrich batch failed: %s", e)
+                # With the traceback: a blind catch over a batch that fetches AND writes can be
+                # a Plex shape change, a DB error or our own bug, and `%s` told them apart badly.
+                log.exception("the Plex people-enrich batch failed")
             # Probe a bounded batch of HW-decodable transcode items for bit-depth/HDR, so the
             # playback-class filter's HW/SW split is precise (a 10-bit/HDR HEVC → software) over
             # successive runs. Only HW-decodable-codec transcodes are probed — the rest are SW by codec.
             try:
                 stats["media_probed"] = _probe_media_batch(db, plex)
-            except Exception as e:  # best-effort — never fail the sync over the backfill
+            except Exception:  # best-effort; never fail the sync over a backfill
-                log.warning("Plex media probe batch failed: %s", e)
+                log.exception("the Plex media-probe batch failed")
+    # silent-ok: the return value IS the report. `{"skipped": …}` goes to the scheduler, which
+    # logs the job result, and to the admin route, which answers it to the caller.
     except PlexNotConfigured as e:
         return {"skipped": str(e)}
     except PlexError as e:
@@ -285,8 +281,8 @@ def sync(db: Session) -> dict:
     # appear in ANY item's cast/directors — drop those rows, mirroring how Plex drops orphaned metadata.
     try:
         stats["people_pruned"] = people.prune_orphan_people(db)
-    except Exception as e:  # best-effort housekeeping — never fail the sync over it
+    except Exception:  # best-effort housekeeping; never fail the sync over it
-        log.warning("Plex people prune failed: %s", e)
+        log.exception("the Plex people prune failed")
     log.info("Plex sync done: %s", stats)
     return stats
 
@@ -325,7 +321,9 @@ def _probe_media_batch(db: Session, plex: PlexClient, limit: int = _PROBE_BATCH)
     for row in rows:
         try:
             meta = plex.metadata(row.rating_key) or {}
-        except Exception as e:  # transient — leave media_probed_at NULL so it retries next run
+        # No traceback and only debug: this is per item over a whole library, and the stack is
+        # the same `plex.metadata` call every time. The retry next run is the handling.
+        except Exception as e:  # noqa: BLE001 — transient; leave media_probed_at NULL to retry
             log.debug("Plex media probe fetch failed for %s: %s", row.rating_key, e)
             continue
         bit_depth, hdr = _video_stream_props(meta)
@@ -372,7 +370,8 @@ def _enrich_people_batch(db: Session, plex: PlexClient, limit: int = _ENRICH_BAT
     for row in [*movies, *shows]:
         try:
             meta = plex.metadata(row.rating_key) or {}
-        except Exception as e:
+        # As in `_probe_media_batch` above: per item, so the message without the stack.
+        except Exception as e:  # noqa: BLE001 — transient; leave people_enriched_at NULL
             log.debug("Plex people enrich fetch failed for %s: %s", row.rating_key, e)
             continue  # leave people_enriched_at NULL so it's retried on a later run
         people.enrich_row_people(db, row, people.cast_from_meta(meta))
@@ -419,7 +418,10 @@ def _paginate(plex: PlexClient, key: str, item_type: int, progress: dict | None
     if progress is not None:
         # A truthy `total` we reached = a trustworthy complete listing. `total` unknown/zero-with-rows
         # (a Plex that doesn't report totalSize) is treated as NOT complete, so prune stays conservative.
+        # `total` itself is kept for a caller to whom a server that SAID 0 is news it can trust (the
+        # watch reconcile); an empty mirror or overlay must not read it so (E32 S2).
         progress["complete"] = bool(total) and start >= total
+        progress["total"] = total
 
 
 # The section listing exposes at most the top ~20 cast (Plex's own cap); a stored cast longer than this
@@ -436,6 +438,10 @@ def _apply_facet_fields(row, meta: dict) -> None:
     row.studio = (meta.get("studio") or None) and str(meta.get("studio"))[:255]
     row.originally_available_at = _date(meta.get("originallyAvailableAt"))
     row.genres = _tags(meta, "Genre")
+    # Every row has Hungarian genres, not only one with an HU overlay twin (the overlay pass, which runs
+    # after this, puts the twin's own list over it): the HU genre facets and filter read this column,
+    # and an English word there could never match a filter read in Hungarian.
+    row.genres_hu = hu_genres(row.genres)
     row.directors = _tags(meta, "Director")
     # The listing carries only the top ~20 cast, so re-applying it would DOWNGRADE an already-enriched
     # full cast (up to 24, grown by the rich backfill/info-page enrich), silently breaking the
@@ -643,59 +649,6 @@ def _sync_shows(db: Session, plex: PlexClient, lib: PlexLibrary, stats: dict) ->
     _prune_stale(db, shows, shows_seen, stats, complete=shows_progress.get("complete", False))
 
 
-def _sync_overlay(
-    db: Session, plex: PlexClient, lib: PlexLibrary, en_lib_id: int, stats: dict
-) -> None:
-    """Write an overlay section's Hungarian metadata onto its EN twin's rows, matched by guid. Creates
-    NO browsable rows — an overlay library is never listed on its own. Only title/summary/genres differ
-    by language; the language-neutral people_text feeds both search vectors.
-
-    This is a full RECONCILE: every guid-bearing EN row's HU fields are cleared first, then the HU
-    section repopulates the matches — so an EN row whose HU twin vanished (or was re-guid'd) falls back
-    to EN instead of keeping stale Hungarian metadata. HU items with no EN match are skipped."""
-    if lib.kind == "movie":
-        rows = db.execute(
-            select(PlexItem).where(
-                PlexItem.library_id == en_lib_id,
-                PlexItem.kind == "movie",
-                PlexItem.guid.is_not(None),
-            )
-        ).scalars()
-        item_type = _TYPE_MOVIE
-    else:
-        rows = db.execute(
-            select(PlexShow).where(PlexShow.library_id == en_lib_id, PlexShow.guid.is_not(None))
-        ).scalars()
-        item_type = _TYPE_SHOW
-    # guid -> the EN rows carrying it (a list: two editions of one film can share a guid; both overlay).
-    by_guid: dict[str, list] = {}
-    for r in rows:
-        by_guid.setdefault(r.guid, []).append(r)
-    # Reconcile step: clear HU fields on all guid-bearing rows. Only rows that actually held HU data
-    # become dirty (SQLAlchemy skips a no-net-change UPDATE), so this is cheap; matches are re-set below.
-    for group in by_guid.values():
-        for row in group:
-            row.title_hu = row.summary_hu = row.genres_hu = None
-    seen: set[str] = (
-        set()
-    )  # first HU entry for a guid wins; a dup HU item can't overwrite or double-count
-    for meta in _paginate(plex, lib.plex_key, item_type):
-        g = _canonical_guid(meta)
-        targets = by_guid.get(g) if g else None
-        if not targets or g in seen:
-            continue
-        seen.add(g)
-        title_hu = (meta.get("title") or "").strip() or None
-        summary_hu = meta.get("summary")
-        genres_hu = _tags(meta, "Genre")
-        for row in targets:
-            row.title_hu = title_hu
-            row.summary_hu = summary_hu
-            row.genres_hu = genres_hu
-        stats["overlaid"] = stats.get("overlaid", 0) + 1
-    db.flush()
-
-
 def _upsert_collection(
     db: Session, existing: PlexCollection | None, rk: str, lib_id: int, meta: dict
 ) -> PlexCollection:
@@ -721,9 +674,11 @@ def resync_collection(db: Session, plex: PlexClient, lib: PlexLibrary, rk: str)
     meta = plex.metadata(rk) or {}
     col = _upsert_collection(db, by_rating_key(db, PlexCollection, rk), rk, lib.id, meta)
     Model = PlexItem if lib.kind == "movie" else PlexShow
-    new_members = {
+    children = _believed_children(plex, rk, meta)
-        str(c.get("ratingKey")) for c in plex.collection_children(rk) if c.get("ratingKey")
+    if children is None:
-    }
+        db.flush()
+        return col
+    new_members = {str(c.get("ratingKey")) for c in children if c.get("ratingKey")}
     current = db.execute(select(Model).where(Model.collection_keys.contains([rk]))).scalars().all()
     for row in current:  # drop rk from rows that are no longer members
         if row.rating_key not in new_members:
@@ -736,6 +691,22 @@ def resync_collection(db: Session, plex: PlexClient, lib: PlexLibrary, rk: str)
     return col
 
 
+def _believed_children(plex: PlexClient, rk: str, meta: dict) -> list[dict] | None:
+    """A collection's members, or None when the answer is not believed: no members from a collection
+    whose own listing counts some (`childCount`) is a glitch, not a collection emptied — the same
+    empty-answer rule as the collection listing's (E32 review). Rebuilding on it showed the
+    collection empty until the next sync."""
+    children = plex.collection_children(rk)
+    count = str(meta.get("childCount") or "0")
+    if children or not count.isdigit() or int(count) == 0:
+        return children
+    log.warning(
+        "Plex listed no members for a non-empty collection; keeping its membership",
+        extra=logs.event("plex.sync.collection_children_empty", collection=rk),
+    )
+    return None
+
+
 def delete_collection_local(db: Session, lib: PlexLibrary, rk: str) -> None:
     """Local cleanup after a collection is deleted from Plex: strip rk from members + drop the row."""
     Model = PlexItem if lib.kind == "movie" else PlexShow
@@ -747,6 +718,19 @@ def delete_collection_local(db: Session, lib: PlexLibrary, rk: str) -> None:
     db.flush()
 
 
+def _rebuild_membership(
+    db: Session, lib: PlexLibrary, membership: dict[str, list[str]], kept: set[str]
+) -> None:
+    """Rebuild collection_keys on every member row of this library (clear + set), carrying over the
+    keys of a collection whose empty answer was not believed (`_believed_children`)."""
+    for Model in (PlexItem, PlexShow):
+        for row in db.execute(select(Model).where(Model.library_id == lib.id)).scalars():
+            held = [k for k in (row.collection_keys or []) if k in kept]
+            keys = [*membership.get(row.rating_key, []), *held]
+            row.collection_keys = list(dict.fromkeys(keys)) or None
+    db.flush()
+
+
 def _sync_collections(db: Session, plex: PlexClient, lib: PlexLibrary, stats: dict) -> None:
     """Mirror a library's collections + membership. Reads never touch Plex afterwards. Membership is
     written as `collection_keys` on member movies (plex_items) / shows (plex_shows). A full reconcile:
@@ -758,7 +742,17 @@ def _sync_collections(db: Session, plex: PlexClient, lib: PlexLibrary, stats: di
             select(PlexCollection).where(PlexCollection.library_id == lib.id)
         ).scalars()
     }
+    if not cols and existing:
+        # The catalog's own rule (see the module doc): an empty listing while we hold rows is a Plex
+        # glitch, not a library emptied of collections. Reconciling on it dropped every collection
+        # (its user-set `editable` flag with it) and every member's keys (E32 S2).
+        log.warning(
+            "Plex listed no collections for a library that has some; keeping them this run",
+            extra=logs.event("plex.sync.collections_empty", library=lib.plex_key),
+        )
+        return
     seen: set[str] = set()
+    kept: set[str] = set()  # collections whose empty children answer is not believed
     membership: dict[str, list[str]] = {}  # member rating_key -> [collection rating_key, ...]
     for meta in cols:
         rk = str(meta.get("ratingKey") or "")
@@ -766,7 +760,10 @@ def _sync_collections(db: Session, plex: PlexClient, lib: PlexLibrary, stats: di
             continue
         _upsert_collection(db, existing.get(rk), rk, lib.id, meta)
         seen.add(rk)
-        for child in plex.collection_children(rk):
+        children = _believed_children(plex, rk, meta)
+        if children is None:
+            kept.add(rk)
+        for child in children or []:
             crk = str(child.get("ratingKey") or "")
             if crk:
                 membership.setdefault(crk, []).append(rk)
@@ -775,9 +772,4 @@ def _sync_collections(db: Session, plex: PlexClient, lib: PlexLibrary, stats: di
     for rk, col in existing.items():
         if rk not in seen:
             db.delete(col)
-    # Rebuild collection_keys on every member row of this library (clear + set).
+    _rebuild_membership(db, lib, membership, kept)
-    for row in db.execute(select(PlexItem).where(PlexItem.library_id == lib.id)).scalars():
-        row.collection_keys = membership.get(row.rating_key) or None
-    for row in db.execute(select(PlexShow).where(PlexShow.library_id == lib.id)).scalars():
-        row.collection_keys = membership.get(row.rating_key) or None
-    db.flush()

backend/app/plex/tmdb.py

@@ -10,21 +10,23 @@ One-parser rule (mirrors app/plex/detail.py): `parse_tmdb` is the pure parser, `
 the row. The enrich batch + scheduler entry live here too so the whole TMDB integration is one module.
 """
 
-import logging
+import re
 import time
 from datetime import datetime, timezone
+from typing import cast
 
 import httpx
 from sqlalchemy import func, select
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.orm import Session
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.config import settings
 from app.models import PlexItem, PlexPerson, PlexShow
+from app.plex.cards import display_title
+from app.plex.people import people_by_credits, warm_portraits
 
-log = logging.getLogger("siftlode.tmdb")
+log = logs.get_logger(__name__)
-
 # Politeness throttle: TMDB tolerates ~50 req/s; ~33/s leaves comfortable headroom for a shared host.
 _MIN_INTERVAL = 0.03
 
@@ -126,9 +128,17 @@ class TmdbClient:
         """Resolve an external id (imdb_id / tvdb_id) to TMDB items via /find."""
         return self._get(f"/find/{external_id}", {"external_source": source})
 
+    def countries(self) -> list[dict] | None:
+        """TMDB's country list: `[{iso_3166_1, english_name, native_name}]` (a JSON array, which
+        `_get`'s dict annotation does not describe)."""
+        return cast("list[dict] | None", self._get("/configuration/countries"))
+
     def fetch_person(self, person_id: int) -> dict | None:
-        """One person's bio + combined credits (E5 S6), in a single append_to_response request."""
+        """One person's bio + combined credits (E5 S6) + the bio's translations (E30 S4), in a single
-        return self._get(f"/person/{person_id}", {"append_to_response": "combined_credits"})
+        append_to_response request."""
+        return self._get(
+            f"/person/{person_id}", {"append_to_response": "combined_credits,translations"}
+        )
 
 
 def guid_to_tmdb_id(guid: str | None, media_type: str, tmdb: TmdbClient) -> int | None:
@@ -142,6 +152,12 @@ def guid_to_tmdb_id(guid: str | None, media_type: str, tmdb: TmdbClient) -> int
         try:
             return int(ext)
         except ValueError:
+            # The caller stamps the watermark on a None, so this title is never looked at again:
+            # one unreadable guid costs it its artwork and its cast permanently.
+            log.info(
+                "a Plex guid named TMDB but did not carry a numeric id",
+                extra=logs.event("plex.tmdb.guid_unparsable", provider=provider),
+            )
             return None
     source = {"imdb": "imdb_id", "tvdb": "tvdb_id"}.get(provider)
     if not source:
@@ -250,8 +266,12 @@ def parse_tmdb(payload: dict, media_type: str) -> dict:
             c.get("name") for c in (payload.get("production_companies") or []) if c.get("name")
         ][:8]
         or None,
+        # The ISO code beside TMDB's English name (E30): the client names the country in the reader's
+        # language from the code; the name is the fallback for a code it cannot name.
         "production_countries": [
-            c.get("name") for c in (payload.get("production_countries") or []) if c.get("name")
+            {"code": c.get("iso_3166_1") or None, "name": c["name"]}
+            for c in (payload.get("production_countries") or [])
+            if c.get("name")
         ]
         or None,
         "recommendations": recs or None,
@@ -297,8 +317,8 @@ def _link_people(db: Session, links: dict[str, int]) -> None:
 
 
 def enrich_batch(db: Session, tmdb: TmdbClient, limit: int) -> int:
-    """Enrich up to `limit` not-yet-TMDB-enriched titles (movies first, shows fill the remainder —
+    """Enrich up to `limit` not-yet-TMDB-enriched titles (movies first, shows fill the
-    mirrors _enrich_people_batch). A transient TmdbError leaves the watermark NULL for retry; a
+    remainder). A transient TmdbError leaves the watermark NULL for retry; a
     definitive no-match (unresolvable guid / TMDB 404) still stamps the watermark so it isn't re-fetched
     every run. One commit at the end. Returns the number processed (matched OR definitively unmatched)."""
     movies = (
@@ -352,13 +372,62 @@ def enrich_batch(db: Session, tmdb: TmdbClient, limit: int) -> int:
     return processed
 
 
+def _country(c: dict | str) -> dict:
+    """One stored production country as served: `{code, name}`. A row enriched before E30 stored the
+    bare English name; until `backfill_country_codes` reaches it, it serves with no code (the client
+    shows the name)."""
+    return c if isinstance(c, dict) else {"code": None, "name": c}
+
+
+def backfill_country_codes(db: Session, tmdb: TmdbClient) -> int:
+    """Give the countries stored before E30 (bare English names) their ISO codes, from TMDB's own
+    country list — the names came from that list, so they match it exactly. One request, and only
+    while such a row is left: re-fetching each title instead would take the enrichment batch days.
+    A name the list does not hold keeps no code. Returns the number of rows rewritten."""
+    stale = [
+        row
+        for model in (PlexItem, PlexShow)
+        for row in db.execute(
+            select(model).where(func.jsonb_typeof(model.production_countries[0]) == "string")
+        ).scalars()
+    ]
+    if not stale:
+        return 0
+    codes = {
+        c["english_name"]: c["iso_3166_1"]
+        for c in tmdb.countries() or []
+        if c.get("english_name") and c.get("iso_3166_1")
+    }
+    for row in stale:
+        row.production_countries = [
+            {"code": codes.get(c), "name": c} if isinstance(c, str) else c
+            for c in row.production_countries or []
+        ]
+    db.commit()
+    return len(stale)
+
+
 def run_enrich(db: Session) -> dict:
     """Scheduler entry point. No-op (never an error) when the TMDB key isn't configured."""
     if not sysconfig.get_str(db, "tmdb_api_key"):
         return {"enriched": 0, "skipped": "tmdb_api_key not set"}
     with TmdbClient(db) as tmdb:
+        coded = backfill_country_codes(db, tmdb)
         n = enrich_batch(db, tmdb, settings.tmdb_enrich_batch)
-    return {"enriched": n}
+    return {"enriched": n, **({"countries_coded": coded} if coded else {})}
+
+
+def run_people_enrich(db: Session) -> dict:
+    """Scheduler entry point of people_enrich (E30 S6): a batch of person bios (the most-credited
+    first) and a batch of portraits. Each half is a no-op — never an error — without its TMDB key /
+    without Plex."""
+    batch = settings.people_enrich_batch
+    if sysconfig.get_str(db, "tmdb_api_key"):
+        with TmdbClient(db) as tmdb:
+            bios = enrich_people_batch(db, tmdb, batch)
+    else:
+        bios = {"skipped": "tmdb_api_key not set"}
+    return {"bios": bios, "portraits": warm_portraits(db, batch)}
 
 
 def resolve_library_guids(
@@ -402,10 +471,11 @@ def resolve_library_guids(
     return movies, shows
 
 
-def resolve_recs(db: Session, recs: list[dict] | None) -> list[dict]:
+def resolve_recs(db: Session, recs: list[dict] | None, *, lang: str) -> list[dict]:
     """Tag each recommendation with whether it's in OUR library (matched by guid `tmdb:<id>`, bucketed
     by media_type — see resolve_library_guids) and, if so, its rating_key — so the client can deep-link
-    the ones we have and merely display the rest."""
+    the ones we have and merely display the rest. One we have is named as the library names it, in
+    `lang` (E30); one we lack keeps the title TMDB gave."""
     if not recs:
         return []
     movies, shows = resolve_library_guids(db, recs)
@@ -413,13 +483,19 @@ def resolve_recs(db: Session, recs: list[dict] | None) -> list[dict]:
     for r in recs:
         g = f"tmdb:{r.get('tmdb_id')}"
         row = shows.get(g) if r.get("media_type") == "tv" else movies.get(g)
+        named = {"title": display_title(row, lang)} if row is not None else {}
         out.append(
-            {**r, "in_library": row is not None, "rating_key": row.rating_key if row else None}
+            {
+                **r,
+                **named,
+                "in_library": row is not None,
+                "rating_key": row.rating_key if row else None,
+            }
         )
     return out
 
 
-def serve_block(db: Session, row) -> dict | None:
+def serve_block(db: Session, row, *, lang: str) -> dict | None:
     """The TMDB extras for a movie/show detail response, served from the row (no live TMDB call).
     None until the row has been TMDB-enriched (the client then shows nothing extra)."""
     if row.tmdb_enriched_at is None:
@@ -433,30 +509,58 @@ def serve_block(db: Session, row) -> dict | None:
         "keywords": row.keywords or [],
         "crew": row.crew or [],
         "studios": row.studios or [],
-        "production_countries": row.production_countries or [],
+        "production_countries": [_country(c) for c in row.production_countries or []],
-        "recommendations": resolve_recs(db, row.recommendations),
+        "recommendations": resolve_recs(db, row.recommendations, lang=lang),
         "backdrops": row.backdrops or [],
     }
 
 
 # --- Person enrichment (E5 S6) -------------------------------------------------------------------
-# A person's bio + birth facts + a "Known for" credit pool, fetched once (the first person-page open)
+# A person's bio + birth facts + a "Known for" credit pool, fetched once — by the people_enrich job, or
-# and served from the plex_people row. `known_for` stores the top combined-credits ids; the person
+# by the first person-page open if the job has not reached them yet — and served from the plex_people row. `known_for` stores the top combined-credits ids; the person
 # endpoint resolves them to library rows by guid at serve time (like recommendations).
 
-_KNOWN_FOR_MAX = 15
+# The candidate pool kept per person. Only the titles that are in the library are shown (at most
+# routes.plex.people._KNOWN_FOR_SHOWN), so the pool is deeper than the rail: a star's 15 most-voted
+# credits are often not the 15 a given library holds.
+_KNOWN_FOR_MAX = 40
+# TMDB genre ids of the shows a person visits as a guest (News, Reality, Talk). Ranked by popularity,
+# talk-show visits crowded out the films (Tom Cruise: 14 of 15 were talk shows; E30 S6).
+_GUEST_GENRES = {10763, 10764, 10767}
+# A credit in which the person plays themselves ("Self", "Himself - Guest", "Herself (archive footage)").
+_SELF_CREDIT = re.compile(r"^\s*(self|himself|herself|themselves)\b", re.IGNORECASE)
+# The crew jobs that make a title the person's own work (a director is known for what they directed).
+_KNOWN_FOR_CREW_JOBS = {"Director", "Creator"}
+
+
+def _known_for_credit(c: dict, crew: bool) -> bool:
+    if crew:
+        return c.get("job") in _KNOWN_FOR_CREW_JOBS
+    if _SELF_CREDIT.match(c.get("character") or ""):
+        return False
+    return not _GUEST_GENRES.intersection(c.get("genre_ids") or [])
 
 
 def parse_person(payload: dict) -> dict:
     """TMDB /person/{id} (+combined_credits) → the persisted bio fields. `known_for` = the person's
-    most-popular movie/tv credits (deduped by id), top `_KNOWN_FOR_MAX`, resolved to library later."""
+    most-voted movie/tv credits — roles and directed works, not guest visits as themselves — deduped,
-    credits = (payload.get("combined_credits") or {}).get("cast") or []
+    the top `_KNOWN_FOR_MAX`, resolved to library rows at serve time. Ranked by vote count, as TMDB's
+    own "Known for" is: popularity is this week's attention, which a talk-show visit wins."""
+    combined = payload.get("combined_credits") or {}
+    credits = [c for c in combined.get("cast") or [] if _known_for_credit(c, crew=False)] + [
+        c for c in combined.get("crew") or [] if _known_for_credit(c, crew=True)
+    ]
     # Dedup on (id, media_type), not the bare id: TMDB movie ids and tv ids are independent sequences
     # (id 603 can be both a movie and a show — see resolve_library_guids), so a same-id movie+show pair
     # is two distinct credits and both must survive.
     seen: set[tuple[int, str]] = set()
     known: list[dict] = []
-    for c in sorted(credits, key=lambda c: c.get("popularity") or 0, reverse=True):
+    ranked = sorted(
+        credits,
+        key=lambda c: (c.get("vote_count") or 0, c.get("popularity") or 0),
+        reverse=True,
+    )
+    for c in ranked:
         tid, mt = c.get("id"), c.get("media_type")
         if not tid or mt not in ("movie", "tv") or (tid, mt) in seen:
             continue
@@ -466,6 +570,7 @@ def parse_person(payload: dict) -> dict:
             break
     return {
         "biography": (payload.get("biography") or "").strip() or None,
+        "biography_hu": _translated_bio(payload, "hu"),
         "birthday": payload.get("birthday") or None,
         "deathday": payload.get("deathday") or None,
         "place_of_birth": payload.get("place_of_birth") or None,
@@ -475,8 +580,17 @@ def parse_person(payload: dict) -> dict:
     }
 
 
+def _translated_bio(payload: dict, lang: str) -> str | None:
+    """The `lang` biography from the appended `translations` (TMDB leaves an untranslated one blank)."""
+    for tr in (payload.get("translations") or {}).get("translations") or []:
+        if tr.get("iso_639_1") == lang:
+            return ((tr.get("data") or {}).get("biography") or "").strip() or None
+    return None
+
+
 def apply_person(row: PlexPerson, parsed: dict) -> None:
     row.biography = parsed["biography"]
+    row.biography_hu = parsed["biography_hu"]
     row.birthday = parsed["birthday"]
     row.deathday = parsed["deathday"]
     row.place_of_birth = parsed["place_of_birth"]
@@ -486,11 +600,22 @@ def apply_person(row: PlexPerson, parsed: dict) -> None:
     row.tmdb_enriched_at = datetime.now(timezone.utc)
 
 
+def _fetch_person_into(tmdb: TmdbClient, person: PlexPerson) -> None:
+    """Fetch one person from TMDB and write the row (not committed). A 404 (deleted person) stamps the
+    watermark with empty fields so it isn't fetched again. Raises TmdbError on a transient failure,
+    leaving the row untouched and `tmdb_enriched_at` NULL for a retry."""
+    payload = tmdb.fetch_person(cast(int, person.tmdb_person_id))
+    if payload is None:
+        person.tmdb_enriched_at = datetime.now(timezone.utc)
+    else:
+        apply_person(person, parse_person(payload))
+
+
 def enrich_person(db: Session, person: PlexPerson) -> bool:
-    """Lazily fetch + persist one person's TMDB bio (first person-page open), committing the write so it
+    """Lazily fetch + persist one person's TMDB bio (a person page opened before the background job
-    survives regardless of the read path. No-op (False) without a key or a `tmdb_person_id`; a transient
+    reached them), committing the write so it survives regardless of the read path. No-op (False)
-    error leaves `tmdb_enriched_at` NULL so the next open retries. A 404 (deleted person) stamps the
+    without a key or a `tmdb_person_id`; a transient error leaves `tmdb_enriched_at` NULL so the next
-    watermark with empty fields so we don't refetch it every open."""
+    open retries."""
     if not person.tmdb_person_id or not sysconfig.get_str(db, "tmdb_api_key"):
         return False
     try:
@@ -498,12 +623,54 @@ def enrich_person(db: Session, person: PlexPerson) -> bool:
         # longer blocks a serve; still cap the outbound TMDB call with a short timeout so a hanging TMDB
         # doesn't tie up a threadpool worker for the default 30s.
         with TmdbClient(db, timeout=8.0) as tmdb:
-            payload = tmdb.fetch_person(person.tmdb_person_id)
+            _fetch_person_into(tmdb, person)
-    except TmdbError:
+    except TmdbError as exc:
+        # False leaves the person un-enriched and unstamped, so it is retried — but the page in
+        # front of the user right now shows an empty biography with nothing to say why.
+        log.warning(
+            "the TMDB person lookup failed",
+            extra=logs.event(
+                "plex.tmdb.person_enrich_failed",
+                person_id=person.id,
+                reason=type(exc).__name__,
+            ),
+        )
         return False
-    if payload is None:
-        person.tmdb_enriched_at = datetime.now(timezone.utc)
-    else:
-        apply_person(person, parse_person(payload))
     db.commit()
     return True
+
+
+# Lookups failing this many people in a row end a people batch: TMDB itself is degraded, and hammering
+# it helps nobody. One person failing alone is skipped, so an id TMDB keeps refusing never stalls the rest.
+_PEOPLE_FAILS_IN_A_ROW = 3
+
+
+def enrich_people_batch(db: Session, tmdb: TmdbClient, limit: int) -> dict:
+    """Fetch up to `limit` linked, not-yet-enriched people (E30 S6), the most-credited first, so a
+    person page reads its bio from the row instead of waiting on TMDB. Each person is committed as it
+    lands: a page open during the run already sees it. A failed lookup is skipped (left unstamped for the
+    next run); _PEOPLE_FAILS_IN_A_ROW of them in a row end the batch."""
+    ids = people_by_credits(db, "bio", limit)
+    done = failed = in_a_row = 0
+    for pid in ids:
+        if in_a_row >= _PEOPLE_FAILS_IN_A_ROW:
+            break
+        person = db.get(PlexPerson, pid)
+        # A page open may have enriched this person since the select: skip, don't fetch twice.
+        if person is None or person.tmdb_enriched_at is not None:
+            continue
+        try:
+            _fetch_person_into(tmdb, person)
+        except TmdbError as exc:
+            db.rollback()
+            failed, in_a_row = failed + 1, in_a_row + 1
+            log.warning(
+                "a TMDB person lookup in the batch failed",
+                extra=logs.event(
+                    "plex.tmdb.people_batch_failed", person_id=pid, reason=type(exc).__name__
+                ),
+            )
+            continue
+        db.commit()
+        done, in_a_row = done + 1, 0
+    return {"enriched": done, **({"failed": failed} if failed else {})}

backend/app/plex/watch_sync.py

@@ -17,16 +17,19 @@ and the incremental pull resolves a genuine divergence to Plex's state (see `_pu
 running a last-write-wins comparison.
 """
 
-import logging
+# On the PlexError lines in this module: none of them carries a traceback, deliberately. A
+# PlexError is httpx's own message — the method, the URL and the status of one call — and the stack
+# beneath it is always this `except` plus `PlexClient._get`. The attrs beside the line say which
+# user and which phase, which is the part the message cannot know.
 from datetime import datetime, timezone
 
 from sqlalchemy import func, select, update
 from sqlalchemy.dialects.postgresql import insert as pg_insert
 from sqlalchemy.orm import Session
 
-from app import sysconfig
+from app import logs, sysconfig
 from app.db import SessionLocal
-from app.models import PlexItem, PlexLink, PlexState
+from app.models import PlexItem, PlexLibrary, PlexLink, PlexState
 from app.plex.client import PlexClient, PlexError, PlexNotConfigured
 from app.plex.sync import (
     _TYPE_EPISODE,
@@ -36,8 +39,7 @@ from app.plex.sync import (
     _paginate,
 )
 
-log = logging.getLogger("siftlode.plex.watch")
+log = logs.get_logger(__name__)
-
 # A resume position below this many seconds isn't worth mirroring — matches the player's own
 # _PROGRESS_MIN_S (a stray few-second seek is not "in progress").
 _PROGRESS_MIN_S = 5
@@ -70,13 +72,27 @@ def _rk_to_id(db: Session) -> dict[str, int]:
 
 def _scan_plex_states(
     db: Session, plex: PlexClient
-) -> tuple[dict[int, tuple[str, int, datetime | None, datetime | None]], int]:
+) -> tuple[dict[int, tuple[str, int, datetime | None, datetime | None]], int, set[str], set[str]]:
     """Scan the enabled movie/show sections and return ``{item_id: (status, pos, watched_at,
-    prog_at)}`` for every mirrored leaf Plex holds a watch signal for, plus the scanned-leaf count.
+    prog_at)}`` for every mirrored leaf Plex holds a watch signal for, the scanned-leaf count, the
-    Shared by the one-time import and the full reconcile (both need Plex's whole current picture)."""
+    keys of the sections SETTLED — walked to the server's total (or said empty by a server we mirror
+    nothing from) — and the keys of every section walked.
+    Shared by the one-time import and the full reconcile (both need Plex's whole current picture).
+    The settled keys matter to the reconcile alone: it reads "absent from the scan" as "Plex
+    un-watched it", which a walk cut short by a Plex hiccup does not mean (see `_paginate`). Settled
+    is decided PER SECTION: one section that never settles (a library emptied on Plex while its
+    mirrored items are kept) must not stop un-watches syncing in all the others (E32 review)."""
     item_id_by_rk = _rk_to_id(db)
     wanted = _enabled_section_keys(db)
+    # The sections we mirror something from: a total of 0 from one of those contradicts the mirror.
+    mirrored = set(
+        db.execute(
+            select(PlexLibrary.plex_key).join(PlexItem, PlexItem.library_id == PlexLibrary.id)
+        ).scalars()
+    )
     scanned = 0
+    settled: set[str] = set()
+    walked: set[str] = set()
     out: dict[int, tuple[str, int, datetime | None, datetime | None]] = {}
     for s in plex.sections():
         stype = s.get("type")
@@ -86,7 +102,10 @@ def _scan_plex_states(
         if wanted is not None and key not in wanted:
             continue
         item_type = _TYPE_MOVIE if stype == "movie" else _TYPE_EPISODE
-        for meta in _paginate(plex, key, item_type):
+        walked.add(key)
+        progress: dict = {}
+        before = scanned
+        for meta in _paginate(plex, key, item_type, progress):
             scanned += 1
             item_id = item_id_by_rk.get(str(meta.get("ratingKey") or ""))
             if item_id is None:
@@ -96,7 +115,15 @@ def _scan_plex_states(
             )
             if res is not None:
                 out[item_id] = res
-    return out, scanned
+        # A section the server SAYS is empty (total 0, nothing listed) settles too. Believed only for
+        # a section we mirror nothing from: one whose items we hold answering 0 is a Plex that lost
+        # the library for a moment, not one that un-watched all of it (E32 review) — so it is left
+        # unsettled, and its rows are not cleared.
+        listed = scanned > before
+        said_empty = progress.get("total") == 0 and not listed and key not in mirrored
+        if (progress.get("complete") and listed) or said_empty:
+            settled.add(key)
+    return out, scanned, settled, walked
 
 
 def import_owner_watch_state(db: Session, link: PlexLink) -> dict:
@@ -109,7 +136,8 @@ def import_owner_watch_state(db: Session, link: PlexLink) -> dict:
     stats = {"watched": 0, "in_progress": 0, "scanned": 0}
     try:
         with PlexClient(db) as plex:
-            scanned_states, stats["scanned"] = _scan_plex_states(db, plex)
+            # A partial scan is still safe here: the import only ever ADDS what Plex reports.
+            scanned_states, stats["scanned"], _settled, _walked = _scan_plex_states(db, plex)
             # One row per item (keyed by item_id → inherently de-duped) then bulk-UPSERT. An upsert is
             # idempotent AND concurrency-safe: a repeated/overlapping enable (the toggle firing twice
             # while the multi-second import runs) can't raise a duplicate-key error — it re-updates.
@@ -148,6 +176,8 @@ def import_owner_watch_state(db: Session, link: PlexLink) -> dict:
             link.initial_import_done = True
             link.last_watch_sync_at = datetime.now(timezone.utc)
             db.commit()
+    # silent-ok: the `{"skipped": …}` return is the report — the admin route answers it and the
+    # scheduler logs it on the job line.
     except PlexNotConfigured as e:
         return {"skipped": str(e)}
     except PlexError as e:
@@ -326,8 +356,18 @@ def _resolve_account_id(db: Session, plex: PlexClient, link: PlexLink) -> int:
             if int(a.get("id") or -1) == 1:
                 username = a.get("name")
                 break
-    except PlexError:
+    except PlexError as exc:
-        pass
+        # The fallback account id 1 is right for the admin-token setup, so sync keeps working —
+        # but the username is never stored, and the UI then has no name to show for the link,
+        # exactly as if the server had not offered one.
+        log.info(
+            "the Plex account list could not be read; the link keeps the default account",
+            extra=logs.event(
+                "plex.watch.accounts_probe_failed",
+                user_id=link.user_id,
+                reason=type(exc).__name__,
+            ),
+        )
     link.plex_account_id = acct_id
     if username:
         link.plex_username = username
@@ -566,6 +606,7 @@ def run_plex_watch_sync(db: Session) -> dict:
                 total["repushed"] += _repush_dirty(db, plex, link)
                 link.last_watch_sync_at = now
                 db.commit()
+        # silent-ok: the `{"skipped": …}` return is the report, as above.
         except PlexNotConfigured as e:
             return {"skipped": str(e)}
         except PlexError as e:
@@ -575,6 +616,28 @@ def run_plex_watch_sync(db: Session) -> dict:
     return total
 
 
+def _clear_unwatched(db: Session, user_id: int, plex_states: dict, settled: set[str]) -> int:
+    """Clear the user's rows Plex no longer has a record for — the mirrored ones (Plex un-watched
+    them); hidden and dirty rows stay (a dirty one is re-pushed by `_repush_dirty`). Only rows of the
+    sections the scan SETTLED: a walk a Plex hiccup cut short leaves out rows it merely failed to
+    reach, and clearing those un-watched them here until the next complete run (E32 S2); an item of
+    a library the allowlist leaves out is absent from every scan, and was cleared on every run (E32
+    review)."""
+    n = 0
+    rows = db.execute(
+        select(PlexState)
+        .join(PlexItem, PlexItem.id == PlexState.item_id)
+        .join(PlexLibrary, PlexLibrary.id == PlexItem.library_id)
+        .where(PlexState.user_id == user_id, PlexLibrary.plex_key.in_(settled))
+    )
+    for st in rows.scalars().all():
+        if st.item_id in plex_states or st.status == "hidden" or not st.synced_to_plex:
+            continue
+        db.delete(st)
+        n += 1
+    return n
+
+
 def run_plex_watch_reconcile(db: Session) -> dict:
     """Scheduler job (default daily) — the authoritative full reconcile. Rescans every section and,
     per item, uses `synced_to_plex` to settle what the incremental feed can't:
@@ -593,7 +656,7 @@ def run_plex_watch_reconcile(db: Session) -> dict:
         try:
             with PlexClient(db) as plex:
                 _resolve_account_id(db, plex, link)
-                plex_states, scanned = _scan_plex_states(db, plex)
+                plex_states, scanned, settled, walked = _scan_plex_states(db, plex)
                 total["scanned"] += scanned
 
                 for item_id, (status, pos, watched_at, prog_at) in plex_states.items():
@@ -601,22 +664,24 @@ def run_plex_watch_reconcile(db: Session) -> dict:
                     if ch:
                         total["applied"] += 1
 
-                # Siftlode rows Plex no longer has a record for: clear the ones that were mirrored
+                total["cleared"] += _clear_unwatched(db, link.user_id, plex_states, settled)
-                # (Plex un-watched them); leave hidden + dirty (dirty gets re-pushed just below).
+                if unsettled := sorted(walked - settled):
-                for st in (
+                    total["incomplete"] = total.get("incomplete", 0) + 1
-                    db.execute(select(PlexState).where(PlexState.user_id == link.user_id))
+                    log.warning(
-                    .scalars()
+                        "Plex sections %s were cut short or answered empty; their rows are not cleared"
-                    .all()
+                        " this run",
-                ):
+                        ", ".join(unsettled),
-                    if st.item_id in plex_states or st.status == "hidden":
+                        extra=logs.event(
-                        continue
+                            "plex.watch.reconcile_incomplete",
-                    if st.synced_to_plex:
+                            user_id=link.user_id,
-                        db.delete(st)
+                            sections=unsettled,
-                        total["cleared"] += 1
+                        ),
+                    )
 
                 total["repushed"] += _repush_dirty(db, plex, link)
                 link.last_watch_sync_at = datetime.now(timezone.utc)
                 db.commit()
+        # silent-ok: the `{"skipped": …}` return is the report, as above.
         except PlexNotConfigured as e:
             return {"skipped": str(e)}
         except PlexError as e:

backend/app/progress.py

@@ -8,7 +8,7 @@ its own thread/context) never cross wires. When no sink is bound (e.g. a manual
 """
 
 import contextvars
-from typing import Callable, Optional
+from typing import Callable, Literal, Optional
 
 Sink = Callable[[int, Optional[int], Optional[str]], None]
 
@@ -24,6 +24,19 @@ def report(current: int, total: int | None = None, phase: str | None = None) ->
         sink(current, total, phase)
 
 
+class JobFailed(RuntimeError):
+    """A job's own failure, said so the reader's language can name it (E31): the client reads
+    `scheduler.failure.<reason>` and interpolates `params` (keep them to values the reader may see —
+    numbers, ids, short codes). The message is the English line the log and `last_error` keep. An
+    upstream's own error (a timeout, a refused connection) is not one of these: it stays as raised,
+    being diagnostic."""
+
+    def __init__(self, reason: str, message: str, **params: object):
+        super().__init__(message)
+        self.reason = reason
+        self.params = params
+
+
 class bind:
     """Context manager that binds a progress sink for the duration of a job run."""
 
@@ -35,7 +48,9 @@ class bind:
         self._token = _sink.set(self._sink)
         return self
 
-    def __exit__(self, *exc) -> bool:
+    # Literal[False], not bool: this never swallows the job's exception, and saying so is what lets
+    # a type checker know a name assigned inside `with bind(...)` is bound after it.
+    def __exit__(self, *exc) -> Literal[False]:
         if self._token is not None:
             _sink.reset(self._token)
         return False

backend/app/quota.py

@@ -7,11 +7,13 @@ always be allowed; expensive backfill must yield to the remaining budget.
 
 import contextlib
 import contextvars
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
+from typing import Any, cast
 from zoneinfo import ZoneInfo
 
-from sqlalchemy import func, select
+from sqlalchemy import delete, func, select
 from sqlalchemy.dialects.postgresql import insert as pg_insert
+from sqlalchemy.engine import CursorResult
 from sqlalchemy.orm import Session
 
 from app import sysconfig
@@ -37,6 +39,7 @@ class QuotaAction:
     PLAYLISTS_DELETE = "playlists_delete"
     VIDEOS_BACKFILL_RECENT = "videos_backfill_recent"
     VIDEOS_BACKFILL_FULL = "videos_backfill_full"
+    VIDEOS_FEED_FALLBACK = "videos_feed_fallback"
     VIDEOS_ENRICH = "videos_enrich"
     VIDEOS_LOOKUP = "videos_lookup"
     VIDEOS_SEARCH = "videos_search"
@@ -95,6 +98,20 @@ def pacific_today():
     return datetime.now(_PACIFIC).date()
 
 
+def gc(db: Session, retention_days: int) -> int:
+    """Delete quota-history rows older than `retention_days` (<= 0 keeps them forever) and return how
+    many went. Flushes only; the cleanup job commits. Today's rows are never in reach: the per-user
+    search cap counts them, and the floor of two days outlasts any Pacific day (25 h at most)."""
+    if retention_days <= 0:
+        return 0
+    cutoff = datetime.now(timezone.utc) - timedelta(days=max(2, retention_days))
+    # A DELETE answers with a CursorResult; `execute` is typed as the general Result, without rowcount.
+    result = cast(
+        CursorResult[Any], db.execute(delete(QuotaEvent).where(QuotaEvent.created_at < cutoff))
+    )
+    return result.rowcount or 0
+
+
 def pacific_day_start_utc() -> datetime:
     """UTC instant of the current Pacific day's midnight — aligns 'today' with the budget reset."""
     start = datetime.now(_PACIFIC).replace(hour=0, minute=0, second=0, microsecond=0)

backend/app/ratelimit.py

@@ -17,10 +17,14 @@ from app.config import settings
 
 
 class RateLimiter:
-    def __init__(self, max_events: int, window_seconds: float):
+    """At most `max_events` attempts per `window_seconds` per key — or, when each attempt passes a
+    `cost`, at most that much total cost (a byte budget counts bytes, not calls)."""
+
+    def __init__(self, max_events: float, window_seconds: float):
         self.max_events = max_events
         self.window = window_seconds
-        self._hits: dict[str, list[float]] = {}
+        # key → (monotonic time, cost) of every counted attempt still inside the window
+        self._hits: dict[str, list[tuple[float, float]]] = {}
         self._lock = threading.Lock()
 
     def reset(self) -> None:
@@ -36,27 +40,27 @@ class RateLimiter:
         count."""
         cutoff = time.monotonic() - self.window
         with self._lock:
-            return sum(1 for t in self._hits.get(key, []) if t > cutoff) >= self.max_events
+            return sum(c for t, c in self._hits.get(key, []) if t > cutoff) >= self.max_events
 
-    def allow(self, key: str) -> bool:
+    def allow(self, key: str, cost: float = 1) -> bool:
-        """Record an attempt for ``key`` and return whether it is within the limit.
+        """Record an attempt costing ``cost`` for ``key`` and return whether it is within the limit.
 
-        True  -> under the cap (the attempt is counted).
+        True  -> within the cap (the attempt is counted).
-        False -> the cap for the current window is already reached (attempt NOT counted, so a
+        False -> the attempt would pass the cap for the current window (attempt NOT counted, so a
                  blocked caller can't keep pushing the window forward)."""
         now = time.monotonic()
         cutoff = now - self.window
         with self._lock:
-            hits = [t for t in self._hits.get(key, []) if t > cutoff]
+            hits = [(t, c) for t, c in self._hits.get(key, []) if t > cutoff]
-            if len(hits) >= self.max_events:
+            if sum(c for _, c in hits) + cost > self.max_events:
                 self._hits[key] = hits
                 return False
-            hits.append(now)
+            hits.append((now, cost))
             self._hits[key] = hits
             # Opportunistic cleanup so abandoned keys don't accumulate unboundedly.
             if len(self._hits) > 4096:
                 for k in list(self._hits):
-                    fresh = [t for t in self._hits[k] if t > cutoff]
+                    fresh = [(t, c) for t, c in self._hits[k] if t > cutoff]
                     if fresh:
                         self._hits[k] = fresh
                     else:
@@ -69,6 +73,8 @@ def _norm_ip(s: str) -> ipaddress.IPv4Address | ipaddress.IPv6Address | None:
     plain-IPv4 allowlist still matches a peer that surfaces in mapped form. None for a non-IP string."""
     try:
         ip = ipaddress.ip_address(s.strip())
+    # silent-ok: parsing an untrusted header value, where "not an IP" is a documented answer this
+    # function returns — and a line per request is what an attacker would aim for.
     except ValueError:
         return None
     return ip.ipv4_mapped if getattr(ip, "ipv4_mapped", None) else ip

backend/app/realtime.py

@@ -13,12 +13,12 @@ mutated from a single thread.
 """
 
 import asyncio
-import logging
 from collections import defaultdict
 
 from starlette.websockets import WebSocket
+from app import logs
 
-log = logging.getLogger("siftlode.realtime")
+log = logs.get_logger(__name__)
 
 
 class ConnectionManager:
@@ -42,9 +42,16 @@ class ConnectionManager:
         for ws in list(self._conns.get(user_id, ())):
             try:
                 await ws.send_json(payload)
-            except Exception:
+            except Exception as exc:  # noqa: BLE001 — one dead socket must not stop the others
-                # Stale socket; it'll be cleaned up by its own disconnect handler.
+                # Stale socket; it'll be cleaned up by its own disconnect handler. But this is
-                pass
+                # also where a live update goes missing: the UI simply never changes, and there
+                # is no request behind it whose failure anyone could look up.
+                log.debug(
+                    "a live update could not be pushed to a socket",
+                    extra=logs.event(
+                        "realtime.push_failed", user_id=user_id, reason=type(exc).__name__
+                    ),
+                )
 
     def push(self, user_id: int, payload: dict) -> None:
         """Schedule a push onto the WebSocket event loop. Safe to call from a sync route
@@ -54,7 +61,7 @@ class ConnectionManager:
             return
         try:
             asyncio.run_coroutine_threadsafe(self._push(user_id, payload), loop)
-        except Exception:
+        except Exception:  # noqa: BLE001 — a closed loop or socket must not fail the caller
             log.debug("ws push failed", exc_info=True)
 
 

backend/app/refusal.py

@@ -0,0 +1,37 @@
+"""A request the app refuses, said in a way the UI can put in the reader's language.
+
+Every refusal the API answers carries a structured `detail` — `{code, reason, …params}` — and the client
+translates it (`lib/api/core.ts` `localizeDetail`). A bare English string reached the Hungarian UI
+verbatim: the blocking error dialog, and every inline surface that printed `HttpError.detail`, showed
+whatever sentence the route happened to hand-write. `refused` is the generic family
+(`errors.refused.<reason>`, the params interpolated into the sentence); the older specific codes
+(`not_found`, `youtube`, `playlist`, the admin refusals, …) stay, since each is already translated.
+
+A test (`test_http_details.py`) reads every `HTTPException` in the app and fails on a string detail,
+and the frontend's contract test requires every `refused(…, "<reason>")` to be named in both
+languages — so the rule is enforced, not described.
+"""
+
+from fastapi import HTTPException
+
+
+def refused(status: int, reason: str, **params: object) -> HTTPException:
+    """The refusal to raise: `raise refused(409, "view_name_taken", name=name)`. `params` travel in
+    the detail for the sentence to interpolate (`{{name}}`); keep them to values the reader may see."""
+    return HTTPException(status_code=status, detail={"code": "refused", "reason": reason, **params})
+
+
+# `sysconfig.InvalidSetting.reason` → the refusal it answers as. A dict of literals, so the frontend's
+# contract test can read every reason this can produce.
+_SETTING_REASONS = {
+    "not_a_number": "setting_not_a_number",
+    "too_low": "setting_too_low",
+    "too_high": "setting_too_high",
+    "not_a_choice": "setting_not_a_choice",
+    "no_encryption_key": "no_encryption_key",
+}
+
+
+def setting_refused(exc) -> HTTPException:
+    """A value the admin entered that `sysconfig.set_value` refused (`exc`: its `InvalidSetting`)."""
+    return refused(400, _SETTING_REASONS[exc.reason], key=exc.key, **exc.params)

backend/app/routes/_common.py

@@ -6,13 +6,16 @@ and "fetch a row this user owns or 404" idioms have one definition, one status c
 one existence-leak policy.
 """
 
-from typing import Any, TypeVar
+from typing import IO, Any, TypeVar
 
-from fastapi import HTTPException
+from fastapi import HTTPException, Request
 from sqlalchemy.orm import Session
 
 from app.db import Base
 from app.models import User
+from app.issues import gitea
+from app.refusal import refused
+from app.youtube.client import YouTubeError, refusal_reason
 
 M = TypeVar("M", bound=Base)
 
@@ -25,10 +28,13 @@ def not_found(what: str = "generic") -> HTTPException:
     return HTTPException(status_code=404, detail={"code": "not_found", "reason": what})
 
 
-def get_or_404(db: Session, model: type[M], id_: Any, *, what: str = "generic") -> M:
+def get_or_404(
+    db: Session, model: type[M], id_: Any, *, what: str = "generic", lock: bool = False
+) -> M:
     """Fetch a row by primary key or raise 404. Replaces the `db.get(Model, id); if None: 404`
-    idiom repeated across routes."""
+    idiom repeated across routes. `lock` takes it FOR UPDATE (bypassing the identity map), so two
-    obj = db.get(model, id_)
+    requests that read-then-write its children (a playlist's item positions) run one after the other."""
+    obj = db.get(model, id_, with_for_update=True if lock else None)
     if obj is None:
         raise not_found(what)
     return obj
@@ -42,12 +48,101 @@ def owned_or_404(
     *,
     owner_attr: str = "user_id",
     what: str = "generic",
+    lock: bool = False,
 ) -> M:
-    """Fetch a row by primary key that `user` owns, or raise 404. A row owned by someone
+    """Fetch a row by primary key that `user` owns, or raise 404 (`lock`: see `get_or_404`). A row owned by someone
     else 404s (NOT 403) on purpose — a 403 would confirm the id exists to a stranger.
-    Generalizes the near-identical per-route ownership helpers (e.g. plex's
+    It replaced the near-identical per-route ownership helpers, and is called directly — a thin
-    `_own_playlist_or_404`, the saved-views owner check)."""
+    `_own_x` wrapper around it is the ad-hoc shape CLAUDE.md's patterns table rules out."""
-    obj = get_or_404(db, model, id_, what=what)
+    obj = get_or_404(db, model, id_, what=what, lock=lock)
     if getattr(obj, owner_attr) != user.id:
         raise not_found(what)
     return obj
+
+
+def youtube_refusal(exc: Exception, *, playlist: bool = False) -> HTTPException:
+    """The answer to an action YouTube refused: a 422 whose reason the UI translates
+    (`errors.youtube.<reason>`). One definition, because the routes had grown three shapes of it —
+    a bare English string quoting the upstream message, a playlist-scoped reason, and nothing at
+    all (a raw 500 from the manual syncs). The upstream failure is already logged where it
+    happened (`youtube.api.error`). A failure that is not the Data API's (the search scraper's) is
+    YouTube being unavailable to us — the retryable answer. `playlist`: the route acts on one
+    playlist the user picked, so a missing playlist is THAT one, gone (`refusal_reason`)."""
+    reason = (
+        refusal_reason(exc, playlist=playlist) if isinstance(exc, YouTubeError) else "unavailable"
+    )
+    return HTTPException(status_code=422, detail={"code": "youtube", "reason": reason})
+
+
+async def capped_body(request: Request, limit: int) -> bytes:
+    """The request body, refused (413 `too_large`) the moment it passes `limit` bytes while still
+    arriving. Counted, not taken from Content-Length: a declared length can be absent (chunked) or
+    wrong, and counting stops reading at the same point a declared-length check would have."""
+    body = bytearray()
+    async for chunk in request.stream():
+        body += chunk
+        if len(body) > limit:
+            raise HTTPException(status_code=413, detail={"code": "too_large"})
+    return bytes(body)
+
+
+async def capped_stream_to(
+    request: Request, out: IO[bytes], limit: int, refusal: HTTPException
+) -> int:
+    """`capped_body`'s sibling for a body too large to hold: the same counting rule, written
+    straight to `out` instead of accumulated, returning the size.
+
+    Two deliberate differences from `capped_body`. It never builds a `bytearray` — the upload route
+    admits 500 MiB by default, so holding the body would trade a disk problem for a memory one. And
+    the refusal is the CALLER'S, passed in rather than assumed: `capped_body`'s flat 413 is right
+    for a screenshot, while the upload route answers 422 `upload_too_large` carrying the admin
+    knob's own number, which is what its dialog shows the user. One counting rule, two answers,
+    neither of them a second copy of the loop.
+
+    The caller owns `out` and owns cleaning it up: a body refused half-way has already written what
+    arrived before the limit was passed.
+
+    **The refusal DRAINS the rest of the body before it is raised, and that is not politeness.**
+    Answering mid-upload and closing the connection is what the client sees as a network failure,
+    not as a refusal: the browser's XHR fires `onerror` before it can read the response, so a file
+    over the limit reported itself as "the server couldn't be reached" instead of naming the limit —
+    measured in a real browser, not reasoned about. Reading the remainder to nowhere lets the client
+    finish sending, read the 422 and say the true thing.
+
+    The bytes are discarded, never written, and the whole request is already bounded by
+    `bodylimit`'s per-route ceiling — so the worst case is bandwidth an authenticated caller was
+    sending anyway, against the old behaviour which wrote exactly those bytes to disk."""
+    size = 0
+    over = False
+    async for chunk in request.stream():
+        size += len(chunk)
+        if over:
+            continue  # draining — see the docstring; counted, not stored
+        if size > limit:
+            over = True
+            continue
+        out.write(chunk)
+    if over:
+        raise refusal
+    return size
+
+
+def issue_refused(exc: gitea.GiteaError) -> HTTPException:
+    """The refusal an issue-tracker failure (E16) answers as — the send, the preview and the
+    admin's connection test alike. Literal reasons, one per case: the frontend's contract test reads
+    them to require a sentence for each in both languages."""
+    match exc.reason:
+        case "not_configured":
+            return refused(400, "issues_not_configured")
+        case "unreachable":
+            return refused(422, "issues_unreachable")
+        case "bad_token":
+            return refused(422, "issues_bad_token")
+        case "repo_not_found":
+            return refused(422, "issues_repo_not_found")
+        case "public_repo":
+            return refused(422, "issues_public_repo")
+        case "issues_disabled":
+            return refused(422, "issues_disabled")
+        case _:
+            return refused(422, "issues_rejected")

backend/app/routes/admin.py

@@ -39,6 +39,7 @@ from app.routes.auth import (
     send_verification_link,
 )
 from app.utils import canonical_email, iso, valid_email
+from app.refusal import refused
 
 router = APIRouter(prefix="/api/admin", tags=["admin"])
 
@@ -169,7 +170,7 @@ def _activate_user_by_email(db: Session, email: str) -> None:
     user = next((u for u in candidates if u.email == email), None) or next(iter(candidates), None)
     if user is not None and not user.is_active:
         user.is_active = True
-        db.commit()
+        db.flush()  # the handler commits (CLAUDE.md: helpers flush)
 
 
 def _decide(db: Session, invite_id: int, admin: User, approved: bool) -> Invite:
@@ -179,7 +180,7 @@ def _decide(db: Session, invite_id: int, admin: User, approved: bool) -> Invite:
     # sending a duplicate decision mail, even flipping an already-approved invite to "denied" and
     # mailing a rejection to an already-active account.
     if inv.status != "pending":
-        raise HTTPException(status_code=409, detail="Invite already decided")
+        raise refused(409, "invite_decided")
     inv.status = "approved" if approved else "denied"
     inv.decided_at = datetime.now(timezone.utc)
     inv.decided_by = admin.email
@@ -191,7 +192,7 @@ def _decide(db: Session, invite_id: int, admin: User, approved: bool) -> Invite:
         target_id=inv.id,
         summary=inv.email,
     )
-    db.commit()
+    db.flush()  # the handler commits, with the activation that goes with it
     return inv
 
 
@@ -210,6 +211,9 @@ def approve_invite(
     link, has_account, hours = _welcome_mail_parts(db, inv.email)
     lang = mailtext.lang_for_email(db, inv.email)
     background.add_task(email_mod.send_access_approved, inv.email, lang, link, has_account, hours)
+    # Before the response, so the mail (a background task, after it) never describes a decision
+    # the database does not hold yet.
+    db.commit()
     return _serialize(inv, has_account=bool(_accounts_for(db, [inv.email])))
 
 
@@ -226,6 +230,7 @@ def deny_invite(
     background.add_task(
         email_mod.send_access_denied, inv.email, mailtext.lang_for_email(db, inv.email)
     )
+    db.commit()
     return _serialize(inv, has_account=bool(_accounts_for(db, [inv.email])))
 
 
@@ -242,7 +247,7 @@ def add_invite(
     """Manually whitelist an email (approved straight away). Convenience for the admin."""
     email = body.email.strip().lower()
     if not valid_email(email):
-        raise HTTPException(status_code=400, detail="Enter a valid email address.")
+        raise refused(400, "invalid_email")
     # Reuse a canonically-equal row instead of adding a second one for the same mailbox. Prod carries
     # exactly that pair for one person from 2026-06-16 (hartmann.peter1986@ / hartmannpeter1986@),
     # which is how an admin ends up approving the same human twice and still seeing them bounce.
@@ -259,11 +264,13 @@ def add_invite(
     inv.status = "approved"
     inv.decided_at = datetime.now(timezone.utc)
     inv.decided_by = admin.email
+    db.flush()  # a new invite's id, for the audit row — before it, the row named no target
     audit.record(
         db, admin.id, AuditAction.INVITE_ADD, target_type="invite", target_id=inv.id, summary=email
     )
-    db.commit()
+    # One transaction: the whitelisting and the activation it implies land together or not at all.
     _activate_user_by_email(db, email)
+    db.commit()
     return _serialize(inv, has_account=bool(_accounts_for(db, [email])))
 
 
@@ -366,10 +373,7 @@ def delete_invite(
     promoted to admin on its next Google sign-in). The admin has to hear that, not guess it."""
     inv = get_or_404(db, Invite, invite_id, what="invite")
     if _accounts_for(db, [inv.email]):
-        raise HTTPException(
+        raise refused(409, "address_has_account")
-            status_code=409,
-            detail="That address already has an account — delete the account instead.",
-        )
     still_allowed_by_env = _env_allowed(inv.email)
     audit.record(
         db,
@@ -408,6 +412,8 @@ def _serialize_user(u: User) -> dict:
         # confusing sign-in.
         "google_email": u.google_email if u.google_email != u.email else None,
         "created_at": iso(u.created_at),
+        # NULL = no sign-in since the column existed (0085), or the shared demo account.
+        "last_login_at": iso(u.last_login_at),
     }
 
 
@@ -433,19 +439,19 @@ def set_user_role(
     account, and can't remove the last remaining admin. Emails the user when it actually changes."""
     role = body.role
     if role not in ("user", "admin"):
-        raise HTTPException(status_code=400, detail="role must be 'user' or 'admin'")
+        raise refused(400, "invalid_role")
     target = get_or_404(db, User, user_id, what="user")
     if target.is_demo:
-        raise HTTPException(status_code=400, detail="The demo account's role can't be changed.")
+        raise refused(400, "demo_role")
     if target.id == admin.id:
-        raise HTTPException(status_code=400, detail="You can't change your own role.")
+        raise refused(400, "own_role")
     if (
         target.role == "admin"
         and role == "user"
         and not target.is_suspended
         and count_admins(db, active_only=True) <= 1
     ):
-        raise HTTPException(status_code=400, detail="Can't remove the last admin.")
+        raise refused(400, "last_admin")
     changed = target.role != role
     if changed:
         audit.record(
@@ -491,16 +497,16 @@ def set_user_suspended(
     suspended = body.suspended
     target = get_or_404(db, User, user_id, what="user")
     if target.is_demo:
-        raise HTTPException(status_code=400, detail="The demo account can't be suspended.")
+        raise refused(400, "demo_suspend")
     if target.id == admin.id:
-        raise HTTPException(status_code=400, detail="You can't suspend your own account.")
+        raise refused(400, "own_suspend")
     if (
         suspended
         and target.role == "admin"
         and not target.is_suspended
         and count_admins(db, active_only=True) <= 1
     ):
-        raise HTTPException(status_code=400, detail="Can't suspend the last active admin.")
+        raise refused(400, "last_active_admin")
     was_suspended = target.is_suspended
     if was_suspended != suspended:
         audit.record(
@@ -535,17 +541,15 @@ def admin_delete_user(
     not the demo account, not yourself (use Settings → Account), and not the last admin."""
     target = get_or_404(db, User, user_id, what="user")
     if target.is_demo:
-        raise HTTPException(status_code=400, detail="The demo account can't be deleted.")
+        raise refused(400, "demo_delete")
     if target.id == admin.id:
-        raise HTTPException(
+        raise refused(400, "own_delete")
-            status_code=400, detail="Delete your own account from Settings → Account."
-        )
     if (
         target.role == "admin"
         and not target.is_suspended
         and count_admins(db, active_only=True) <= 1
     ):
-        raise HTTPException(status_code=400, detail="Can't delete the last admin.")
+        raise refused(400, "last_admin")
     # Record BEFORE the purge — the target row (and its email) is gone afterwards.
     audit.record(
         db,
@@ -689,7 +693,7 @@ def activate_user(
     `_activate_user_by_email`). What was missing was any way to know they had to."""
     target = get_or_404(db, User, user_id, what="user")
     if target.is_demo:
-        raise HTTPException(status_code=400, detail="The demo account can't be activated.")
+        raise refused(400, "demo_activate")
     if target.is_active:
         return _serialize_user(target)  # idempotent: a double click is not an error
     if not is_allowed(db, target.email):
@@ -786,7 +790,7 @@ def add_demo_whitelist(
     sign-in). Idempotent — re-adding an existing email just returns it."""
     email = body.email.strip().lower()
     if not valid_email(email):
-        raise HTTPException(status_code=400, detail="Enter a valid email address.")
+        raise refused(400, "invalid_email")
     row = db.execute(select(DemoWhitelist).where(DemoWhitelist.email == email)).scalar_one_or_none()
     if row is None:
         row = DemoWhitelist(
@@ -830,10 +834,14 @@ def reset_demo(admin: User = Depends(admin_user), db: Session = Depends(get_db))
     demo = get_or_create_demo_user(db)
     seeded = reset_demo_state(db, demo)
     audit.record(
-        db, admin.id, AuditAction.DEMO_RESET, summary=f"{seeded} sample playlists re-seeded"
+        db,
+        admin.id,
+        AuditAction.DEMO_RESET,
+        summary=f"{seeded} sample playlists re-seeded",
+        after={"playlists_seeded": seeded},  # the count as data: named in the reader's language
     )
     db.commit()
-    return {"reset": True, "playlists_seeded": seeded}
+    return {"ok": True, "reset": True, "playlists_seeded": seeded}
 
 
 @router.post("/purge-discovery")
@@ -852,4 +860,4 @@ def purge_discovery(admin: User = Depends(admin_user), db: Session = Depends(get
         after=result,
     )
     db.commit()
-    return result
+    return {"ok": True, **result}

backend/app/routes/admin_logs.py

@@ -0,0 +1,216 @@
+"""The admin diagnostics console's read side (E22.6): query `log_events`, the server's stored log.
+
+Admin-only and read-only — rows leave the table through the retention job (`log_retention_days`),
+never through here. Reading is not audited: it changes nothing, and the console's live tail polls,
+so a row per read would bury the audit log's real entries.
+
+What a row holds was redacted when it was written (`logs.record_fields`, and the ingest route for a
+browser's rows), so this serves the columns as they are; a value the redactor took out reads as
+`[redacted]` / `[email]`, which is the point — the console shows that something was there.
+
+Paging is a keyset over `(ts desc, id desc)`: a browser row's `ts` is clamped into its consent
+window, so `ts` alone is not unique, and an offset would skip or repeat rows while the writer adds
+new ones at the top. The live tail re-reads the first page rather than asking for "ids after N":
+several processes write this table, and a batch's ids are taken when it inserts, not when it
+commits, so a slower batch can land below an id the tail already passed — and would never show.
+"""
+
+import base64
+import binascii
+import json
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Literal
+
+from fastapi import APIRouter, Depends, HTTPException, Query
+from sqlalchemy import Text, and_, cast, func, or_, select
+from sqlalchemy.orm import Session
+
+from app import logsink
+from app.auth import admin_user
+from app.db import get_db
+from app.models import LogEvent, User
+from app.utils import like_escape
+
+router = APIRouter(prefix="/api/admin/logs", tags=["admin"])
+
+PAGE_MAX = 200
+# A facet list is a chip row, not a report: past this many distinct events the rest are reachable
+# through the free-text search, and the response stays small whatever a self-host logs.
+FACET_EVENTS_MAX = 100
+
+Level = Literal["TRACE", "DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"]
+# The filters a facet can leave out of its own count (`LogFilters.conditions`).
+Facet = Literal["levels", "sources", "user", "event"]
+# `none` selects the rows that have no account / no event code (a scheduler line, an anonymous
+# request) — a real value to filter on, not the absence of a filter.
+NONE = "none"
+
+
+def _utc(value: datetime | None) -> datetime | None:
+    """A bound without a zone is UTC — the column is `timestamptz`, and a naive value would be read
+    in the database session's zone instead."""
+    if value is None or value.tzinfo is not None:
+        return value
+    return value.replace(tzinfo=timezone.utc)
+
+
+def _encode_cursor(row: LogEvent) -> str:
+    return base64.urlsafe_b64encode(json.dumps([row.ts.isoformat(), row.id]).encode()).decode()
+
+
+def _decode_cursor(cursor: str) -> tuple[datetime, int]:
+    try:
+        ts, row_id = json.loads(base64.urlsafe_b64decode(cursor.encode()))
+        return _utc(datetime.fromisoformat(ts)), int(row_id)
+    except (binascii.Error, json.JSONDecodeError, ValueError, TypeError) as exc:
+        raise HTTPException(status_code=400, detail={"code": "bad_cursor"}) from exc
+
+
+def _accounts(db: Session, user_ids: set[int]) -> dict[str, str]:
+    """The accounts a page names, by id → address. An id with no account left (erased after the
+    line was written) is simply absent, and the client shows the bare id."""
+    if not user_ids:
+        return {}
+    rows = db.execute(select(User.id, User.email).where(User.id.in_(user_ids))).all()
+    return {str(uid): email for uid, email in rows}
+
+
+@dataclass(frozen=True)
+class LogFilters:
+    """The console's filters, shared by the page and its facet counts."""
+
+    levels: list[str]
+    sources: list[str]
+    user: str | None
+    event: str | None
+    request_id: str | None
+    job_id: int | None
+    q: str | None
+    since: datetime | None
+    until: datetime | None
+
+    def conditions(self, skip: Facet | None = None) -> list:
+        """The WHERE conditions, ANDed by the caller. `skip` leaves one filter out: a facet counts
+        its own values under every OTHER filter, so picking a level does not zero the other
+        levels' counts (the way back), while a request pivot does narrow them (honest counts)."""
+        out = []
+        if self.levels and skip != "levels":
+            out.append(LogEvent.level.in_(self.levels))
+        if self.sources and skip != "sources":
+            out.append(LogEvent.source.in_(self.sources))
+        if self.user is not None and skip != "user":
+            out.append(
+                LogEvent.user_id.is_(None)
+                if self.user == NONE
+                else LogEvent.user_id == int(self.user)
+            )
+        if self.event is not None and skip != "event":
+            out.append(
+                LogEvent.event.is_(None) if self.event == NONE else LogEvent.event == self.event
+            )
+        if self.request_id is not None:
+            out.append(LogEvent.request_id == self.request_id)
+        if self.job_id is not None:
+            out.append(LogEvent.job_id == self.job_id)
+        if self.since is not None:
+            out.append(LogEvent.ts >= _utc(self.since))
+        if self.until is not None:
+            out.append(LogEvent.ts < _utc(self.until))
+        if self.q and self.q.strip():
+            like = f"%{like_escape(self.q.strip())}%"
+            out.append(
+                or_(
+                    LogEvent.msg.ilike(like),
+                    LogEvent.event.ilike(like),
+                    LogEvent.request_id.ilike(like),
+                    LogEvent.exc.ilike(like),
+                    cast(LogEvent.attrs, Text).ilike(like),
+                )
+            )
+        return out
+
+
+def log_filters(
+    levels: list[Level] = Query(default=[]),
+    sources: list[str] = Query(default=[], max_length=8),
+    user: str | None = Query(default=None, pattern=r"^(none|[0-9]{1,10})$"),
+    event: str | None = Query(
+        default=None, max_length=64, pattern=r"^(none|[a-z][a-z0-9_]*(\.[a-z0-9_]+)*)$"
+    ),
+    request_id: str | None = Query(default=None, pattern=r"^[A-Za-z0-9._-]{8,64}$"),
+    job_id: int | None = Query(default=None, ge=1),
+    q: str | None = Query(default=None, max_length=200),
+    since: datetime | None = None,
+    until: datetime | None = None,
+) -> LogFilters:
+    return LogFilters(list(levels), list(sources), user, event, request_id, job_id, q, since, until)
+
+
+@router.get("")
+def list_logs(
+    filters: LogFilters = Depends(log_filters),
+    before: str | None = Query(default=None, max_length=200),
+    limit: int = Query(default=100, ge=1, le=PAGE_MAX),
+    _: User = Depends(admin_user),
+    db: Session = Depends(get_db),
+) -> dict:
+    """One page of rows, newest first, matching every filter given (they AND together); `before`
+    continues from a previous page's `next_cursor`."""
+    query = select(LogEvent).where(*filters.conditions())
+    if before is not None:
+        ts, row_id = _decode_cursor(before)
+        # The `ts <=` conjunct is implied by the OR; it is there so the planner can bound the
+        # `ts` index scan instead of filtering every newer row first.
+        query = query.where(
+            and_(
+                LogEvent.ts <= ts,
+                or_(LogEvent.ts < ts, and_(LogEvent.ts == ts, LogEvent.id < row_id)),
+            )
+        )
+
+    # One more than asked, to know whether there is a next page without a second count query.
+    rows = (
+        db.execute(query.order_by(LogEvent.ts.desc(), LogEvent.id.desc()).limit(limit + 1))
+        .scalars()
+        .all()
+    )
+    more = len(rows) > limit
+    rows = rows[:limit]
+    return {
+        "items": [logsink.serialize(r) for r in rows],
+        "accounts": _accounts(db, {r.user_id for r in rows if r.user_id is not None}),
+        "next_cursor": _encode_cursor(rows[-1]) if more else None,
+    }
+
+
+@router.get("/facets")
+def log_facets(
+    filters: LogFilters = Depends(log_filters),
+    _: User = Depends(admin_user),
+    db: Session = Depends(get_db),
+) -> dict:
+    """What the filter chips offer, with counts: the levels, sources and accounts present, and the
+    most frequent events — each counted under every filter but its own (`LogFilters.conditions`).
+    `total` is the "all levels" count: every filter but the level one."""
+
+    def counts(col, skip: Facet, cap: int | None = None) -> list[dict]:
+        n = func.count().label("n")
+        stmt = select(col, n).where(*filters.conditions(skip)).group_by(col).order_by(n.desc(), col)
+        if cap is not None:
+            stmt = stmt.limit(cap)
+        return [
+            {"value": NONE if value is None else value, "count": count}
+            for value, count in db.execute(stmt).all()
+        ]
+
+    users = counts(LogEvent.user_id, "user")
+    total = select(func.count()).select_from(LogEvent).where(*filters.conditions("levels"))
+    return {
+        "levels": counts(LogEvent.level, "levels"),
+        "sources": counts(LogEvent.source, "sources"),
+        "events": counts(LogEvent.event, "event", FACET_EVENTS_MAX),
+        "users": [{**u, "value": str(u["value"])} for u in users],
+        "accounts": _accounts(db, {u["value"] for u in users if u["value"] != NONE}),
+        "total": db.scalar(total) or 0,
+    }

backend/app/routes/audit.py

@@ -66,6 +66,12 @@ def list_audit(
 def clear_audit(admin: User = Depends(admin_user), db: Session = Depends(get_db)) -> dict:
     """Clear the audit log. Leaves ONE row recording who cleared it + how many (tamper-evidence)."""
     n = db.execute(delete(AuditLog)).rowcount or 0
-    audit.record(db, admin.id, AuditAction.AUDIT_CLEAR, summary=f"Cleared {n} audit entries")
+    audit.record(
+        db,
+        admin.id,
+        AuditAction.AUDIT_CLEAR,
+        summary=f"Cleared {n} audit entries",
+        after={"cleared": n},  # the count as data: the log names it in the reader's language
+    )
     db.commit()
-    return {"cleared": n}
+    return {"ok": True, "cleared": n}

backend/app/routes/auth.py

@@ -17,6 +17,7 @@ from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 
 from app import email as email_mod
+from app import logs
 from app import mailtext
 from app import sysconfig
 from app.auth import (
@@ -38,10 +39,11 @@ from app.auth import (
 )
 from app.config import settings
 from app.db import SessionLocal, get_db
-from app.models import AuthToken, DemoWhitelist, Invite, OAuthToken, User
+from app.models import AuthToken, DemoWhitelist, Invite, OAuthToken, User, display_text
 from app.ratelimit import RateLimiter, client_key
 from app.security import encrypt, hash_password, hash_token, verify_password
 from app.utils import canonical_email, valid_email
+from app.refusal import refused
 
 VERIFY_TTL = timedelta(hours=24)
 # A confirmation link that rides an APPROVAL or an invitation lives longer than one the person is
@@ -377,7 +379,7 @@ async def callback(
 
     userinfo = token.get("userinfo")
     if not userinfo or not userinfo.get("sub"):
-        raise HTTPException(status_code=400, detail="No user info returned by Google")
+        raise refused(400, "google_no_userinfo")
 
     sub = userinfo["sub"]
     email = (userinfo.get("email") or "").lower()
@@ -386,7 +388,15 @@ async def callback(
         return _complete_link(request, db, link_uid, link_explicit, sub, userinfo, token)
 
     if not is_allowed(db, email):
-        log.warning("Login denied (not approved): %s", email or "<no email>")
+        # No uid to name — having no account here is the branch. And the address is left out on
+        # purpose: `redact_text` turns it into `[email]` before either sink sees it, so a line
+        # whose only variable is always `[email]` says nothing. WHO asked is in the pending
+        # invite recorded just below, which is where an admin looks anyway.
+        log.warning(
+            "Google sign-in denied: address not approved",
+            # log-pii-ok: whether an address came back, not what it was
+            extra=logs.event("auth.login.denied", reason="not_approved", has_email=bool(email)),
+        )
         # A denied Google login doubles as an access request: record it for the admin and
         # land the user on a friendly "request received" screen instead of a raw 403.
         if email:
@@ -419,15 +429,20 @@ async def callback(
             if existing.google_sub is not None and existing.google_sub != sub:
                 # A different Google identity is already bound to this email — refuse rather than
                 # hijack it. (Near-impossible with real Google accounts: email↔sub is stable.)
-                log.warning("Google login email collision (different sub): %s", email)
+                log.warning(
+                    "Google sign-in refused: a different Google identity holds this address",
+                    extra=logs.event("auth.login.denied", reason="sub_collision", uid=existing.id),
+                )
                 return RedirectResponse(url=_access_notice_url("denied"))
             if not userinfo.get("email_verified", False):
                 # Adopting (and activating) a pre-existing password account requires Google to
                 # actually attest the address — otherwise an unverified Google email that merely
                 # collides with a pending registration could seize that account. Defense-in-depth.
                 log.warning(
-                    "Google login: unverified email, refusing to adopt account: %s",
+                    "Google sign-in refused: unverified address, not adopting the account",
-                    email,
+                    extra=logs.event(
+                        "auth.login.denied", reason="email_unverified", uid=existing.id
+                    ),
                 )
                 return RedirectResponse(url=_access_notice_url("denied"))
             user = existing
@@ -445,7 +460,10 @@ async def callback(
     # A suspended account can't sign in by any method. Block before establishing the session and
     # tell the (verified) owner why — a successful Google auth proves they control this account.
     if user.is_suspended:
-        log.info("Suspended account blocked at Google login: %s", email)
+        log.info(
+            "Google sign-in blocked: the account is suspended",
+            extra=logs.event("auth.login.denied", reason="suspended", uid=user.id),
+        )
         _notify_suspended(background, user)
         return RedirectResponse(url="/?login=suspended")
     # Record what Google actually answers with, beside the address the user registered with. The two
@@ -454,7 +472,7 @@ async def callback(
     if email:
         user.google_email = email
     _sync_google_email(db, user, email)
-    user.display_name = userinfo.get("name")
+    user.display_name = display_text(userinfo.get("name"))
     user.avatar_url = userinfo.get("picture")
     # ADMIN_EMAILS (env) is the bootstrap admin list and always wins, so the configured admin can't
     # be locked out. Everyone else's role is managed in the admin UI and stored in the DB, so a
@@ -479,17 +497,12 @@ async def callback(
     db.flush()
 
     _store_token(db, user, token)
+    _establish_session(request, user)
     db.commit()
-
+    log.info(
-    # Multi-session: remember every account that has authenticated in this browser so the
+        "signed in",
-    # user can switch between them without going through Google again. Only accounts that
+        extra=logs.event("auth.login.ok", method="google", uid=user.id, role=user.role),
-    # actually completed OAuth here land in this list (it's the proof they may be switched to).
+    )
-    accounts = [a for a in (request.session.get("account_ids") or []) if a != user.id]
-    accounts.append(user.id)
-    request.session["account_ids"] = accounts[-MAX_SESSION_ACCOUNTS:]
-    request.session["user_id"] = user.id
-    _remember_epoch(request, user)
-    log.info("Login: %s (id=%s, role=%s)", email, user.id, user.role)
     return RedirectResponse(url="/")
 
 
@@ -581,7 +594,7 @@ def _complete_link(
     if linked_email:
         user.google_email = linked_email  # see the sign-in callback: the other half of the identity
     if not user.display_name:
-        user.display_name = userinfo.get("name")
+        user.display_name = display_text(userinfo.get("name"))
     if not user.avatar_url:
         user.avatar_url = userinfo.get("picture")
     _store_token(db, user, token)
@@ -641,15 +654,15 @@ async def request_access(
     endpoint can't be probed for throttling either."""
     email = (body.email or "").strip().lower()
     if not valid_email(email):
-        raise HTTPException(status_code=400, detail="Enter a valid email address.")
+        raise refused(400, "invalid_email")
     if not _request_access_limiter.allow(client_key(request)):
-        return {"status": "pending"}
+        return {"ok": True, "status": "pending"}
     if is_allowed(db, email):
-        return {"status": "pending"}  # already allowed — don't reveal it
+        return {"ok": True, "status": "pending"}  # already allowed — don't reveal it
     req = upsert_pending_invite(db, email, body.lang)
     if req.is_new:
         notify_admins_of_request(db, background, email)
-    return {"status": "pending"}
+    return {"ok": True, "status": "pending"}
 
 
 @router.post("/demo")
@@ -660,7 +673,7 @@ def demo_login(body: EmailIn, request: Request, db: Session = Depends(get_db)) -
     exactly like a non-match (authenticated=false) so the field can't be hammered as an
     enumeration oracle. On a match it sets the session and the client reloads into the app."""
     if not _demo_limiter.allow(client_key(request)):
-        return {"authenticated": False}
+        return {"ok": True, "authenticated": False}
     email = (body.email or "").strip().lower()
     if valid_email(email) and is_demo_allowed(db, email):
         demo = get_or_create_demo_user(db)
@@ -669,19 +682,25 @@ def demo_login(body: EmailIn, request: Request, db: Session = Depends(get_db)) -
         request.session.clear()
         request.session["user_id"] = demo.id
         _remember_epoch(request, demo)
-        log.info("Demo login (key=%s, demo_id=%s)", email, demo.id)
+        # The demo key IS an address, so it is not in the line — see the denied branch above.
-        return {"authenticated": True}
+        log.info("signed in", extra=logs.event("auth.login.ok", method="demo", uid=demo.id))
-    return {"authenticated": False}
+        return {"ok": True, "authenticated": True}
+    return {"ok": True, "authenticated": False}
 
 
 def _establish_session(request: Request, user: User) -> None:
-    """Sign `user` in for this browser session, mirroring the Google callback's multi-account
+    """Sign `user` in for this browser session — the ONE sign-in path, shared by the Google callback
-    handling so password sign-in joins the same account switcher."""
+    and password login, so both join the same account switcher and both stamp `last_login_at` (the
+    caller commits). Multi-session: every account that authenticated in this browser is remembered so
+    the user can switch between them; only accounts that actually proved themselves HERE land in that
+    wallet (it is the proof they may be switched to). The demo entry deliberately does not come
+    through here: it clears the wallet instead, and its shared account has no meaningful last login."""
     accounts = [a for a in (request.session.get("account_ids") or []) if a != user.id]
     accounts.append(user.id)
     request.session["account_ids"] = accounts[-MAX_SESSION_ACCOUNTS:]
     request.session["user_id"] = user.id
     _remember_epoch(request, user)
+    user.last_login_at = datetime.now(timezone.utc)
 
 
 def _app_base() -> str:
@@ -797,23 +816,20 @@ def register(
     password = body.password or ""
     # Input validation is safe to reveal (independent of whether the email exists).
     if not valid_email(email):
-        raise HTTPException(status_code=400, detail="Enter a valid email address.")
+        raise refused(400, "invalid_email")
     if len(password) < PASSWORD_MIN_LEN:
-        raise HTTPException(
+        raise refused(400, "weak_password", min=PASSWORD_MIN_LEN)
-            status_code=400,
-            detail=f"Password must be at least {PASSWORD_MIN_LEN} characters.",
-        )
     if not sysconfig.get_bool(db, "allow_registration"):
-        raise HTTPException(status_code=403, detail="Registration is currently closed.")
+        raise refused(403, "registration_closed")
     if not _register_limiter.allow(client_key(request)):
-        return {"status": "ok"}  # silently throttle; uniform response
+        return {"ok": True, "status": "ok"}  # silently throttle; uniform response
 
     # Do the existence check + account creation OFF the response path (a background task with its own
     # session). The create path hashes the password + writes several rows + schedules emails; doing it
     # inline would make a NEW email respond measurably slower than an already-registered one (which
     # skips all that) — a timing oracle for enumeration. Off-path, any valid email responds the same.
     background.add_task(_register_account, email, password, body.lang)
-    return {"status": "ok"}
+    return {"ok": True, "status": "ok"}
 
 
 def _register_account(email: str, password: str, lang: str | None = None) -> None:
@@ -883,7 +899,12 @@ def _register_account(email: str, password: str, lang: str | None = None) -> Non
         except IntegrityError:
             # A concurrent registration for the same new email won the race between the pre-check
             # above and this flush (the unique-email constraint fires). Same already-registered
-            # no-op as the pre-check hit — roll back and stop.
+            # no-op as the pre-check hit — roll back and stop. The line carries nothing: the
+            # answer to the caller is deliberately identical either way, and so is this.
+            log.info(
+                "a registration lost the race for its address",
+                extra=logs.event("auth.register.race_lost"),
+            )
             db.rollback()
             return
         upsert_pending_invite(db, email)  # admin-approval gate
@@ -974,11 +995,11 @@ def verify_email_resend(
     mail had no way back. Uniform response + timing regardless of whether the email has a pending
     account (the lookup runs off the response path), so it can't probe for registered emails."""
     if not _verify_resend_limiter.allow(client_key(request)):
-        return {"status": "ok"}  # silently throttle; uniform response
+        return {"ok": True, "status": "ok"}  # silently throttle; uniform response
     email = (body.email or "").strip().lower()
     if valid_email(email):
         background.add_task(_resend_verify_if_eligible, email)
-    return {"status": "ok"}
+    return {"ok": True, "status": "ok"}
 
 
 def _resend_verify_if_eligible(email: str) -> None:
@@ -1051,7 +1072,11 @@ def password_login(
     if not user.is_active:
         raise _login_refused(403, "pending_approval", user)
     _establish_session(request, user)
-    log.info("Password login: %s (id=%s, role=%s)", email, user.id, user.role)
+    db.commit()
+    log.info(
+        "signed in",
+        extra=logs.event("auth.login.ok", method="password", uid=user.id, role=user.role),
+    )
     return {"ok": True}
 
 
@@ -1064,14 +1089,14 @@ def password_reset_request(
     """Request a password-reset link. Uniform response + timing regardless of whether the email has a
     password account (the lookup runs off the response path), so it can't probe for registered emails."""
     if not _reset_limiter.allow(client_key(request)):
-        return {"status": "ok"}
+        return {"ok": True, "status": "ok"}
     email = (body.email or "").strip().lower()
     # Do the account lookup + token issue + email OFF the response path (a background task with its
     # own session), so the endpoint takes the same time for any valid email whether or not it has a
     # password account — no timing-based enumeration. Any valid email schedules the same task.
     if valid_email(email):
         background.add_task(_send_reset_if_eligible, email)
-    return {"status": "ok"}
+    return {"ok": True, "status": "ok"}
 
 
 def _send_reset_if_eligible(email: str) -> None:
@@ -1266,13 +1291,10 @@ def set_password(
         raise HTTPException(status_code=403, detail={"code": "demo"})
     new_password = body.password or ""
     if len(new_password) < PASSWORD_MIN_LEN:
-        raise HTTPException(
+        raise refused(400, "weak_password", min=PASSWORD_MIN_LEN)
-            status_code=400,
-            detail=f"Password must be at least {PASSWORD_MIN_LEN} characters.",
-        )
     if user.password_hash:
         if not verify_password(body.current_password or "", user.password_hash):
-            raise HTTPException(status_code=403, detail="Current password is incorrect.")
+            raise refused(403, "wrong_password")
     user.password_hash = hash_password(new_password)
     user.password_revoked_at = None  # same as the reset path: a password exists again
     # SA4: a password change logs out every OTHER session; re-stamp this cookie at the new epoch so

backend/app/routes/channels.py

@@ -2,15 +2,16 @@
 hidden, tags) and per-channel sync state. Local reads/writes plus YouTube-side writes
 (subscribe / unsubscribe) that require the optional write scope."""
 
-import logging
 from datetime import datetime, timezone
 
-from fastapi import APIRouter, BackgroundTasks, Depends, HTTPException
+from fastapi import APIRouter, BackgroundTasks, Depends
 from pydantic import BaseModel
-from sqlalchemy import func, select, update
+from sqlalchemy import Select, func, select, update
 from sqlalchemy.orm import Session
 
-from app import quota
+from app.collation import collated, sorted_words
+from app.mailtext import lang_for_user
+from app import logs, quota
 from app.auth import admin_user, current_user, require_human, require_write_scope
 from app.db import get_db
 from app.models import (
@@ -26,16 +27,33 @@ from app.models import (
     User,
     Video,
 )
-from app.routes._common import get_or_404, not_found, owned_or_404
+from app.routes._common import get_or_404, not_found, owned_or_404, youtube_refusal
 from app.sync.explore import explore_ingest_page
 from app.sync.runner import backfill_channel_now, run_recent_backfill
-from app.sync.subscriptions import apply_channel_details
+from app.sync.subscriptions import apply_channel_details, details_stale
 from app.utils import iso
 from app.youtube.client import YouTubeClient, YouTubeError
+from app.refusal import refused
 
 router = APIRouter(prefix="/api/channels", tags=["channels"])
 
-log = logging.getLogger("siftlode.api")
+log = logs.get_logger(__name__)
+
+
+def channel_aggregates(channel_ids: list[str]) -> Select:
+    """Per-channel stored count, last upload and total duration, in one grouped pass. Every column it
+    reads is in `ix_videos_channel_pub_cover` (migration 0095), so Postgres answers it from the index
+    alone; a test holds that (test_channel_list_index_only)."""
+    return (
+        select(
+            Video.channel_id,
+            func.count(Video.id),
+            func.max(Video.published_at),
+            func.coalesce(func.sum(Video.duration_seconds), 0),
+        )
+        .where(Video.channel_id.in_(channel_ids))
+        .group_by(Video.channel_id)
+    )
 
 
 @router.get("")
@@ -44,27 +62,18 @@ def list_channels(user: User = Depends(current_user), db: Session = Depends(get_
         select(Subscription, Channel)
         .join(Channel, Channel.id == Subscription.channel_id)
         .where(Subscription.user_id == user.id)
-        .order_by(Subscription.priority.desc(), func.lower(Channel.title))
+        .order_by(Subscription.priority.desc(), collated(Channel.title, lang_for_user(user)))
     ).all()
 
-    # Per-channel aggregates over the shared catalog in ONE grouped pass: stored count, last
+    # Per-channel aggregates over the shared catalog (channel_aggregates), computed on read rather than
-    # upload, total duration. Cheap — a single grouped scan, so computed on read rather than
+    # denormalised: an Index Only Scan over a ~20 MB index, where it used to read every video's heap page
-    # denormalised (measure-first: reads aren't frequent/hot enough to warrant cached columns).
+    # (681 MB on prod, a 2–6 s first open once something else had pushed it out of memory, E31 S4d).
     # (The Normal/Shorts/Live breakdown moved to the channel-detail page in R14, so it is no longer
     # computed here — the manager table dropped its N/S/L column.)
     channel_ids = [c.id for _, c in rows]
     agg: dict[str, dict] = {}
     if channel_ids:
-        for cid, n, last_up, total_dur in db.execute(
+        for cid, n, last_up, total_dur in db.execute(channel_aggregates(channel_ids)).all():
-            select(
-                Video.channel_id,
-                func.count(Video.id),
-                func.max(Video.published_at),
-                func.coalesce(func.sum(Video.duration_seconds), 0),
-            )
-            .where(Video.channel_id.in_(channel_ids))
-            .group_by(Video.channel_id)
-        ).all():
             agg[cid] = {
                 "stored": n,
                 "last_video_at": iso(last_up),
@@ -139,7 +148,7 @@ def _discovery_rows(db: Session, user: User) -> list:
     return db.execute(
         q.group_by(Channel.id).order_by(
             func.count(func.distinct(PlaylistItem.video_id)).desc(),
-            func.lower(Channel.title),
+            collated(Channel.title, lang_for_user(user)),
         )
     ).all()
 
@@ -166,7 +175,15 @@ def discover_channels(
                 user.yt_channel_id = own
                 db.commit()
         except YouTubeError as exc:
-            log.warning("discovery: own-channel resolve failed (user %s): %s", user.id, exc)
+            # No traceback anywhere in this file's YouTube handlers: for an upstream answer
+            # `_upstream_failure` already logged the structured `youtube.api.error`, and the
+            # message names the call, its status and its reason (or, with no answer at all, the
+            # transport failure). The stack under it is this `except` and the client call.
+            log.warning(
+                "discovery: own-channel resolve failed: %s",
+                exc,
+                extra=logs.event("channels.discovery.own_channel_failed", reason=exc.reason),
+            )
 
     rows = _discovery_rows(db, user)
 
@@ -182,15 +199,23 @@ def discover_channels(
                 quota.attribute(user.id, quota.QuotaAction.CHANNELS_DISCOVER),
                 YouTubeClient(db, user) as yt,
             ):
-                apply_channel_details(db, yt.get_channels(need))
+                apply_channel_details(db, yt.get_channels(need), need)
             db.commit()
-            # Match _discovery_rows' ORDER BY exactly: video-count DESC, then title (NULLs LAST,
+            # Match _discovery_rows' ORDER BY exactly: video-count DESC, then title in the reader's
-            # as Postgres orders them — coercing NULL→"" would wrongly float untitled channels up).
+            # alphabet (NULLs LAST, as Postgres orders them — coercing NULL→"" would wrongly float
+            # untitled channels up).
+            titles = [r[0].title for r in rows if r[0].title is not None]
+            alpha = {w: i for i, w in enumerate(sorted_words(db, titles, lang_for_user(user)))}
             rows = sorted(
-                rows, key=lambda r: (-int(r[1]), r[0].title is None, (r[0].title or "").lower())
+                rows,
+                key=lambda r: (-int(r[1]), r[0].title is None, alpha.get(r[0].title or "", 0)),
             )
         except YouTubeError as exc:
-            log.warning("discovery: channel enrich failed (user %s): %s", user.id, exc)
+            log.warning(
+                "discovery: channel enrich failed: %s",
+                exc,
+                extra=logs.event("channels.discovery.enrich_failed", reason=exc.reason),
+            )
             db.rollback()
 
     return [
@@ -284,21 +309,30 @@ def channel_detail(
     """One channel's "About" detail for the channel page header/About tab, plus this user's
     relationship to it (subscribed / explored) and how many of its videos are in the catalog.
     Lazily fills the About metadata (view count, join date, banner, links) the first time it's
-    needed — a stub, or a channel last enriched before those fields existed (published_at is a
+    needed (a stub) and refreshes it once it is stale (`DETAILS_MAX_AGE`), which is how a channel nobody subscribes to (an explored one, outside the
-    reliable sentinel: every real channel has a creation date). The enrich uses the API key
+    subscription import) gets a banner that still resolves. The enrich uses the API key (1 unit,
-    (1 unit, cached after), so it's skipped for the quota-less demo account."""
+    cached after), so it's skipped for the quota-less demo account."""
     channel = get_or_404(db, Channel, channel_id, what="channel")
-    if channel.published_at is None and not user.is_demo:
+    # No `published_at is None` sentinel: a terminated channel never gets one, so it re-fetched on
+    # every view; rows enriched before that field existed are long past DETAILS_MAX_AGE anyway.
+    stale = details_stale(channel, datetime.now(timezone.utc))
+    if stale and not user.is_demo:
         try:
             with (
                 quota.attribute(user.id, quota.QuotaAction.CHANNELS_DISCOVER),
                 YouTubeClient(db, user) as yt,
             ):
-                apply_channel_details(db, yt.get_channels([channel_id]))
+                apply_channel_details(db, yt.get_channels([channel_id]), [channel_id])
             db.commit()
             channel = db.get(Channel, channel_id)
         except YouTubeError as exc:
-            log.warning("channel detail enrich failed (%s): %s", channel_id, exc)
+            log.warning(
+                "channel detail enrich failed: %s",
+                exc,
+                extra=logs.event(
+                    "channels.detail.enrich_failed", channel_id=channel_id, reason=exc.reason
+                ),
+            )
             db.rollback()
     sub = db.execute(
         select(Subscription).where(
@@ -366,12 +400,9 @@ def explore_channel(
     backfill quota reserve; blocked for the demo account (it spends shared quota)."""
     channel = get_or_404(db, Channel, channel_id, what="channel")
     if _is_blocked(db, user, channel_id):
-        raise HTTPException(status_code=403, detail="You've blocked this channel.")
+        raise refused(403, "channel_blocked")
     if quota.backfill_reserve_reached():
-        raise HTTPException(
+        raise refused(429, "quota_reserve")
-            status_code=429,
-            detail="The shared daily quota reserve is reached — try exploring later.",
-        )
 
     # Record/refresh the exploration (DB only) before ingesting, so the explorer can see the
     # videos as soon as they land. Flag the channel ephemeral unless it's organically followed.
@@ -397,18 +428,17 @@ def explore_channel(
             YouTubeClient(db, user) as yt,
         ):
             # Enrich for the About panel + the uploads playlist id (needed to ingest videos).
-            if (
+            # Stale, not "a field is missing": a terminated channel never gets an uploads playlist
-                channel.details_synced_at is None
+            # or a creation date, and asking again on every explore would spend quota on a miss.
-                or channel.uploads_playlist_id is None
+            if details_stale(channel, datetime.now(timezone.utc)):
-                or channel.published_at is None
+                apply_channel_details(db, yt.get_channels([channel_id]), [channel_id])
-            ):
-                apply_channel_details(db, yt.get_channels([channel_id]))
                 db.commit()
                 channel = db.get(Channel, channel_id)
             result = explore_ingest_page(db, yt, channel, page_token)
     except YouTubeError as exc:
-        raise HTTPException(status_code=422, detail=f"YouTube couldn't load this channel: {exc}")
+        raise youtube_refusal(exc) from exc
     return {
+        "ok": True,
         "channel_id": channel_id,
         "ingested": result["ingested"],
         "next_page_token": result["next_page_token"],
@@ -490,6 +520,7 @@ def update_channel(
         background.add_task(backfill_channel_now, channel_id, user.id)
 
     return {
+        "ok": True,
         "id": channel_id,
         "priority": sub.priority,
         "hidden": sub.hidden,
@@ -524,7 +555,7 @@ def reset_backfill(
     db.commit()
     with quota.attribute(user.id, quota.QuotaAction.VIDEOS_BACKFILL_RECENT):
         run_recent_backfill(db, [channel], max_channels=1)
-    return {"id": channel_id, "reset": True}
+    return {"ok": True, "id": channel_id, "reset": True}
 
 
 @router.post("/{channel_id}/subscribe")
@@ -545,7 +576,7 @@ def subscribe(
         )
     ).scalar_one_or_none()
     if existing is not None:
-        raise HTTPException(status_code=409, detail="Already subscribed to this channel.")
+        raise refused(409, "already_subscribed")
     with (
         quota.attribute(user.id, quota.QuotaAction.CHANNELS_SUBSCRIBE),
         YouTubeClient(db, user) as yt,
@@ -559,22 +590,31 @@ def subscribe(
             # subscription resync fills in the resource id. Any other YouTube error is a real
             # problem — surface it with a clear message (422, not a 502 that the client would
             # mistake for a transient "connection lost").
-            msg = str(exc).lower()
+            # On the reason code, never on the prose: the error no longer carries the body (it is
-            if "already exists" in msg or "duplicate" in msg:
+            # the request quoted back), and a message YouTube rewords would silently turn "already
+            # following" back into a 422 the user cannot act on. `reason` is comma-joined when the
+            # body listed several.
+            if "subscriptionDuplicate" in exc.reason.split(","):
                 yt_sub_id = ""
             else:
-                raise HTTPException(
+                raise youtube_refusal(exc) from exc
-                    status_code=422, detail=f"YouTube couldn't subscribe you: {exc}"
-                )
         # Enrich a stub channel (title/thumbnail/subscriber count) so it shows up properly right
         # away, on BOTH the normal and the desync path — a resurfaced stub still needs metadata to
         # render. Keep the client open (we're still inside the `with`); a metadata failure here
         # must not block recording the subscription the user asked for.
         if channel.details_synced_at is None:
             try:
-                apply_channel_details(db, yt.get_channels([channel_id]))
+                apply_channel_details(db, yt.get_channels([channel_id]), [channel_id])
             except YouTubeError as exc:
-                log.warning("subscribe: stub enrich failed (%s): %s", channel_id, exc)
+                log.warning(
+                    "subscribe: stub enrich failed: %s",
+                    exc,
+                    extra=logs.event(
+                        "channels.subscribe.enrich_failed",
+                        channel_id=channel_id,
+                        reason=exc.reason,
+                    ),
+                )
     db.add(
         Subscription(
             user_id=user.id,
@@ -594,7 +634,7 @@ def subscribe(
             .values(via_explore=False)
         )
     db.commit()
-    return {"subscribed": channel_id}
+    return {"ok": True, "subscribed": channel_id}
 
 
 @router.delete("/{channel_id}/subscription")
@@ -608,10 +648,7 @@ def unsubscribe(
     should Hide instead."""
     sub = _user_subscription(db, user, channel_id)
     if not sub.yt_subscription_id:
-        raise HTTPException(
+        raise refused(400, "no_subscription_id")
-            status_code=400,
-            detail="No YouTube subscription id on record — sync subscriptions first.",
-        )
     try:
         with (
             quota.attribute(user.id, quota.QuotaAction.CHANNELS_UNSUBSCRIBE),
@@ -621,10 +658,10 @@ def unsubscribe(
     except YouTubeError as exc:
         # 422 (a real, explainable error → speaking modal), not 502 which the client treats as
         # a transient "connection lost" gateway blip.
-        raise HTTPException(status_code=422, detail=f"YouTube couldn't unsubscribe you: {exc}")
+        raise youtube_refusal(exc) from exc
     db.delete(sub)
     db.commit()
-    return {"unsubscribed": channel_id}
+    return {"ok": True, "unsubscribed": channel_id}
 
 
 @router.get("/blocked/list")
@@ -635,7 +672,7 @@ def list_blocked(user: User = Depends(current_user), db: Session = Depends(get_d
             select(Channel)
             .join(BlockedChannel, BlockedChannel.channel_id == Channel.id)
             .where(BlockedChannel.user_id == user.id)
-            .order_by(func.lower(Channel.title))
+            .order_by(collated(Channel.title, lang_for_user(user)))
         )
         .scalars()
         .all()
@@ -665,7 +702,7 @@ def block_channel(
             )
         )
         db.commit()
-    return {"blocked": channel_id}
+    return {"ok": True, "blocked": channel_id}
 
 
 @router.delete("/{channel_id}/block")
@@ -680,7 +717,7 @@ def unblock_channel(
         )
     )
     db.commit()
-    return {"unblocked": channel_id}
+    return {"ok": True, "unblocked": channel_id}
 
 
 class AttachTagIn(BaseModel):
@@ -715,7 +752,7 @@ def attach_tag(
             )
         )
         db.commit()
-    return {"channel_id": channel_id, "tag_id": body.tag_id}
+    return {"ok": True, "channel_id": channel_id, "tag_id": body.tag_id}
 
 
 @router.delete("/{channel_id}/tags/{tag_id}")
@@ -735,4 +772,4 @@ def detach_tag(
     if row is not None:
         db.delete(row)
         db.commit()
-    return {"channel_id": channel_id, "tag_id": tag_id}
+    return {"ok": True, "channel_id": channel_id, "tag_id": tag_id}

backend/app/routes/config.py

@@ -4,18 +4,23 @@ app.sysconfig. Secret values (e.g. SMTP password) are write-only — never retur
 import re
 from typing import Any
 
-from fastapi import APIRouter, Depends, HTTPException
+from fastapi import APIRouter, Depends
 from pydantic import BaseModel
 from sqlalchemy.orm import Session
 
 from app import audit
 from app import email as email_mod
+from app import logs
 from app import mailtext
 from app import sysconfig
 from app.audit import AuditAction
 from app.auth import admin_user
 from app.db import get_db
 from app.models import User
+from app.refusal import refused
+from app.issues import gitea
+from app.routes._common import issue_refused, not_found
+from app.refusal import setting_refused
 
 router = APIRouter(prefix="/api/admin/config", tags=["admin"])
 
@@ -29,6 +34,14 @@ def _redact(v):
     return _URL_USERINFO.sub(r"\1***@", v) if isinstance(v, str) else v
 
 
+def _apply_runtime(db: Session, key: str) -> None:
+    """Some knobs change how THIS process behaves, not just what a later read returns — they have to
+    take effect now, without a restart. One dispatch for both call sites (set and reset), so the next
+    such key can't be wired into one of them and forgotten in the other."""
+    if key == "log_level":
+        logs.set_level(sysconfig.get_str(db, "log_level"))
+
+
 @router.get("")
 def get_config(_: User = Depends(admin_user), db: Session = Depends(get_db)) -> dict:
     return sysconfig.describe(db)
@@ -50,19 +63,16 @@ def set_config(
 ) -> dict:
     s = sysconfig.spec(key)
     if s is None:
-        raise HTTPException(status_code=404, detail="Unknown config key")
+        raise not_found("config_key")
     if "value" not in body.model_fields_set:
-        raise HTTPException(status_code=400, detail="Missing 'value'")
+        raise refused(400, "value_missing")
     if s.secret and not sysconfig.secrets_manageable():
-        raise HTTPException(
+        raise refused(400, "no_encryption_key")
-            status_code=400,
-            detail="Set TOKEN_ENCRYPTION_KEY to store secrets in the database.",
-        )
     old = None if s.secret else sysconfig.get(db, key)
     try:
         sysconfig.set_value(db, key, body.value)
-    except (TypeError, ValueError) as exc:
+    except sysconfig.InvalidSetting as exc:
-        raise HTTPException(status_code=400, detail=str(exc) or "Invalid value")
+        raise setting_refused(exc) from None
     # Never log a secret's value — only that it was changed, by whom. Non-secret values are logged
     # but with URL credentials redacted (a proxy/URL setting can embed user:pass). Skip the row when
     # a non-secret value didn't actually change (a no-op re-save shouldn't add trail noise); secrets
@@ -80,6 +90,7 @@ def set_config(
             after=None if s.secret else {"value": _redact(new)},
         )
         db.commit()
+    _apply_runtime(db, key)
     return sysconfig.describe(db)
 
 
@@ -91,7 +102,7 @@ def reset_config(
 ) -> dict:
     s = sysconfig.spec(key)
     if s is None:
-        raise HTTPException(status_code=404, detail="Unknown config key")
+        raise not_found("config_key")
     old = None if s.secret else sysconfig.get(db, key)
     sysconfig.reset(db, key)
     audit.record(
@@ -104,6 +115,7 @@ def reset_config(
         before=None if s.secret else {"value": _redact(old)},
     )
     db.commit()
+    _apply_runtime(db, key)
     return sysconfig.describe(db)
 
 
@@ -115,8 +127,21 @@ def send_test_email(
     """Send a test email to the requesting admin using the current (DB or env) SMTP config —
     lets the admin verify the settings actually work."""
     if not email_mod.email_enabled():
-        raise HTTPException(status_code=400, detail="SMTP is not configured.")
+        raise refused(400, "smtp_off")
     sent = email_mod.send_test(user.email, mailtext.lang_for_user(user))
     if not sent:
-        raise HTTPException(status_code=422, detail="SMTP send failed — check the settings.")
+        raise refused(422, "smtp_failed")
-    return {"sent": True, "to": user.email}
+    return {"ok": True, "sent": True, "to": user.email}
+
+
+@router.post("/test-issues")
+def test_issue_tracker(
+    _: User = Depends(admin_user),
+    db: Session = Depends(get_db),
+) -> dict:
+    """Check the configured issue tracker (E16): the repository answers with this token, takes
+    issues, and is PRIVATE — the one it must be, since an issue carries its reporter's log."""
+    try:
+        return {"ok": True, **gitea.check(db)}
+    except gitea.GiteaError as exc:
+        raise issue_refused(exc) from None

backend/app/routes/debug_sessions.py

@@ -0,0 +1,418 @@
+"""Debug sessions (E16): Start → reproduce → Stop, the screenshots, and the evidence under review.
+
+The session itself — what tags a request, its cap, its level — is `app/debugsession.py`. These routes
+are the user's handle on their OWN sessions only: another user's id 404s (`owned_or_404`), including
+for an admin, who reads everyone's lines in the diagnostics console instead. The shared demo account
+cannot record: its log would interleave every visitor under one account.
+
+Screenshots arrive as a raw image body (`image/png`, `image/jpeg`, `image/webp`), not a multipart
+form — one image per request, capped while it arrives, and kept only if its bytes really are one of
+those types, whatever the Content-Type claimed.
+"""
+
+import json
+from typing import Any
+
+from fastapi import APIRouter, Depends, HTTPException, Request, Response
+from pydantic import BaseModel, Field
+from sqlalchemy import select
+from sqlalchemy.exc import IntegrityError
+from sqlalchemy.orm import Session
+
+from app import audit, debugsession, logs, logsink
+from app.audit import AuditAction
+from app.auth import require_human
+from app.db import get_db, release_read_connection
+from app.issues import bundle, gitea
+from app.ratelimit import RateLimiter
+from app.refusal import refused
+from app.models import DebugScreenshot, DebugSession, User
+from app.routes._common import capped_body, issue_refused, not_found, owned_or_404
+from app.utils import iso
+
+router = APIRouter(prefix="/api/debug-sessions", tags=["diagnostics"])
+log = logs.get_logger(__name__)
+
+# Sessions listed in Settings → Diagnosztika, newest first: enough to reach last week's unsent draft.
+LIST_MAX = 20
+
+
+class StartIn(BaseModel):
+    level: debugsession.SessionLevel
+
+
+_view = debugsession.view
+
+# New sessions a user may open per hour (joining the running one is not counted): a session is a
+# place to keep screenshots, and this — with the per-user storage cap — bounds how many there are.
+_new_sessions = RateLimiter(max_events=20, window_seconds=3600)
+
+
+@router.post("")
+def start_session(
+    body: StartIn,
+    user: User = Depends(require_human),
+    db: Session = Depends(get_db),
+) -> dict:
+    """Start recording, or join the recording already running (another tab started it)."""
+    joining = debugsession.open_session(db, user) is not None
+    if not joining and not _new_sessions.allow(str(user.id)):
+        raise refused(429, "debug_session_rate_limited", limit=_new_sessions.max_events)
+    try:
+        session = debugsession.start(db, user, body.level)
+        db.commit()
+    except IntegrityError:
+        # A concurrent Start won the one-open-per-user index: join that one.
+        db.rollback()
+        session = debugsession.open_session(db, user)
+        if session is None:
+            raise refused(409, "debug_session_conflict") from None
+    # The rest of THIS request is part of the session too.
+    level, until = debugsession.level_for(user, session)
+    logs.set_user(user.id, level, until, session.id)
+    log.info(
+        "debug session started",
+        extra=logs.event("debug.session.started", level=session.level, until=iso(session.until)),
+    )
+    return {"ok": True, "session": _view(db, session)}
+
+
+@router.post("/{sid}/stop")
+def stop_session(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> dict:
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    if session.ended_at is None:
+        # Logged BEFORE the session closes, while this request still carries its tag: the last line
+        # of the evidence is the Stop itself.
+        log.info("debug session stopped", extra=logs.event("debug.session.stopped"))
+        debugsession.stop(session)
+        db.commit()
+        logs.end_session_elevation(session.id)
+    return {"ok": True, "session": _view(db, session)}
+
+
+@router.get("")
+def list_sessions(user: User = Depends(require_human), db: Session = Depends(get_db)) -> dict:
+    sessions = db.scalars(
+        select(DebugSession)
+        .where(DebugSession.user_id == user.id)
+        .order_by(DebugSession.started_at.desc())
+        .limit(LIST_MAX)
+    ).all()
+    return {"items": [_view(db, s) for s in sessions]}
+
+
+@router.get("/{sid}")
+def get_session(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> dict:
+    return _view(db, owned_or_404(db, DebugSession, sid, user, what="debug_session"))
+
+
+@router.get("/{sid}/events")
+def session_events(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> dict:
+    """The rows an issue from this session would carry — browser and server, in time order."""
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    rows, total = debugsession.bundle(db, session)
+    return {"items": [logsink.serialize(r) for r in rows], "total": total}
+
+
+# --- the draft and its export ---------------------------------------------------------------
+# The composer's own fields (steps, description…) are its business, kept whole so it reopens as it
+# was — bounded here so a draft cannot become a place to store anything else.
+FORM_MAX_BYTES = 64 * 1024
+# Zips a user may build per minute: each holds every screenshot in memory while it is built.
+_exports = RateLimiter(max_events=6, window_seconds=60)
+
+
+class DraftIn(BaseModel):
+    title: str = Field(max_length=200)
+    # The rendered markdown the export and the issue carry.
+    body: str = Field(max_length=50_000)
+    # Log rows taken out of the evidence in the review.
+    excluded: list[int] = Field(default_factory=list, max_length=debugsession.BUNDLE_MAX)
+    form: dict[str, Any] = Field(default_factory=dict)
+
+
+@router.put("/{sid}/draft")
+def save_draft(
+    sid: str,
+    body: DraftIn,
+    user: User = Depends(require_human),
+    db: Session = Depends(get_db),
+) -> dict:
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    if len(json.dumps(body.form, ensure_ascii=False).encode()) > FORM_MAX_BYTES:
+        raise HTTPException(status_code=413, detail={"code": "too_large"})
+    # Under the lock, and refused while a send is under way: the issue is being built from the draft
+    # as it was saved, and this would change it under the send.
+    debugsession.lock(db, session)
+    if debugsession.sending(session):
+        raise refused(409, "issue_sending")
+    if session.issue_url is not None:
+        # Sent: the issue is on the tracker now, and a draft edited after it would say otherwise.
+        raise refused(409, "issue_already_sent")
+    debugsession.save_draft(session, body.model_dump())
+    db.commit()
+    return {"ok": True, "session": _view(db, session)}
+
+
+@router.get("/{sid}/export")
+def export_session(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> Response:
+    """The issue as a zip — the text, the two logs and the screenshots — for an instance without an
+    issue tracker, or for sending it by hand. Built from the SAVED draft: what the composer shows is
+    what it saved."""
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    if not _exports.allow(str(user.id)):
+        raise refused(429, "export_rate_limited")
+    data = bundle.zipped(bundle.files(db, session))
+    name = f"siftlode-debug-{session.started_at:%Y%m%d-%H%M}-{session.id[:6]}.zip"
+    return Response(
+        content=data,
+        media_type="application/zip",
+        headers={
+            "Content-Disposition": f'attachment; filename="{name}"',
+            "Cache-Control": "no-store",
+        },
+    )
+
+
+# --- the issue tracker (S4) -----------------------------------------------------------------
+# Issues a user may open per hour — counted once the tracker is ASKED to open one (an answer that
+# goes missing may still have created it, so an attempt cannot be given back); finishing a half-sent
+# one, or an attempt refused before the tracker was asked, is not counted. The composer sends one at
+# a time; this is room for a bad afternoon, not a pace anything legitimate keeps up.
+_new_issues = RateLimiter(max_events=5, window_seconds=3600)
+# Attempts on the tracker per hour, creating or finishing. Deliberately NOT the budget above: a
+# half-sent issue must always be finishable, so the pace that guards creating one must never be what
+# stands between a user and their own report. This one only bounds a loop — every attempt re-reads
+# the whole bundle into memory and re-offers the pending file to the tracker, and a retry that is
+# refused for good (the tracker's own limits) never leaves `pending`. No hand reaches it.
+_sends = RateLimiter(max_events=60, window_seconds=3600)
+# The composer's live preview, rendered BY the tracker: one authenticated call each, with our token.
+# The composer's toggle fires it once per switch INTO preview — the editor is the other branch, so
+# no keystroke can change the query key while the preview is up — and react-query's cache makes an
+# unchanged draft free. (It does NOT debounce, whatever this comment used to say: a key that changes
+# per keystroke is one fetch per keystroke, measured. Showing the preview BESIDE the fields would
+# need a real debounce first.) This bounds what a loop can aim at the tracker.
+_previews = RateLimiter(max_events=60, window_seconds=60)
+
+
+def _issues_rate_limited() -> HTTPException:
+    return refused(429, "issues_rate_limited", limit=_new_issues.max_events)
+
+
+@router.post("/{sid}/issue")
+def send_issue(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> dict:
+    """Send the session's issue to the tracker — from the SAVED draft, as the export is built — or
+    finish a send that stopped half-way (the issue exists; only what is missing is sent)."""
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    # The session is CLAIMED for the send rather than locked through it: the tracker is slow (a
+    # dozen calls, 20 s each), and a row lock held across them would queue this user's Stop, their
+    # autosave and their screenshots behind it. They see the claim instead and are refused at once.
+    debugsession.lock(db, session)
+    if debugsession.sending(session):
+        raise refused(409, "issue_sending")
+    new = session.issue_url is None
+    if not new and not debugsession.unfinished(session):
+        return {"ok": True, "session": _view(db, session)}
+    key = str(user.id)
+    if new:
+        if not ((session.draft or {}).get("title") or "").strip():
+            raise refused(422, "issue_title_required")
+        if _new_issues.exhausted(key):
+            raise _issues_rate_limited()
+    if not _sends.allow(key):
+        raise refused(429, "issue_rate_limited", limit=_sends.max_events)
+    token = debugsession.claim_send(session)
+    db.commit()  # the claim is what everything else sees; the lock ends here
+
+    def count_new_issue() -> None:
+        if not _new_issues.allow(key):
+            raise _issues_rate_limited()
+
+    try:
+        gitea.deliver(
+            db,
+            session,
+            token=token,
+            checkpoint=db.commit,
+            before_create=count_new_issue,
+            on_created=lambda: _record_created(db, user, session),
+        )
+    except debugsession.SendClaimLost:
+        # This send outlived its lease and another one took the session over while it was still
+        # talking to the tracker. It stopped rather than writing over that one's progress — and a
+        # send IS under way, just not ours, which is what the caller needs to hear.
+        raise refused(409, "issue_sending") from None
+    except gitea.GiteaError as exc:
+        # Keep what did happen: an issue created before an upload failed stays linked, and the
+        # retry does only what is still missing.
+        raise issue_refused(exc) from None
+    except HTTPException:
+        raise  # the pace, refused before the tracker was asked
+    finally:
+        _end_send(db, session, token)
+    return {"ok": True, "session": _view(db, session)}
+
+
+def _end_send(db: Session, session: DebugSession, token: str) -> None:
+    """Give the claim back, whatever else happened. A failure here must not replace the failure that
+    brought us: a dead transaction (the database went away mid-send) is rolled back, and the claim
+    then simply ages out."""
+    try:
+        debugsession.lock(db, session)
+        debugsession.release_send(session, token)
+        db.commit()
+    except Exception:  # noqa: BLE001 — whatever went wrong here, the send's own failure is the news
+        db.rollback()
+        log.warning(
+            "a send could not give its claim back",
+            extra=logs.event("issues.claim.stuck", session_id=session.id),
+        )
+
+
+def _record_created(db: Session, user: User, session: DebugSession) -> None:
+    """The trail for the one act that takes a user's log off this instance. Staged inside the send,
+    so the checkpoint that first writes `issue_url` commits BOTH — recorded later it could be lost
+    on its own (the process dies mid-send), and no retry could write it afterwards: the next send
+    sees the link and knows the issue is not new."""
+    audit.record(
+        db,
+        user.id,
+        AuditAction.ISSUE_CREATE,
+        target_type="debug_session",
+        target_id=session.id,
+        summary=session.issue_url or "",
+    )
+
+
+class PreviewIn(BaseModel):
+    text: str = Field(max_length=50_000)
+
+
+@router.post("/{sid}/preview")
+def preview_issue(
+    sid: str,
+    body: PreviewIn,
+    user: User = Depends(require_human),
+    db: Session = Depends(get_db),
+) -> dict:
+    """The body as the tracker will render it. Gitea sanitises its own output; the composer still
+    shows it in a sandboxed frame, never in the app's own document."""
+    owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    if not _previews.allow(str(user.id)):
+        raise refused(429, "preview_rate_limited")
+    try:
+        return {"ok": True, "html": gitea.preview(db, body.text)}
+    except gitea.GiteaError as exc:
+        raise issue_refused(exc) from None
+
+
+# --- screenshots ---------------------------------------------------------------------------
+def _shot_view(shot: DebugScreenshot) -> dict:
+    return {"id": shot.id, "mime": shot.mime, "size": shot.size, "created_at": iso(shot.created_at)}
+
+
+@router.get("/{sid}/screenshots")
+def list_screenshots(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> dict:
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    return {"items": [_shot_view(s) for s in debugsession.screenshots(db, session)]}
+
+
+def _upload_target(
+    sid: str, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> DebugSession:
+    return owned_or_404(db, DebugSession, sid, user, what="debug_session")
+
+
+async def _image_body(request: Request, db: Session = Depends(get_db)) -> bytes:
+    # The owner check above has already run, and it was a SELECT — so this session is holding a
+    # pooled connection `idle in transaction` and would hold it for the whole 10 MiB upload, from a
+    # client that may be on a phone. Hand it back first; the handler's own work re-takes one. The
+    # `session` object the handler holds is expired by this and reloads on its next attribute read,
+    # which is the correct moment to touch the database again.
+    release_read_connection(db)
+    return await capped_body(request, debugsession.SCREENSHOT_MAX_BYTES)
+
+
+@router.post("/{sid}/screenshots")
+def add_screenshot(
+    # The owner check FIRST: dependencies resolve in order, so a stranger's upload is refused
+    # before a byte of it is read.
+    session: DebugSession = Depends(_upload_target),
+    data: bytes = Depends(_image_body),
+    db: Session = Depends(get_db),
+) -> dict:
+    try:
+        shot = debugsession.add_screenshot(db, session, data)
+    except debugsession.ScreenshotRefused as exc:
+        raise _screenshot_refused(exc.code) from None
+    db.commit()
+    return {"ok": True, "screenshot": _shot_view(shot)}
+
+
+def _screenshot_refused(code: str) -> HTTPException:
+    # Literal reasons, one per case: the frontend's contract test reads them to require a sentence
+    # for each in both languages.
+    if code == "already_sent":
+        return refused(409, "screenshot_already_sent")
+    if code == "too_many":
+        return refused(422, "screenshot_too_many", limit=debugsession.SCREENSHOTS_MAX)
+    if code == "sending":
+        return refused(409, "issue_sending")
+    if code == "storage_full":
+        mb = debugsession.SCREENSHOT_USER_MAX_BYTES // (1024 * 1024)
+        return refused(422, "screenshot_storage_full", limit=mb)
+    if code == "empty":
+        return refused(422, "screenshot_empty")
+    return refused(422, "screenshot_not_an_image")
+
+
+def _session_screenshot(db: Session, sid: str, shot_id: int, user: User) -> DebugScreenshot:
+    return _screenshot_of(
+        db, owned_or_404(db, DebugSession, sid, user, what="debug_session"), shot_id
+    )
+
+
+def _screenshot_of(db: Session, session: DebugSession, shot_id: int) -> DebugScreenshot:
+    shot = db.get(DebugScreenshot, shot_id)
+    if shot is None or shot.session_id != session.id:
+        raise not_found("screenshot")
+    return shot
+
+
+@router.get("/{sid}/screenshots/{shot_id}")
+def get_screenshot(
+    sid: str, shot_id: int, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> Response:
+    shot = _session_screenshot(db, sid, shot_id, user)
+    # `nosniff`: the type was checked against the bytes at upload, and the browser must not guess.
+    return Response(
+        content=shot.data,
+        media_type=shot.mime,
+        headers={"Cache-Control": "private, max-age=3600", "X-Content-Type-Options": "nosniff"},
+    )
+
+
+@router.delete("/{sid}/screenshots/{shot_id}")
+def delete_screenshot(
+    sid: str, shot_id: int, user: User = Depends(require_human), db: Session = Depends(get_db)
+) -> dict:
+    session = owned_or_404(db, DebugSession, sid, user, what="debug_session")
+    shot = _screenshot_of(db, session, shot_id)
+    try:
+        debugsession.remove_screenshot(db, session, shot)
+    except debugsession.ScreenshotRefused as exc:
+        raise _screenshot_refused(exc.code) from None
+    db.commit()
+    return {"ok": True, "deleted": shot_id}

backend/app/routes/diagnostics.py

@@ -0,0 +1,190 @@
+"""The browser's log, sent while its user has a raised diagnostics level (E22 S5).
+
+The client keeps its log in memory (frontend `lib/log.ts`). Only after the user consented to Részletes
+or Teljes in Settings → Diagnosztika does it stream that log here, in batches, and each batch lands in
+`log_events` beside the server's own rows: `source="web"`, the account from the SESSION (never from
+the body), and the request id each event carries, which is what joins a browser line to the server
+line of the same request.
+
+Nothing from the body is trusted as-is:
+  * only a JSON Content-Type is read, and the whole body is capped before it is parsed, and so are
+    the number of events and each event's attributes;
+  * the pace is capped per account, and so is what an account may store per hour;
+  * everything is redacted AGAIN with the server's own rules — the client redacts at capture, but a
+    modified client would not;
+  * an event below the consented level is dropped (a Részletes consent does not store what only
+    Teljes may carry), and so is the whole batch once the level has lapsed — where an open debug
+    session the request names counts as consent to ITS level (E16), and only its own events keep
+    its tag;
+  * the timestamp is clamped to the consent window, so a row cannot dodge the retention job by
+    claiming to come from the future.
+"""
+
+import json
+from datetime import datetime, timedelta, timezone
+from typing import Any, Literal
+
+from fastapi import APIRouter, Depends, HTTPException, Request
+from pydantic import BaseModel, Field, ValidationError
+from sqlalchemy import insert
+from sqlalchemy.orm import Session
+
+from app import debugsession, diagnostics, logs, logsink
+from app.auth import current_user
+from app.db import get_db, release_read_connection
+from app.models import DebugSession, LogEvent, User
+from app.ratelimit import RateLimiter
+from app.routes._common import capped_body
+
+router = APIRouter(prefix="/api/diagnostics", tags=["diagnostics"])
+
+BODY_MAX_BYTES = 512 * 1024
+EVENTS_MAX = 200
+ATTRS_MAX_BYTES = 16 * 1024
+# The client flushes every few seconds and when its buffer fills; this is room for a burst, not a
+# pace anything legitimate keeps up for a minute.
+_batches = RateLimiter(max_events=30, window_seconds=60)
+# What one account may STORE per hour. The pace cap alone still lets a self-consented account write
+# ~15 MiB a minute into the one shared database, kept for the whole retention window; a real session
+# at Teljes stores a few hundred kB an hour. Past the budget a batch is accepted and not stored (a
+# refusal would only make the client resend it), and one notice per hour says so.
+STORED_BYTES_PER_HOUR = 5 * 1024 * 1024
+_stored_bytes = RateLimiter(max_events=STORED_BYTES_PER_HOUR, window_seconds=3600)
+_budget_notices = RateLimiter(max_events=1, window_seconds=3600)
+
+ClientLevel = Literal["TRACE", "DEBUG", "INFO", "WARNING", "ERROR", "CRITICAL"]
+
+
+class ClientEvent(BaseModel):
+    ts: datetime
+    level: ClientLevel
+    event: str = Field(max_length=64, pattern=r"^[a-z][a-z0-9_]*(\.[a-z0-9_]+)*$")
+    msg: str = Field(max_length=8_000)
+    attrs: dict[str, Any] | None = None
+    request_id: str | None = Field(default=None, pattern=r"^[A-Za-z0-9._-]{8,64}$")
+    tab_id: str | None = Field(default=None, pattern=r"^[A-Za-z0-9_-]{1,32}$")
+    # The debug session the event was captured in (E16); kept only when it is the one this request
+    # names in `X-Debug-Session` and the auth dependency found it the caller's own and open.
+    debug_session_id: str | None = Field(default=None, pattern=r"^[0-9a-f]{16}$")
+    version: str | None = Field(default=None, max_length=64)
+
+
+class ClientEventsIn(BaseModel):
+    events: list[ClientEvent] = Field(max_length=EVENTS_MAX)
+
+
+async def _parsed_events(request: Request, db: Session = Depends(get_db)) -> ClientEventsIn:
+    """The body as the model — parsed only after the size cap, which is why this is a dependency and
+    not a plain body parameter (FastAPI reads a declared body in full before anything else runs).
+
+    JSON only, by its Content-Type, as a declared body would be. Parsing any body as JSON would let a
+    CORS-simple `text/plain` form post from a same-site page write events into a signed-in user's
+    log; the app's own client always sends `application/json`, the unload beacon included."""
+    media_type = request.headers.get("content-type", "").split(";")[0].strip().lower()
+    if media_type != "application/json" and not media_type.endswith("+json"):
+        raise HTTPException(status_code=415, detail={"code": "json_required"})
+    # `current_user` resolved above on a SELECT, so this session is holding a pooled connection.
+    # Give it back before awaiting the body — the sink is a background beacon and its sender may be
+    # on a link far slower than the 512 KiB suggests. See `db.release_read_connection`.
+    release_read_connection(db)
+    raw = await capped_body(request, BODY_MAX_BYTES)
+    try:
+        return ClientEventsIn.model_validate_json(raw)
+    except ValidationError as exc:
+        raise HTTPException(status_code=422, detail=json.loads(exc.json())) from exc
+
+
+def _row(
+    item: ClientEvent,
+    user: User,
+    now: datetime,
+    window_start: datetime,
+    session: DebugSession | None,
+) -> dict[str, Any]:
+    attrs = logs.redact(item.attrs) if item.attrs else None
+    if attrs is not None:
+        size = len(json.dumps(attrs, ensure_ascii=False, default=str).encode())
+        if size > ATTRS_MAX_BYTES:
+            attrs = {"attrs_dropped_bytes": size}
+    if item.tab_id:
+        attrs = {**(attrs or {}), "tab_id": item.tab_id}
+    ts = item.ts if item.ts.tzinfo else item.ts.replace(tzinfo=timezone.utc)
+    return logsink.to_row(
+        {
+            "ts": min(max(ts, window_start), now),
+            "level": item.level,
+            "sev": logs.SEVERITY[item.level],
+            "source": "web",
+            "logger": "web",
+            "event": item.event,
+            "msg": logs.redact_text(item.msg),
+            "attrs": attrs,
+            "request_id": item.request_id,
+            "user_id": user.id,
+            "debug_session_id": (
+                session.id if session and item.debug_session_id == session.id else None
+            ),
+            "version": item.version,
+        }
+    )
+
+
+@router.post("/events")
+def ingest_events(
+    # The account first: dependencies resolve in order, so an anonymous caller is refused before a
+    # byte of its body is read.
+    request: Request,
+    user: User = Depends(current_user),
+    body: ClientEventsIn = Depends(_parsed_events),
+    db: Session = Depends(get_db),
+) -> dict:
+    # A debug session the request names raises the level for its own events (E16): starting one is
+    # the consent to that level for its duration, whatever the user's own level is.
+    session = debugsession.from_header(db, user, request.headers.get(debugsession.HEADER))
+    level, _ = debugsession.level_for(user, session)
+    if level == diagnostics.BASE:
+        # Lapsed or withdrawn: the client stops streaming when it sees this.
+        raise HTTPException(status_code=409, detail={"code": "diagnostics_off"})
+    if not _batches.allow(str(user.id)):
+        raise HTTPException(status_code=429, detail={"code": "rate_limited"})
+    now = datetime.now(timezone.utc)
+    window_start = now - timedelta(hours=diagnostics.ELEVATION_HOURS)
+    floor = logs.SEVERITY[level]
+    rows = [
+        _row(item, user, now, window_start, session)
+        for item in body.events
+        if logs.SEVERITY[item.level] >= floor
+    ]
+    if not rows:
+        return {"ok": True, "stored": 0}
+    size = sum(len(json.dumps(row, ensure_ascii=False, default=str).encode()) for row in rows)
+    stored = len(rows)
+    if not _stored_bytes.allow(str(user.id), size):
+        stored = 0
+        if not _budget_notices.allow(str(user.id)):
+            return {"ok": True, "stored": 0}
+        rows = [_budget_notice(user, now)]
+    # Under the erasure guard until the commit: an account erased meanwhile gets no rows back.
+    with logs.erasure_guard():
+        db.execute(insert(LogEvent), logsink.without_erased(rows))
+        db.commit()
+    return {"ok": True, "stored": stored}
+
+
+def _budget_notice(user: User, now: datetime) -> dict[str, Any]:
+    return logsink.to_row(
+        {
+            "ts": now,
+            "level": "WARNING",
+            "sev": logs.SEVERITY["WARNING"],
+            "source": "web",
+            "logger": "web",
+            "event": "diagnostics.budget_spent",
+            # What the budget does, exactly: it is a SLIDING hour, and a refused batch is not
+            # charged to it, so smaller batches keep landing while there is room for them. Saying
+            # "the rest of the hour is not kept" would be a row contradicting the rows beneath it.
+            "msg": "The browser log passed its hourly storage budget; a batch bigger than what is left of it is not kept",
+            "attrs": {"budget_bytes": STORED_BYTES_PER_HOUR},
+            "user_id": user.id,
+        }
+    )

backend/app/routes/discover.py

@@ -11,9 +11,9 @@ thing it exists to remove.
 
 from typing import Literal
 
-from fastapi import APIRouter, Depends, HTTPException
+from fastapi import APIRouter, Depends
 from pydantic import BaseModel
-from sqlalchemy import select, update
+from sqlalchemy import select
 from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 
@@ -23,7 +23,8 @@ from app.discover import hero as hero_mod
 from app.discover import pins, rails
 from app.discover.budget import Budget, DEFAULT_BUDGET
 from app.models import ComfortPin, User
-from app.routes._common import owned_or_404
+from app.plex.cards import norm_lang
+from app.routes._common import owned_or_404, not_found
 
 router = APIRouter(prefix="/api/discover", tags=["discover"])
 
@@ -45,6 +46,7 @@ class ReorderIn(BaseModel):
 def get_discover(
     budget: Budget = DEFAULT_BUDGET,
     seed: int = 0,
+    lang: str = "en",
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
@@ -52,11 +54,14 @@ def get_discover(
 
     `resume` and `shelf` are built once and handed to the hero, so the two rails the hero draws from
     are not queried twice for a single page load."""
-    resume = rails.continue_watching(db, user)
+    lang = norm_lang(lang)
-    shelf = pins.shelf(db, user)
+    resume = rails.continue_watching(db, user, lang=lang)
+    # Plex off → the shelf rail (and so the hero's comfort tier) goes YouTube-only like the rails
+    # beside it; the pins are kept and `GET /pins` still lists them (`pins.shelf`).
+    shelf = pins.shelf(db, user, include_plex=rails.plex_available(db), lang=lang)
     return {
         "budget": budget,
-        "hero": hero_mod.pick(db, user, budget, seed, resume=resume, shelf=shelf),
+        "hero": hero_mod.pick(db, user, budget, seed, resume=resume, shelf=shelf, lang=lang),
         "continue_watching": resume,
         "comfort": shelf,
     }
@@ -78,6 +83,7 @@ def get_pick(
     seed: int = 0,
     exclude: str = "",
     start: int = 0,
+    lang: str = "en",
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
@@ -97,18 +103,25 @@ def get_pick(
     return {
         "budget": budget,
         "hero": hero_mod.pick(
-            db, user, budget, seed, exclude=set(list(skip)[:MAX_EXCLUDE]), start=start
+            db,
+            user,
+            budget,
+            seed,
+            exclude=set(list(skip)[:MAX_EXCLUDE]),
+            start=start,
+            lang=norm_lang(lang),
         ),
     }
 
 
 @router.get("/pins")
 def list_pins(
+    lang: str = "en",
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
     """The shelf on its own — for the shelf editor, which needs it without the rest of the page."""
-    return {"items": pins.shelf(db, user)}
+    return {"items": pins.shelf(db, user, lang=norm_lang(lang))}
 
 
 @router.post("/pins")
@@ -125,7 +138,7 @@ def add_pin(
     constraint is the mechanism, not the message."""
     resolved = pins.resolve_ref(db, user, body.kind, body.ref)
     if resolved is None:
-        raise HTTPException(status_code=404, detail="Unknown item")
+        raise not_found("item")
     column, value = resolved
 
     existing = db.execute(
@@ -157,7 +170,7 @@ def add_pin(
             # between `resolve_ref` confirming it and this INSERT (a Plex sync dropping a removed
             # item), which trips the FK. Semantically that is the same "no such thing" the pre-check
             # returns above — re-raising handed the user a 500 they could do nothing with.
-            raise HTTPException(status_code=404, detail="Unknown item")
+            raise not_found("item") from None
         return {"ok": True, "id": winner.id, "created": False}
     return {"ok": True, "id": pin.id, "created": True}
 
@@ -180,18 +193,11 @@ def reorder_pins(
     user: User = Depends(require_human),
     db: Session = Depends(get_db),
 ) -> dict:
-    """Rewrite the shelf order from a full list of pin ids.
+    """Rewrite the shelf order from the list of pin ids the client shows.
 
     Whole-list rather than a move-one delta: the shelf is short, and a full rewrite cannot leave two
-    tiles claiming the same slot the way an interrupted sequence of single moves can. Ids that are not
+    tiles claiming the same slot the way an interrupted sequence of single moves can. The list may be
-    this user's are ignored rather than 404'd — a stale tab replaying an order that includes a pin
+    PARTIAL (hidden Plex pins, a stale tab) — `pins.reorder` keeps what it leaves out in place."""
-    deleted meanwhile should reorder what still exists, not fail wholesale."""
+    pins.reorder(db, user, body.ids)
-    owned = {p.id for p in pins.list_pins(db, user)}
-    for position, pin_id in enumerate(i for i in body.ids if i in owned):
-        db.execute(
-            update(ComfortPin)
-            .where(ComfortPin.id == pin_id, ComfortPin.user_id == user.id)
-            .values(position=position)
-        )
     db.commit()
     return {"ok": True}

backend/app/routes/downloads.py

@@ -9,20 +9,24 @@ serves finished files (range-aware, with the user's custom display name), and sh
 import os
 import re
 import tempfile
+from collections.abc import AsyncIterator
+from dataclasses import dataclass
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
-from urllib.parse import parse_qs, urlparse
+from urllib.parse import parse_qs, unquote, urlparse
 
-from fastapi import APIRouter, Depends, File, Form, HTTPException, Request, UploadFile
+from fastapi import APIRouter, Depends, HTTPException, Request
 from fastapi.responses import FileResponse
 from pydantic import BaseModel
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
+from app.mailtext import lang_for_user
+from app.collation import collated
 from app import audit, sysconfig
 from app.auth import admin_user, require_human
 from app.config import settings
-from app.db import get_db
+from app.db import get_db, release_read_connection
 from app.downloads import credentials as dlcreds
 from app.downloads import edit as editmod
 from app.downloads import links as linksmod
@@ -37,11 +41,13 @@ from app.models import (
     DownloadShare,
     MediaAsset,
     User,
+    display_text,
 )
-from app.routes._common import get_or_404, not_found
+from app.routes._common import capped_stream_to, get_or_404, not_found, owned_or_404
 from app.security import hash_password
 from app.userscope import is_messageable_user, messageable_clauses
 from app.utils import iso
+from app.refusal import refused
 
 router = APIRouter(prefix="/api/downloads", tags=["downloads"])
 admin_router = APIRouter(prefix="/api/admin/downloads", tags=["admin-downloads"])
@@ -63,7 +69,7 @@ def resolve_source(raw: str) -> tuple[str, str]:
     as a raw URL for yt-dlp's generic extractor."""
     raw = (raw or "").strip()
     if not raw:
-        raise HTTPException(status_code=400, detail="A video link or id is required.")
+        raise refused(400, "link_required")
     if _VIDEO_ID.match(raw):
         return "youtube", raw
     if raw.startswith(("http://", "https://")):
@@ -93,7 +99,7 @@ def resolve_source(raw: str) -> tuple[str, str]:
             if _VIDEO_ID.match(vid):
                 return "youtube", vid
         return "url", raw
-    raise HTTPException(status_code=400, detail="That doesn't look like a valid video link.")
+    raise refused(400, "invalid_link")
 
 
 def _serialize(job: DownloadJob, asset: MediaAsset | None) -> dict:
@@ -166,15 +172,8 @@ def _serialize_job(db: Session, job: DownloadJob) -> dict:
     return _serialize(job, asset)
 
 
-def _own_job(db: Session, user: User, job_id: int) -> DownloadJob:
-    job = db.get(DownloadJob, job_id)
-    if job is None or job.user_id != user.id:
-        raise HTTPException(status_code=404, detail="Unknown download")
-    return job
-
-
 def _own_job_for_update(db: Session, user: User, job_id: int) -> DownloadJob:
-    """`_own_job` for a transition that moves its asset's ref count (resume, cancel, delete, redownload)
+    """The ownership check (`owned_or_404`) for a transition that moves its asset's ref count (resume, cancel, delete, redownload)
     or writes the status those decide on (pause). The job is re-read under a row lock, so two
     overlapping requests for one job run one after the other and the second decides on the first's
     committed result — never both releasing the same hold, and never a pause turning a job a cancel
@@ -183,7 +182,7 @@ def _own_job_for_update(db: Session, user: User, job_id: int) -> DownloadJob:
     The asset is locked BEFORE the job: the order a finishing worker child takes them in
     (`service.finalize_if_running`), so the two can never deadlock."""
     while True:
-        asset_id = _own_job(db, user, job_id).asset_id
+        asset_id = owned_or_404(db, DownloadJob, job_id, user, what="download").asset_id
         if asset_id is not None:
             db.execute(select(MediaAsset.id).where(MediaAsset.id == asset_id).with_for_update())
         job = db.execute(
@@ -193,7 +192,7 @@ def _own_job_for_update(db: Session, user: User, job_id: int) -> DownloadJob:
             .execution_options(populate_existing=True)
         ).scalar_one_or_none()
         if job is None:
-            raise HTTPException(status_code=404, detail="Unknown download")
+            raise not_found("download")
         if job.asset_id == asset_id:
             return job
         db.rollback()  # another request moved the job to another asset meanwhile: lock that one
@@ -240,7 +239,7 @@ def create_profile(
     name = body.name.strip()
     spec = body.spec
     if not name:
-        raise HTTPException(status_code=400, detail="A profile name is required.")
+        raise refused(400, "profile_name_required")
     maxpos = db.execute(
         select(func.coalesce(func.max(DownloadProfile.sort_order), 100)).where(
             DownloadProfile.user_id == user.id
@@ -268,16 +267,16 @@ def update_profile(
 ) -> dict:
     p = db.get(DownloadProfile, profile_id)
     if p is None or p.is_builtin or p.user_id != user.id:
-        raise HTTPException(status_code=404, detail="Unknown profile")
+        raise not_found("profile")
     sent = body.model_fields_set
     if "name" in sent:
         name = (body.name or "").strip()
         if not name:
-            raise HTTPException(status_code=400, detail="A profile name is required.")
+            raise refused(400, "profile_name_required")
         p.name = name[:80]
     if "spec" in sent:
         if body.spec is None:
-            raise HTTPException(status_code=400, detail="spec must be an object")
+            raise refused(400, "spec_not_object")
         p.spec = body.spec
     db.commit()
     return _serialize_profile(p)
@@ -289,7 +288,7 @@ def delete_profile(
 ) -> dict:
     p = db.get(DownloadProfile, profile_id)
     if p is None or p.is_builtin or p.user_id != user.id:
-        raise HTTPException(status_code=404, detail="Unknown profile")
+        raise not_found("profile")
     db.delete(p)
     db.commit()
     return {"deleted": profile_id}
@@ -319,11 +318,11 @@ def set_credential(
     """Store (encrypted) this user's cookies.txt for one supported host. Write-only: the value is
     never echoed back, logged, or audited — only the fact that it changed."""
     if not dlcreds.is_supported_host(host_key):
-        raise HTTPException(status_code=404, detail="Unknown host")
+        raise not_found("host")
     try:
         cleaned = dlcreds.validate_cookies(body.cookies)
     except dlcreds.InvalidCookies as e:
-        raise HTTPException(status_code=400, detail={"code": "cookies", "reason": str(e)})
+        raise HTTPException(status_code=400, detail={"code": "cookies", "reason": str(e)}) from e
     dlcreds.upsert_credential(db, user.id, host_key, cleaned)
     audit.record(
         db,
@@ -341,7 +340,7 @@ def delete_credential(
     host_key: str, user: User = Depends(require_human), db: Session = Depends(get_db)
 ) -> dict:
     if not dlcreds.is_supported_host(host_key):
-        raise HTTPException(status_code=404, detail="Unknown host")
+        raise not_found("host")
     if dlcreds.delete_credential(db, user.id, host_key):
         audit.record(
             db,
@@ -362,7 +361,7 @@ def _resolve_spec(
     if profile_id is not None:
         p = db.get(DownloadProfile, profile_id)
         if p is None or (not p.is_builtin and p.user_id != user.id):
-            raise HTTPException(status_code=404, detail="Unknown profile")
+            raise not_found("profile")
         return p.spec, p.id
     if spec is not None:
         return spec, None
@@ -373,7 +372,7 @@ def _resolve_spec(
         .limit(1)
     ).scalar_one_or_none()
     if fallback is None:
-        raise HTTPException(status_code=400, detail="No download profile available.")
+        raise refused(400, "no_profile")
     return fallback.spec, fallback.id
 
 
@@ -404,69 +403,127 @@ def enqueue_download(
                 status_code=422, detail={"code": "url_blocked", "reason": e.reason}
             ) from e
     spec, profile_id = _resolve_spec(db, user, body.profile_id, body.spec)
-    display_name = (body.display_name or "").strip() or None
+    display_name = display_text(body.display_name)
     try:
         job = service.enqueue(db, user.id, source_kind, source_ref, spec, profile_id, display_name)
     except quota.QuotaExceeded as e:
-        raise HTTPException(status_code=422, detail=_quota_detail(e))
+        raise HTTPException(status_code=422, detail=_quota_detail(e)) from e
     return _serialize_job(db, job)
 
 
-# Read the upload in 1 MiB chunks so a large file never sits fully in memory, and the size cap is
+# The upload's filename and display name travel as HEADERS, not as multipart parts. That is not a
-# enforced on the ACTUAL bytes read (a client-set Content-Length can't be trusted to gate it).
+# style choice: a DECLARED body parameter (`file: UploadFile = File(...)`) is read and spooled by
-_UPLOAD_CHUNK = 1024 * 1024
+# FastAPI *before the first dependency runs*, so `require_human` could not gate it and an anonymous
+# caller could make the process write the whole per-file allowance to temp, per request (P1; SEC-20
+# bounded the number but could not stop the writing). Reading the body ourselves, from a dependency
+# declared AFTER the auth one, is the only shape that closes it — the same one
+# `POST /api/debug-sessions/{sid}/screenshots` uses, and `test_body_limit.py` now enforces it for
+# every route with a raised ceiling rather than leaving it to a comment like this one.
+#
+# Headers are latin-1 on the wire, so both are percent-encoded by the client and unquoted here; a
+# Hungarian filename would otherwise be unsendable.
+UPLOAD_FILENAME_HEADER = "x-upload-filename"
+UPLOAD_DISPLAY_NAME_HEADER = "x-upload-display-name"
 
 
-@router.post("/upload")
+@dataclass(frozen=True)
-def upload_download(
+class SpooledUpload:
-    file: UploadFile = File(...),
+    """What the body dependency hands the handler: the bytes already on disk, plus the two names
-    display_name: str | None = Form(None),
+    that came alongside them."""
+
+    path: Path
+    size: int
+    user_id: int
+    filename: str
+    display_name: str | None
+
+
+def _header_text(request: Request, name: str) -> str:
+    """One percent-encoded header, decoded. A header that is not valid UTF-8 once unquoted is the
+    caller's doing, not a server fault — it degrades to empty rather than 500ing, and the route's
+    own fallback names the file."""
+    raw = request.headers.get(name, "")
+    try:
+        return unquote(raw, errors="strict").strip()
+    except (UnicodeDecodeError, ValueError):
+        # silent-ok: a mis-encoded header is the caller's own malformed request, and this one is
+        # cosmetic — it names the file. Refusing the upload over it would fail a 500 MiB transfer on
+        # a label; logging it would hand any client a free log line per request.
+        return ""
+
+
+async def spool_upload(
+    request: Request,
     user: User = Depends(require_human),
     db: Session = Depends(get_db),
-) -> dict:
+) -> AsyncIterator[SpooledUpload]:
-    """Ingest a local file into the caller's library (E7 S5). Any file type is accepted; media
+    """Stream the request body to a staging temp file, enforcing the per-file cap as it arrives.
-    (video/audio) gets the full function set, a generic file is download-only. The file is streamed
-    to a temp path under the download volume (enforcing the per-file cap as it goes), then handed to
-    the non-yt-dlp ingest which probes + places it and returns an already-`done` job.
 
-    A sync (threadpool) handler on purpose: the ingest does blocking work — an ffprobe of up to 30s
+    `user` is taken but unused: it is the whole point. Dependencies resolve in declaration order, so
-    plus sync DB writes — so it must NOT run on the event loop. `file.file` is the sync file object
+    naming `require_human` here makes the auth check a precondition of reading the body at all — a
-    Starlette spools the multipart body into."""
+    stranger is refused before a byte moves. The `yield` shape owns the temp file's life: ingest
+    MOVES it into place on success (the unlink is then a no-op), and on any other exit — cap hit,
+    empty, quota, probe or move error, a client that vanished mid-upload — it is removed."""
     cap = sysconfig.get_int(db, "download_upload_max_bytes")
+    user_id = user.id
+    # Hand the pooled connection back BEFORE the transfer starts — see `db.release_read_connection`
+    # for why, and for the precondition it enforces. Everything this dependency still needs is
+    # already an int or a header in hand: touching an ORM attribute below would reload `user` and
+    # check a connection straight back out.
+    release_read_connection(db)
     staging = Path(settings.download_root) / ".staging" / "uploads"
     staging.mkdir(parents=True, exist_ok=True)
     fd, tmp = tempfile.mkstemp(dir=str(staging))
     tmp_path = Path(tmp)
     try:
-        size = 0
         with os.fdopen(fd, "wb") as out:
-            while chunk := file.file.read(_UPLOAD_CHUNK):
+            size = await capped_stream_to(
-                size += len(chunk)
+                request,
-                if cap and size > cap:
+                out,
-                    raise HTTPException(
+                cap,
-                        status_code=422, detail={"code": "upload_too_large", "limit": cap}
+                HTTPException(status_code=422, detail={"code": "upload_too_large", "limit": cap}),
-                    )
-                out.write(chunk)
-        if size == 0:
-            raise HTTPException(status_code=400, detail={"code": "upload_empty"})
-        try:
-            job = uploadmod.ingest_upload(
-                db,
-                user.id,
-                filename=file.filename or "upload",
-                src_path=tmp_path,
-                size=size,
-                display_name=(display_name or "").strip() or None,
             )
-        except quota.QuotaExceeded as e:
+        yield SpooledUpload(
-            raise HTTPException(status_code=422, detail=_quota_detail(e))
+            path=tmp_path,
-        return _serialize_job(db, job)
+            size=size,
+            user_id=user_id,
+            filename=_header_text(request, UPLOAD_FILENAME_HEADER) or "upload",
+            display_name=display_text(_header_text(request, UPLOAD_DISPLAY_NAME_HEADER)),
+        )
     finally:
-        # Ingest MOVES the temp file into place on success (unlink is then a no-op); on any early exit
-        # — cap hit, empty, quota, probe/move error — this removes the abandoned staging file.
         tmp_path.unlink(missing_ok=True)
 
 
+@router.post("/upload")
+def upload_download(
+    upload: SpooledUpload = Depends(spool_upload),
+    db: Session = Depends(get_db),
+) -> dict:
+    """Ingest a local file into the caller's library (E7 S5). Any file type is accepted; media
+    (video/audio) gets the full function set, a generic file is download-only. The body IS the file
+    (see the header note above); `spool_upload` has already authenticated the caller and streamed it
+    to a temp path under the download volume, so what is left here is the non-yt-dlp ingest which
+    probes + places it and returns an already-`done` job.
+
+    A sync (threadpool) handler on purpose: the ingest does blocking work — an ffprobe of up to 30s
+    plus sync DB writes — so it must NOT run on the event loop. The streaming half is async because
+    reading the body is; keeping them in separate callables is what lets each run where it belongs."""
+    if upload.size == 0:
+        raise HTTPException(status_code=400, detail={"code": "upload_empty"})
+    try:
+        job = uploadmod.ingest_upload(
+            db,
+            upload.user_id,
+            filename=upload.filename,
+            src_path=upload.path,
+            size=upload.size,
+            display_name=upload.display_name,
+        )
+    except quota.QuotaExceeded as e:
+        raise HTTPException(status_code=422, detail=_quota_detail(e)) from e
+    return _serialize_job(db, job)
+
+
 class EnqueueEditIn(BaseModel):
     source_job_id: int = 0
     edit_spec: dict
@@ -484,15 +541,15 @@ def enqueue_edit(
     source file is only read, never modified, so editing a shared video is safe."""
     source_job = _accessible_job(db, user, body.source_job_id)
     if source_job.status != "done" or source_job.asset_id is None:
-        raise HTTPException(status_code=409, detail="You can only edit a finished download.")
+        raise refused(409, "edit_unfinished")
     spec = body.edit_spec
-    display_name = (body.display_name or "").strip() or None
+    display_name = display_text(body.display_name)
     try:
         job = service.enqueue_edit(db, user.id, source_job, spec, display_name)
     except quota.QuotaExceeded as e:
-        raise HTTPException(status_code=422, detail=_quota_detail(e))
+        raise HTTPException(status_code=422, detail=_quota_detail(e)) from e
     except service.EditError as e:
-        raise HTTPException(status_code=400, detail={"code": "edit", "reason": e.reason})
+        raise HTTPException(status_code=400, detail={"code": "edit", "reason": e.reason}) from e
     return _serialize_job(db, job)
 
 
@@ -522,7 +579,20 @@ def list_downloads(
 
 @router.get("/usage")
 def my_usage(user: User = Depends(require_human), db: Session = Depends(get_db)) -> dict:
-    return quota.usage(db, user.id)
+    """The caller's download quota, plus the per-file upload cap.
+
+    The cap rides here rather than in `quota.usage` because it is not a quota value — it is a
+    `sysconfig` knob about one route — and this is the route layer, where composing the two is the
+    job. The upload dialog needs it to refuse an over-cap file BEFORE sending it: both server-side
+    caps answer too late to be pleasant (the edge ceiling refuses the request outright, see
+    `bodylimit.ROUTE_MAX_BODY_BYTES`), and making the user push 600 MB up the wire to be told the
+    limit is 500 is a bad answer even when it is a correct one. It is a convenience, not the
+    boundary: the caps stay where they are and still refuse a client that ignores this.
+    """
+    return {
+        **quota.usage(db, user.id),
+        "upload_max_bytes": sysconfig.get_int(db, "download_upload_max_bytes"),
+    }
 
 
 # Priority when a source has several jobs (e.g. a done one plus a fresh queued one).
@@ -587,6 +657,8 @@ def _clean_url(raw) -> str | None:
     # authority. This one already has a "not a plausible URL" answer, so the typo belongs in it.
     try:
         parsed = urlparse(url)
+    # silent-ok: a typo in a link someone pasted. None is this function's "not a plausible URL"
+    # answer, the caller already has one, and any line here would carry the URL itself.
     except ValueError:
         return None
     if not parsed.netloc:
@@ -612,12 +684,12 @@ def update_download_meta(
 ) -> dict:
     """Update a download's display metadata: title, channel name + link, and extra reference URLs.
     All are display-only overrides — the on-disk file keeps its system-decided, id-bound name."""
-    job = _own_job(db, user, job_id)
+    job = owned_or_404(db, DownloadJob, job_id, user, what="download")
     sent = body.model_fields_set
     if "display_name" in sent:
-        job.display_name = ((body.display_name or "").strip())[:255] or None
+        job.display_name = display_text(body.display_name)
     if "display_uploader" in sent:
-        job.display_uploader = ((body.display_uploader or "").strip())[:255] or None
+        job.display_uploader = display_text(body.display_uploader)
     if "display_uploader_url" in sent:
         job.display_uploader_url = _clean_url(body.display_uploader_url)
     if "extra_links" in sent:
@@ -666,9 +738,9 @@ def redownload_download(
     if job.job_kind in ("edit", "upload"):
         # Neither has a re-fetchable source: an editor clip is derived locally, an upload IS the
         # source. Re-running yt-dlp against source_kind='upload'/'edit' would just fail.
-        raise HTTPException(status_code=409, detail="This item can't be re-downloaded.")
+        raise refused(409, "not_redownloadable")
     if job.status in ("queued", "running", "paused"):
-        raise HTTPException(status_code=409, detail="This download is still in progress.")
+        raise refused(409, "still_downloading")
     service.redownload(db, job)
     return _serialize_job(db, job)
 
@@ -697,7 +769,7 @@ def delete_download(
 def _accessible_job(db: Session, user: User, job_id: int) -> DownloadJob:
     job = db.get(DownloadJob, job_id)
     if job is None:
-        raise HTTPException(status_code=404, detail="Unknown download")
+        raise not_found("download")
     if job.user_id == user.id:
         return job
     shared = db.execute(
@@ -706,7 +778,7 @@ def _accessible_job(db: Session, user: User, job_id: int) -> DownloadJob:
         )
     ).first()
     if shared is None:
-        raise HTTPException(status_code=404, detail="Unknown download")
+        raise not_found("download")
     return job
 
 
@@ -720,10 +792,10 @@ def download_file(
     job = _accessible_job(db, user, job_id)
     asset = db.get(MediaAsset, job.asset_id) if job.asset_id else None
     if asset is None or asset.status != "ready" or not asset.rel_path:
-        raise HTTPException(status_code=409, detail="This download isn't ready.")
+        raise refused(409, "download_not_ready")
     path = storage.safe_abs_path(settings.download_root, asset.rel_path)
     if path is None:
-        raise HTTPException(status_code=404, detail="File is no longer available.")
+        raise not_found("file")
 
     filename = storage.download_filename(
         job.display_name or asset.title or asset.source_ref, asset.container, path
@@ -762,7 +834,7 @@ def storyboard_meta(
     job = _accessible_job(db, user, job_id)
     asset = db.get(MediaAsset, job.asset_id) if job.asset_id else None
     if asset is None or asset.status != "ready" or not asset.rel_path:
-        raise HTTPException(status_code=409, detail="This download isn't ready.")
+        raise refused(409, "download_not_ready")
     meta = editmod.ensure_storyboard(asset, settings.download_root)
     if meta is None:
         return {"available": False}
@@ -783,10 +855,10 @@ def poster_image(job_id: int, user: User = Depends(require_human), db: Session =
     job = _accessible_job(db, user, job_id)
     asset = db.get(MediaAsset, job.asset_id) if job.asset_id else None
     if asset is None or not asset.poster_path:
-        raise HTTPException(status_code=404, detail="No poster.")
+        raise not_found("poster")
     path = storage.safe_abs_path(settings.download_root, asset.poster_path)
     if path is None:
-        raise HTTPException(status_code=404, detail="No poster.")
+        raise not_found("poster")
     return FileResponse(path, media_type="image/jpeg")
 
 
@@ -797,10 +869,10 @@ def storyboard_image(
     job = _accessible_job(db, user, job_id)
     asset = db.get(MediaAsset, job.asset_id) if job.asset_id else None
     if asset is None or not asset.storyboard_path:
-        raise HTTPException(status_code=404, detail="No filmstrip.")
+        raise not_found("filmstrip")
     path = storage.safe_abs_path(settings.download_root, asset.storyboard_path)
     if path is None:
-        raise HTTPException(status_code=404, detail="No filmstrip.")
+        raise not_found("filmstrip")
     return FileResponse(path, media_type="image/jpeg")
 
 
@@ -818,12 +890,12 @@ def share_download(
     user: User = Depends(require_human),
     db: Session = Depends(get_db),
 ) -> dict:
-    job = _own_job(db, user, job_id)
+    job = owned_or_404(db, DownloadJob, job_id, user, what="download")
     if job.status != "done" or job.asset_id is None:
-        raise HTTPException(status_code=409, detail="Only a finished download can be shared.")
+        raise refused(409, "share_unfinished")
     target = body.email.strip().lower()
     if not target:
-        raise HTTPException(status_code=400, detail="A recipient email is required.")
+        raise refused(400, "recipient_required")
     recipient = db.execute(
         select(User).where(func.lower(User.email) == target)
     ).scalar_one_or_none()
@@ -832,7 +904,7 @@ def share_download(
     if not is_messageable_user(recipient):
         raise not_found("recipient")
     if recipient.id == user.id:
-        raise HTTPException(status_code=400, detail="That's already your download.")
+        raise refused(400, "share_own")
     exists = db.execute(
         select(DownloadShare.id).where(
             DownloadShare.job_id == job_id, DownloadShare.to_user_id == recipient.id
@@ -847,11 +919,18 @@ def share_download(
             user_id=recipient.id,
             type="download_shared",
             title=(job.display_name or "A download") + " was shared with you",
-            data={"kind": "download_shared", "from": user.email, "job_id": job_id},
+            # The client names the notice in the reader's language from `data` (E30); the title is the
+            # English fallback only.
+            data={
+                "kind": "download_shared",
+                "from": user.email,
+                "job_id": job_id,
+                "title": job.display_name,
+            },
             commit=False,
         )
         db.commit()
-    return {"shared_with": recipient.email}
+    return {"ok": True, "shared_with": recipient.email}
 
 
 @router.delete("/{job_id}/share/{email}")
@@ -861,7 +940,9 @@ def unshare_download(
     user: User = Depends(require_human),
     db: Session = Depends(get_db),
 ) -> dict:
-    _own_job(db, user, job_id)  # ownership guard (404s if not the caller's job)
+    owned_or_404(
+        db, DownloadJob, job_id, user, what="download"
+    )  # ownership guard (404s if not the caller's job)
     recipient = db.execute(
         select(User).where(func.lower(User.email) == email.strip().lower())
     ).scalar_one_or_none()
@@ -893,7 +974,7 @@ def remove_shared_with_me(
     if row is not None:
         db.delete(row)
         db.commit()
-    return {"removed": job_id}
+    return {"ok": True, "removed": job_id}
 
 
 # --- recipients (for the internal "share with a user" picker) ------------------------------
@@ -909,7 +990,7 @@ def share_recipients(
         db.execute(
             select(User)
             .where(User.id != user.id, *messageable_clauses())
-            .order_by(func.lower(User.email))
+            .order_by(collated(User.email, lang_for_user(user)))
         )
         .scalars()
         .all()
@@ -923,10 +1004,10 @@ def share_recipients(
 def _own_link(db: Session, user: User, link_id: int) -> DownloadLink:
     link = db.get(DownloadLink, link_id)
     if link is None:
-        raise HTTPException(status_code=404, detail="Unknown link")
+        raise not_found("link")
     job = db.get(DownloadJob, link.job_id)
     if job is None or job.user_id != user.id:
-        raise HTTPException(status_code=404, detail="Unknown link")
+        raise not_found("link")
     return link
 
 
@@ -942,7 +1023,7 @@ def _link_expiry(days: int | None) -> datetime | None:
 def list_links(
     job_id: int, user: User = Depends(require_human), db: Session = Depends(get_db)
 ) -> list[dict]:
-    _own_job(db, user, job_id)
+    owned_or_404(db, DownloadJob, job_id, user, what="download")
     rows = (
         db.execute(
             select(DownloadLink)
@@ -973,9 +1054,9 @@ def create_link(
     user: User = Depends(require_human),
     db: Session = Depends(get_db),
 ) -> dict:
-    job = _own_job(db, user, job_id)
+    job = owned_or_404(db, DownloadJob, job_id, user, what="download")
     if job.status != "done" or job.asset_id is None:
-        raise HTTPException(status_code=409, detail="Only a finished download can be shared.")
+        raise refused(409, "share_unfinished")
     pw = body.password.strip()
     link = DownloadLink(
         job_id=job_id,

backend/app/routes/feed.py

@@ -1,6 +1,6 @@
 from datetime import date, datetime, timezone
 
-from fastapi import APIRouter, Depends, HTTPException, Query
+from fastapi import APIRouter, Depends, Query
 from pydantic import BaseModel, ConfigDict, Field
 from sqlalchemy import and_, func, or_, select
 from sqlalchemy.dialects.postgresql import insert as pg_insert
@@ -10,6 +10,7 @@ from sqlalchemy.orm.exc import StaleDataError
 from app import quota
 from app.auth import current_user
 from app.db import get_db
+from app.mailtext import lang_for_user
 from app.feedquery import (
     decode_cursor,
     encode_cursor,
@@ -18,8 +19,11 @@ from app.feedquery import (
     filtered_query,
     keyset_predicate,
     order_by,
+    rank_window_start,
     serialize,
     sort_keys,
+    windowed_rank,
+    windowed_relevance_rows,
 )
 from app.models import (
     POSITION_MAX_S,
@@ -29,10 +33,11 @@ from app.models import (
     Video,
     VideoState,
 )
-from app.routes._common import get_or_404
+from app.routes._common import get_or_404, youtube_refusal, not_found
 from app.sync.videos import parse_iso8601_duration
 from app.utils import iso
 from app.youtube.client import YouTubeClient, YouTubeError
+from app.refusal import refused
 
 router = APIRouter(prefix="/api", tags=["feed"])
 
@@ -107,30 +112,45 @@ def get_feed(
     db: Session = Depends(get_db),
 ) -> dict:
     query, rank_expr = filtered_query(db, user, **params)
-    keys = sort_keys(sort, seed, rank_expr)
+    lang = lang_for_user(user)
-
+    keys = sort_keys(sort, seed, rank_expr, lang)
-    # Expose each key column on the row so we can build the next cursor from the last item.
-    query = query.add_columns(*[k["expr"].label(f"_k{i}") for i, k in enumerate(keys)])
-    query = query.order_by(*order_by(keys))
 
+    vals, cursor_id, window = None, None, None
     if cursor:
         # The read-model raises a plain ValueError on a malformed cursor; translate it to the
         # HTTP 400 here so feedquery stays web-framework-free.
         try:
-            vals, cursor_id = decode_cursor(keys, cursor)
+            vals, cursor_id, window = decode_cursor(keys, cursor)
         except ValueError:
             # `from None`: the ValueError is an expected bad-input signal, not an error to chain.
-            raise HTTPException(status_code=400, detail="Invalid feed cursor") from None
+            raise refused(400, "invalid_cursor") from None
-        for k, v in zip(keys, vals):
+    ranked = sort == "relevance" and rank_expr is not None
-            k["value"] = v
+    if ranked and not cursor:
-        query = query.where(keyset_predicate(keys, cursor_id))
+        window = rank_window_start(db, query)
+    if ranked and window is not None:
+        keys = sort_keys(sort, seed, windowed_rank(rank_expr, window), lang)
 
-    rows = db.execute(query.limit(limit + 1)).all()
+    # Expose each key column on the row so we can build the next cursor from the last item.
+    query = query.add_columns(*[k["expr"].label(f"_k{i}") for i, k in enumerate(keys)])
+    if vals is not None:
+        for k, v in zip(keys, vals, strict=True):
+            k["value"] = v
+    if ranked and window is not None:
+        rows = windowed_relevance_rows(db, query, keys, window, cursor_id, limit + 1)
+    else:
+        query = query.order_by(*order_by(keys))
+        if cursor_id is not None:
+            query = query.where(keyset_predicate(keys, cursor_id))
+        rows = db.execute(query.limit(limit + 1)).all()
     has_more = len(rows) > limit
     page = rows[:limit]
     return {
         "items": [serialize(r) for r in page],
-        "next_cursor": encode_cursor(keys, page[-1]) if (has_more and page) else None,
+        "next_cursor": (
+            encode_cursor(keys, page[-1], window if ranked else None)
+            if (has_more and page)
+            else None
+        ),
         "has_more": has_more,
         "limit": limit,
     }
@@ -235,7 +255,7 @@ def set_video_state(
 ) -> dict:
     status = body.status
     if status not in VALID_STATES:
-        raise HTTPException(status_code=400, detail=f"status must be one of {VALID_STATES}")
+        raise refused(400, "invalid_status")
     get_or_404(db, Video, video_id, what="video")
 
     now = datetime.now(timezone.utc)
@@ -256,9 +276,11 @@ def set_video_state(
                 db.delete(row)
             try:
                 db.commit()
+            # silent-ok: the documented lost race — a concurrent player checkpoint already removed
+            # or moved this row. The state the caller asked for is the state that holds either way.
             except StaleDataError:
                 db.rollback()
-        return {"video_id": video_id, "status": "new"}
+        return {"ok": True, "video_id": video_id, "status": "new"}
 
     # watched / hidden: atomic upsert. The player marks "watched" at end-of-playback at
     # the same instant a progress checkpoint may DELETE the in-progress row; a plain
@@ -279,7 +301,7 @@ def set_video_state(
     )
     db.execute(stmt)
     db.commit()
-    return {"video_id": video_id, "status": status}
+    return {"ok": True, "video_id": video_id, "status": status}
 
 
 @router.delete("/videos/{video_id}/state")
@@ -298,9 +320,10 @@ def clear_video_state(
         db.delete(row)
         try:
             db.commit()
+        # silent-ok: the row went under the delete, which is the outcome the delete wanted.
         except StaleDataError:
             db.rollback()
-    return {"video_id": video_id, "status": "new", "position_seconds": 0}
+    return {"ok": True, "video_id": video_id, "status": "new", "position_seconds": 0}
 
 
 class VideoProgressIn(BaseModel):
@@ -355,9 +378,10 @@ def set_video_progress(
             # have removed/updated this row already — tolerate the lost race.
             try:
                 db.commit()
+            # silent-ok: exactly the race the comment above describes; the response stays true.
             except StaleDataError:
                 db.rollback()
-        return {"video_id": video_id, "position_seconds": 0}
+        return {"ok": True, "video_id": video_id, "position_seconds": 0}
 
     if row is None:
         row = VideoState(user_id=user.id, video_id=video_id)
@@ -365,7 +389,7 @@ def set_video_progress(
     row.position_seconds = position
     row.progress_updated_at = datetime.now(timezone.utc)
     db.commit()
-    return {"video_id": video_id, "position_seconds": position}
+    return {"ok": True, "video_id": video_id, "position_seconds": position}
 
 
 @router.get("/videos/{video_id}")
@@ -398,7 +422,7 @@ def get_video_detail(
     # The DB path above is free; the off-catalog YouTube lookup below spends a quota unit, so the
     # demo account can't reach it — it just sees "unknown" for a video we don't already store.
     if user.is_demo:
-        raise HTTPException(status_code=404, detail="Unknown video")
+        raise not_found("video")
 
     try:
         with (
@@ -407,9 +431,9 @@ def get_video_detail(
         ):
             items = yt.get_videos([video_id])
     except YouTubeError as exc:
-        raise HTTPException(status_code=422, detail=f"YouTube lookup failed: {exc}")
+        raise youtube_refusal(exc) from exc
     if not items:
-        raise HTTPException(status_code=404, detail="Unknown video")
+        raise not_found("video")
 
     snippet = items[0].get("snippet", {})
     stats = items[0].get("statistics", {})

backend/app/routes/me.py

@@ -1,5 +1,7 @@
+from typing import Any
+
 from fastapi import APIRouter, BackgroundTasks, Depends, HTTPException, Request
-from pydantic import BaseModel
+from pydantic import BaseModel, RootModel
 from sqlalchemy import func, select
 from sqlalchemy.orm import Session
 
@@ -15,10 +17,16 @@ from app.auth import (
     optional_current_user,
     purge_user,
 )
+from app import debugsession, diagnostics
+from app.issues import gitea
 from app import email as email_mod
+from app import logs
 from app import sysconfig
 from app.db import get_db
 from app.models import Invite, User
+from app.utils import iso
+from app.refusal import refused
+from app.routes._common import not_found
 
 router = APIRouter(prefix="/api/me", tags=["me"])
 
@@ -65,7 +73,7 @@ def switch_account(
     here) are allowed, and access is re-checked in case the invite was revoked since."""
     target = body.user_id
     if target not in (request.session.get("account_ids") or []):
-        raise HTTPException(status_code=403, detail="Not an account you've signed into here.")
+        raise refused(403, "not_your_account")
     # An account `_session_user` would refuse — gone, deactivated, suspended, or its session revoked — is
     # refused here too, and dropped from the switch list the way `_session_user` drops it. Made the session
     # default instead, the next request's current_user would refuse it as the DEFAULT and clear the whole
@@ -73,22 +81,20 @@ def switch_account(
     u = db.get(User, target)
     if u is None:
         drop_from_wallet(request.session, target)
-        raise HTTPException(status_code=404, detail="That account no longer exists.")
+        raise not_found("account")
     if not is_allowed(db, u.email):
-        raise HTTPException(status_code=403, detail="That account no longer has access.")
+        raise refused(403, "account_no_access")
     if not u.is_active:
         drop_from_wallet(request.session, target)
-        raise HTTPException(status_code=403, detail="That account no longer has access.")
+        raise refused(403, "account_no_access")
     if u.is_suspended:
         drop_from_wallet(request.session, target)
-        raise HTTPException(status_code=403, detail="That account is suspended.")
+        raise refused(403, "account_suspended")
     # Revoked: a password reset/change or "log out everywhere" bumped the account's epoch.
     epochs = request.session.get("epochs") or {}
     if int(epochs.get(str(target), 0)) != (u.session_epoch or 0):
         drop_from_wallet(request.session, target)
-        raise HTTPException(
+        raise refused(403, "account_signed_out")
-            status_code=403, detail="That account was signed out elsewhere — sign in again."
-        )
     request.session["user_id"] = target
     return {"ok": True, "user_id": target}
 
@@ -134,11 +140,17 @@ def get_me(
             user.google_email if user.google_email and user.google_email != user.email else None
         ),
         "has_password": user.password_hash is not None,
+        # The account's latest sign-in on any device — shown in Settings → Account, so a sign-in the
+        # user does not recognise is something they can see (null: none since 0.88, or the demo).
+        "last_login_at": iso(user.last_login_at),
         # Instance-wide: whether Google sign-in / YouTube connect is available at all (the UI hides
         # YouTube-access affordances and the onboarding nudge when it isn't configured).
         "google_enabled": google_enabled(db),
         # Whether the optional Plex module is enabled (UI shows the Plex feed source when so).
         "plex_enabled": sysconfig.get_bool(db, "plex_enabled"),
+        # Which live-search source runs, so the search button can say what a search costs: "scrape"
+        # spends no API quota, "api" spends 100 units a page.
+        "search_source": sysconfig.get_str(db, "search_source"),
         "can_read": has_read_scope(user),
         "can_write": has_write_scope(user),
         "pending_invites": pending_invites,
@@ -147,9 +159,21 @@ def get_me(
         "pending_activations": pending_activations,
         "email_configured": email_configured,
         "preferences": user.preferences or {},
+        # The consented diagnostics level in force and when it lapses (E22 S5). The client applies
+        # it to its own log and streams that log while it is raised.
+        "diagnostics": diagnostics.state(db, user),
+        # The live debug session, if one is recording (E16) — a reload resumes it.
+        "debug_session": _live_session(db, user),
+        # Where a bug report goes (E16 S4): the tracker's repository, or None — download only.
+        "issues": gitea.target(db),
     }
 
 
+def _live_session(db: Session, user: User) -> dict | None:
+    session = debugsession.open_session(db, user)
+    return None if session is None else debugsession.view(db, session)
+
+
 @router.delete("/account")
 def delete_my_account(
     request: Request,
@@ -170,10 +194,7 @@ def delete_my_account(
     if user.is_demo:
         raise HTTPException(status_code=403, detail={"code": "demo", "reason": "delete_account"})
     if user.role == "admin" and count_admins(db) <= 1:
-        raise HTTPException(
+        raise refused(400, "only_admin_delete")
-            status_code=400,
-            detail="You're the only admin — promote another admin before deleting your account.",
-        )
     user_id = user.id
     # Full erasure (cascades) + access-request cleanup + Google-grant revocation; shared with the
     # admin delete path. Capture the id first — `user` is detached once purge_user deletes the row.
@@ -188,16 +209,45 @@ def delete_my_account(
     return {"deleted": True}
 
 
+class PreferencesIn(RootModel[dict[str, Any]]):
+    """Free-form UI preferences — a JSON object of any keys (theme, filters, …), merged into the
+    stored ones. A model rather than a bare `dict` parameter so a non-object body is a 422 at the
+    boundary like every other request body."""
+
+
 @router.put("/preferences")
 def update_preferences(
-    preferences: dict,
+    preferences: PreferencesIn,
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
     # Merge so partial updates don't wipe other keys.
     merged = dict(user.preferences or {})
-    merged.update(preferences)
+    merged.update(preferences.root)
     user.preferences = merged
     db.add(user)
     db.commit()
-    return {"preferences": merged}
+    return {"ok": True, "preferences": merged}
+
+
+class DiagnosticsIn(BaseModel):
+    level: diagnostics.DiagnosticsLevel
+
+
+@router.put("/diagnostics")
+def update_diagnostics(
+    body: DiagnosticsIn,
+    user: User = Depends(current_user),
+    db: Session = Depends(get_db),
+) -> dict:
+    """Raise the diagnostics level (the user consented in Settings → Diagnosztika) or return to the
+    base. The expiry is the server's: a raised level always lapses after
+    `diagnostics.ELEVATION_HOURS`, whatever the client asks. The shared demo account cannot raise
+    it — its log would interleave every visitor's session under one account."""
+    if user.is_demo and body.level != diagnostics.BASE:
+        raise HTTPException(status_code=403, detail={"code": "demo", "reason": "diagnostics"})
+    diagnostics.set_level(user, body.level)
+    db.commit()
+    # The rest of THIS request, and the process's elevation registry, follow the new level now.
+    logs.set_user(user.id, diagnostics.effective_level(user), user.diagnostics_until)
+    return {"ok": True, "diagnostics": diagnostics.state(db, user)}

backend/app/routes/messages.py

@@ -16,11 +16,13 @@ routing and unread counts.
 
 from datetime import datetime, timezone
 
-from fastapi import APIRouter, Depends, HTTPException, WebSocket, WebSocketDisconnect
+from fastapi import APIRouter, Depends, WebSocket, WebSocketDisconnect
 from pydantic import BaseModel
 from sqlalchemy import and_, case, delete, func, or_, select
 from sqlalchemy.orm import Session
 
+from app.mailtext import lang_for_user
+from app.collation import collated
 from app.auth import require_human, resolved_user_id
 from app.db import SessionLocal, get_db
 from app.models import Message, MessageKey, User
@@ -29,6 +31,8 @@ from app.realtime import manager
 from app.security import verify_password
 from app.userscope import is_messageable_user, messageable_clauses
 from app.utils import iso
+from app.refusal import refused
+from app.routes._common import not_found
 
 router = APIRouter(prefix="/api/messages", tags=["messages"])
 
@@ -90,13 +94,31 @@ def _system_partner() -> dict:
     return {"id": SYSTEM_PARTNER_ID, "name": SYSTEM_NAME, "avatar_url": None}
 
 
-def _serialize_msg(m: Message) -> dict:
+# The welcome's opening line in every language — how a stored copy is recognised. It was rendered
+# once, in the language of the day it was created, so an account that later switched language kept
+# reading it in the old one (E31 S1: the owner's, in English). Recognised here, it is served in the
+# reader's language instead. Derived from WELCOME itself, so a change to an OPENING line, or a
+# language taken out of WELCOME, orphans the copies already stored: they are served as stored. The
+# one retired language (German, 2026-07) left none on prod (checked 2026-10-02).
+_WELCOME_OPENINGS = {text.partition("\n")[0] for text in WELCOME.values()}
+
+
+def _system_body(body: str | None, lang: str | None) -> str | None:
+    """A system message's text for a reader of `lang`: the welcome in their language, any other as
+    stored."""
+    if body is None or lang is None or body.partition("\n")[0] not in _WELCOME_OPENINGS:
+        return body
+    return WELCOME.get(lang, WELCOME["en"])
+
+
+def _serialize_msg(m: Message, lang: str | None = None) -> dict:
+    """`lang`: the reader's language, for a system message's text (a user message is ciphertext)."""
     return {
         "id": m.id,
         "kind": m.kind,
         "sender_id": m.sender_id,
         "recipient_id": m.recipient_id,
-        "body": m.body,  # only set for system messages
+        "body": _system_body(m.body, lang) if m.kind == "system" else m.body,
         "ciphertext": m.ciphertext,
         "iv": m.iv,
         "read_at": iso(m.read_at),
@@ -120,7 +142,7 @@ def ensure_welcome(db: Session, user: User) -> None:
     db.add(m)
     db.commit()
     db.refresh(m)
-    manager.push(user.id, {"type": "message", "message": _serialize_msg(m)})
+    manager.push(user.id, {"type": "message", "message": _serialize_msg(m, lang_for_user(user))})
 
 
 # --- E2EE key directory / blob store -----------------------------------------
@@ -152,7 +174,7 @@ def setup_keys(
     """Store this user's public key + passphrase-wrapped private key. Set-once: changing the
     key would orphan every existing conversation key, so re-setup is rejected here."""
     if db.get(MessageKey, user.id) is not None:
-        raise HTTPException(status_code=409, detail="Secure messaging is already set up.")
+        raise refused(409, "messaging_set_up")
     db.add(
         MessageKey(
             user_id=user.id,
@@ -164,7 +186,7 @@ def setup_keys(
         )
     )
     db.commit()
-    return {"configured": True}
+    return {"ok": True, "configured": True}
 
 
 @router.post("/keys/reset")
@@ -185,11 +207,11 @@ def reset_keys(
     less (Google-only) account has nothing to re-verify, so it relies on its live session plus the
     client-side danger confirmation."""
     if not _reset_key_limiter.allow(str(user.id)):
-        raise HTTPException(status_code=429, detail="Too many attempts. Try again shortly.")
+        raise refused(429, "too_many_attempts")
     if user.password_hash and not verify_password(
         payload.current_password or "", user.password_hash
     ):
-        raise HTTPException(status_code=403, detail="Current password is incorrect.")
+        raise refused(403, "wrong_password")
     # Delete the now-dead ciphertext first (both directions), then the wrapped keypair itself so a
     # fresh setup is allowed. Order doesn't matter (no FK between them), but do it in one commit.
     db.execute(
@@ -202,7 +224,7 @@ def reset_keys(
     if key is not None:
         db.delete(key)
     db.commit()
-    return {"configured": False}
+    return {"ok": True, "configured": False}
 
 
 @router.get("/keys/{user_id}")
@@ -223,10 +245,10 @@ def public_key(
     p = db.get(User, user_id)
     has_history = db.scalar(select(Message.id).where(cond).limit(1))
     if p is None or (not is_messageable_user(p) and not has_history):
-        raise HTTPException(status_code=404, detail="User not found")
+        raise not_found("user")
     k = db.get(MessageKey, user_id)
     if k is None:
-        raise HTTPException(status_code=404, detail="That member hasn't set up messaging yet.")
+        raise refused(404, "no_messaging")
     return {"user_id": user_id, "public_key": k.public_key}
 
 
@@ -244,7 +266,7 @@ def list_recipients(
             select(User)
             .join(MessageKey, MessageKey.user_id == User.id)
             .where(User.id != user.id, *messageable_clauses())
-            .order_by(func.lower(func.coalesce(User.display_name, User.email)))
+            .order_by(collated(func.coalesce(User.display_name, User.email), lang_for_user(user)))
         )
         .scalars()
         .all()
@@ -309,7 +331,7 @@ def list_conversations(user: User = Depends(require_human), db: Session = Depend
         items.append(
             {
                 "partner": partner,
-                "last_message": _serialize_msg(m),
+                "last_message": _serialize_msg(m, lang_for_user(user)),
                 "unread": int(unread.get(pid, 0)),
             }
         )
@@ -344,7 +366,7 @@ def get_thread(
         # still opens). Otherwise the same 404 as a non-existent id → no enumeration oracle.
         has_history = db.scalar(select(Message.id).where(cond).limit(1))
         if p is None or (not is_messageable_user(p) and not has_history):
-            raise HTTPException(status_code=404, detail="User not found")
+            raise not_found("user")
         partner = _serialize_user(p)
     msgs = (
         db.execute(select(Message).where(cond).order_by(Message.created_at.asc())).scalars().all()
@@ -362,7 +384,7 @@ def get_thread(
             # refreshes live instead of waiting for the next 20s poll. The client re-fetches the
             # count on this signal, so no number is carried here.
             manager.push(user.id, {"type": "unread"})
-    return {"partner": partner, "items": [_serialize_msg(m) for m in msgs]}
+    return {"partner": partner, "items": [_serialize_msg(m, lang_for_user(user)) for m in msgs]}
 
 
 @router.post("")
@@ -374,20 +396,20 @@ def send_message(
     """Send an end-to-end-encrypted message. The server stores only the ciphertext the client
     produced, then pushes it live to the recipient (and the sender's other devices)."""
     if not _send_limiter.allow(f"msg:{user.id}"):
-        raise HTTPException(status_code=429, detail="Too many messages — slow down a moment.")
+        raise refused(429, "message_rate")
     ct = (payload.ciphertext or "").strip()
     iv = (payload.iv or "").strip()
     if not ct or not iv:
-        raise HTTPException(status_code=400, detail="Message is empty.")
+        raise refused(400, "message_empty")
     if len(ct) > MAX_CIPHERTEXT:
-        raise HTTPException(status_code=400, detail="Message is too long.")
+        raise refused(400, "message_too_long")
     if payload.recipient_id == user.id:
-        raise HTTPException(status_code=400, detail="You can't message yourself.")
+        raise refused(400, "message_self")
     recipient = db.get(User, payload.recipient_id)
     if not is_messageable_user(recipient):
-        raise HTTPException(status_code=404, detail="Recipient not available.")
+        raise not_found("recipient")
     if db.get(MessageKey, recipient.id) is None:
-        raise HTTPException(status_code=404, detail="That member hasn't set up messaging yet.")
+        raise refused(404, "no_messaging")
     m = Message(kind="user", sender_id=user.id, recipient_id=recipient.id, ciphertext=ct, iv=iv)
     db.add(m)
     db.commit()
@@ -449,6 +471,8 @@ async def messages_ws(ws: WebSocket) -> None:
     try:
         while True:
             await ws.receive_text()
+    # silent-ok: the ordinary way a chat socket ends — a closed tab. The `finally` below is the
+    # whole handling, and nothing failed.
     except WebSocketDisconnect:
         pass
     finally:

backend/app/routes/notifications.py

@@ -5,7 +5,7 @@ All endpoints use `current_user` (not `require_human`): notifications and their
 state are personal UI state with no quota or YouTube involvement, so the shared demo account
 manages its own inbox normally."""
 
-from fastapi import APIRouter, Depends, HTTPException, Query
+from fastapi import APIRouter, Depends, Query
 from sqlalchemy import func, select, update
 from sqlalchemy.orm import Session
 
@@ -14,6 +14,7 @@ from app.db import get_db
 from app.models import Notification, User
 from app.notifications import trim_read
 from app.utils import iso
+from app.routes._common import owned_or_404
 
 router = APIRouter(prefix="/api/me/notifications", tags=["notifications"])
 
@@ -72,20 +73,13 @@ def unread_count(user: User = Depends(current_user), db: Session = Depends(get_d
     return {"count": count or 0}
 
 
-def _owned(db: Session, notification_id: int, user: User) -> Notification:
-    n = db.get(Notification, notification_id)
-    if n is None or n.user_id != user.id:
-        raise HTTPException(status_code=404, detail="Notification not found")
-    return n
-
-
 @router.post("/{notification_id}/read")
 def mark_read(
     notification_id: int,
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    n = _owned(db, notification_id, user)
+    n = owned_or_404(db, Notification, notification_id, user, what="notification")
     n.read = True
     db.commit()
     trim_read(db, user.id)
@@ -110,7 +104,7 @@ def dismiss(
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    n = _owned(db, notification_id, user)
+    n = owned_or_404(db, Notification, notification_id, user, what="notification")
     # Dismiss implies read (it no longer counts toward the unread badge).
     n.dismissed = True
     n.read = True

backend/app/routes/playlists.py

@@ -3,10 +3,8 @@ CRUD (create / rename / delete, add / remove / reorder items) plus YouTube two-w
 mirror the user's own YouTube playlists in (`/sync-youtube`) and push/revert local edits back
 out (`/{id}/push-plan`, `/{id}/revert-youtube`), behind the optional write scope."""
 
-from datetime import datetime, timezone
-
 from fastapi import APIRouter, Depends, HTTPException
-from pydantic import BaseModel
+from pydantic import BaseModel, Field
 from sqlalchemy import func, select, and_
 from sqlalchemy.orm import Session, aliased
 
@@ -16,14 +14,18 @@ from app.db import get_db
 from app.feedquery import feed_columns, serialize
 from app.models import Channel, Playlist, PlaylistItem, User, Video, VideoState
 from app.sync.playlists import (
+    PlaylistGone,
+    lock_playlists,
     merge_playlist,
     plan_merge,
     recompute_dirty,
+    pull_user_playlists,
     repull_playlist,
-    sync_user_playlists,
+    sync_claim,
+    user_sync_claim,
 )
 from app.ratelimit import RateLimiter
-from app.routes._common import get_or_404, owned_or_404
+from app.routes._common import get_or_404, not_found, owned_or_404, youtube_refusal
 from app.youtube.client import YouTubeClient, YouTubeError
 
 router = APIRouter(prefix="/api/playlists", tags=["playlists"])
@@ -48,10 +50,6 @@ def _check_pull_cooldown(user: User) -> None:
         raise HTTPException(status_code=429, detail={"code": "sync_cooldown"})
 
 
-def _own_playlist(db: Session, user: User, playlist_id: int) -> Playlist:
-    return owned_or_404(db, Playlist, playlist_id, user, what="playlist")
-
-
 def _refused(status: int, reason: str) -> HTTPException:
     """A playlist action the server refuses, as a reason the UI translates
     (`playlists.errors.<reason>`) — a bare English detail reached the Hungarian UI verbatim."""
@@ -64,8 +62,11 @@ def _touch(pl: Playlist) -> None:
     `UpdatedAtMixin`'s `onupdate` fires on an UPDATE of the PLAYLIST row, and add / remove /
     reorder / move write the item rows instead — so without this the rail's "recently changed"
     sort would only ever see renames. The assignment is what makes the column mean "last time
-    this list changed", which is what the rail promises the reader."""
+    this list changed", which is what the rail promises the reader.
-    pl.updated_at = datetime.now(timezone.utc)
+
+    The DATABASE's clock, like the mixin's own `onupdate`: one sort key written from two clocks
+    (this process's and Postgres's) could order a rename and an item edit the wrong way round."""
+    pl.updated_at = func.now()
 
 
 def _next_playlist_position(db: Session, user_id: int) -> int:
@@ -81,7 +82,10 @@ def _next_playlist_position(db: Session, user_id: int) -> int:
 
 
 def _next_item_position(db: Session, playlist_id: int) -> int:
-    """Append position for a new item at the end of a playlist."""
+    """Append position for a new item at the end of a playlist. Locks the playlist row first: two
+    adds (two tabs, the rail and the page) otherwise both read the same max and land on one
+    position, and the list's order between them is then arbitrary."""
+    lock_playlists(db, playlist_id)
     return (
         db.scalar(
             select(func.coalesce(func.max(PlaylistItem.position), -1)).where(
@@ -94,7 +98,10 @@ def _next_item_position(db: Session, playlist_id: int) -> int:
 
 def _add_item(db: Session, pl: Playlist, video_id: str, *, mark_dirty: bool) -> bool:
     """Append a video to a playlist if not already present (idempotent). Returns whether it
-    was added. `mark_dirty` recomputes the YouTube-sync dirty flag (skip for Watch later)."""
+    was added. `mark_dirty` recomputes the YouTube-sync dirty flag (skip for Watch later).
+    Locked BEFORE the presence check: two adds of one video both saw "absent" and the second one's
+    insert hit uq_playlist_video as a 500."""
+    lock_playlists(db, pl.id)
     exists = db.scalar(
         select(PlaylistItem).where(
             PlaylistItem.playlist_id == pl.id, PlaylistItem.video_id == video_id
@@ -108,13 +115,14 @@ def _add_item(db: Session, pl: Playlist, video_id: str, *, mark_dirty: bool) ->
     if mark_dirty:
         recompute_dirty(db, pl)
     _touch(pl)
-    db.commit()
+    db.flush()  # the handler commits (CLAUDE.md: helpers flush)
     return True
 
 
 def _remove_item(db: Session, pl: Playlist, video_id: str, *, mark_dirty: bool) -> bool:
     """Remove a video from a playlist if present (idempotent). Returns whether it was removed.
     `mark_dirty` recomputes the YouTube-sync dirty flag (skip for Watch later)."""
+    lock_playlists(db, pl.id)
     item = db.scalar(
         select(PlaylistItem).where(
             PlaylistItem.playlist_id == pl.id, PlaylistItem.video_id == video_id
@@ -126,7 +134,7 @@ def _remove_item(db: Session, pl: Playlist, video_id: str, *, mark_dirty: bool)
     if mark_dirty:
         recompute_dirty(db, pl)
     _touch(pl)
-    db.commit()
+    db.flush()  # the handler commits (CLAUDE.md: helpers flush)
     return True
 
 
@@ -280,7 +288,7 @@ def _watch_later(db: Session, user: User) -> Playlist:
             position=_next_playlist_position(db, user.id),
         )
         db.add(pl)
-        db.commit()
+        db.flush()  # the id, for the item the caller adds next; the handler commits
     return pl
 
 
@@ -295,7 +303,8 @@ def add_watch_later(
     get_or_404(db, Video, video_id, what="video")
     pl = _watch_later(db, user)
     _add_item(db, pl, video_id, mark_dirty=False)
-    return {"saved": True, "playlist_id": pl.id}
+    db.commit()
+    return {"ok": True, "saved": True, "playlist_id": pl.id}
 
 
 @router.delete("/watch-later/{video_id}")
@@ -309,7 +318,8 @@ def remove_watch_later(
     )
     if pl is not None:
         _remove_item(db, pl, video_id, mark_dirty=False)
-    return {"saved": False}
+        db.commit()
+    return {"ok": True, "saved": False}
 
 
 @router.post("/sync-youtube")
@@ -321,9 +331,18 @@ def sync_youtube(
     Also runs automatically on the scheduler; this is the manual 'sync now'."""
     if not has_read_scope(user):
         raise _refused(403, "youtube_read_scope")
-    _check_pull_cooldown(user)
+    with user_sync_claim(db, user.id) as free:
-    with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PULL):
+        if not free:
-        return sync_user_playlists(db, user)
+            # Another run (the scheduler, another tab) is pulling right now: "0 synced" would lie.
+            raise _refused(409, "sync_all_busy")
+        # After the claim, like the re-pull's: a refused attempt must not spend one of the pulls.
+        _check_pull_cooldown(user)
+        try:
+            with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PULL):
+                return {"ok": True, **pull_user_playlists(db, user)}
+        except YouTubeError as exc:
+            db.rollback()
+            raise youtube_refusal(exc) from exc
 
 
 @router.post("/{playlist_id}/revert-youtube")
@@ -335,24 +354,31 @@ def revert_youtube(
     """Reset one YouTube-linked playlist to its current YouTube state, discarding local
     edits (and clearing the dirty flag). The user-facing 'undo all my changes' for a list
     once the in-session undo history is gone."""
-    pl = _own_playlist(db, user, playlist_id)
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     if not pl.yt_playlist_id:
         raise _refused(400, "not_on_youtube")
     if not has_read_scope(user):
         raise _refused(403, "youtube_read_scope")
-    _check_pull_cooldown(user)
+    with sync_claim(db, pl.id) as free:
-    try:
+        if not free:
-        with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PULL):
+            raise _refused(409, "sync_busy")
-            return repull_playlist(db, user, pl)
+        # The cooldown is spent only by a pull that runs: checked after the claim, so clicking
-    except YouTubeError:
+        # Revert while a push is in flight does not use up the three pulls per ten minutes.
-        db.rollback()
+        _check_pull_cooldown(user)
-        raise _refused(422, "youtube_failed")
+        try:
+            with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PULL):
+                return {"ok": True, **repull_playlist(db, user, pl)}
+        except PlaylistGone:
+            raise not_found("playlist") from None
+        except YouTubeError as exc:
+            db.rollback()
+            raise youtube_refusal(exc, playlist=True) from exc
 
 
 def _pushable(db: Session, user: User, playlist_id: int) -> Playlist:
     """A local playlist the user may push to YouTube (create or update). The built-in
     Watch later and read-only YouTube mirrors are never pushable."""
-    pl = _own_playlist(db, user, playlist_id)
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     if pl.kind == "watch_later":
         raise _refused(400, "watch_later_not_pushable")
     return pl
@@ -371,8 +397,8 @@ def push_plan(
     try:
         with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PUSH):
             plan = plan_merge(db, user, pl)
-    except YouTubeError:
+    except YouTubeError as exc:
-        raise _refused(422, "youtube_failed")
+        raise youtube_refusal(exc, playlist=True) from exc
     plan["remaining_today"] = quota.remaining_today(db)
     plan["affordable"] = plan["units_estimate"] <= plan["remaining_today"]
     return plan
@@ -388,25 +414,32 @@ def push(
     neither side loses an item). Refuses if the estimated quota exceeds what's left today, so a
     big sync can't strand half-applied. Write scope required (require_write_scope)."""
     pl = _pushable(db, user, playlist_id)
-    try:
+    # One sync of a list at a time, across tabs, devices and the scheduler (`sync_claim`): the
-        with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PUSH):
+    # client's own in-flight flag only covers the menus of one page.
-            # Read the live playlist ONCE (linked playlists only) and share it with both the
+    with sync_claim(db, pl.id) as free:
-            # plan and the merge, halving read-quota and closing the second read→write TOCTOU
+        if not free:
-            # window. A fresh export (no yt_playlist_id) reads nothing.
+            raise _refused(409, "sync_busy")
-            live = None
+        try:
-            if pl.yt_playlist_id:
+            with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_PUSH):
-                with YouTubeClient(db, user) as yt:
+                # Read the live playlist ONCE (linked playlists only) and share it with both the
-                    live = (
+                # plan and the merge, halving read-quota and closing the second read→write TOCTOU
-                        list(yt.iter_playlist_items_with_ids(pl.yt_playlist_id)),
+                # window. A fresh export (no yt_playlist_id) reads nothing.
-                        yt.get_playlist_snippet(pl.yt_playlist_id),
+                live = None
-                    )
+                if pl.yt_playlist_id:
-            plan = plan_merge(db, user, pl, live=live)
+                    with YouTubeClient(db, user) as yt:
-            if plan["units_estimate"] > quota.remaining_today(db):
+                        live = (
-                raise _refused(429, "quota_exceeded")
+                            list(yt.iter_playlist_items_with_ids(pl.yt_playlist_id)),
-            return merge_playlist(db, user, pl, live=live)
+                            yt.get_playlist_snippet(pl.yt_playlist_id),
-    except YouTubeError:
+                        )
-        db.rollback()
+                plan = plan_merge(db, user, pl, live=live)
-        raise _refused(422, "youtube_failed")
+                if plan["units_estimate"] > quota.remaining_today(db):
+                    raise _refused(429, "quota_exceeded")
+                return {"ok": True, **merge_playlist(db, user, pl, live=live)}
+        except PlaylistGone:
+            raise not_found("playlist") from None
+        except YouTubeError as exc:
+            db.rollback()
+            raise youtube_refusal(exc, playlist=True) from exc
 
 
 @router.get("/{playlist_id}")
@@ -415,7 +448,7 @@ def get_playlist(
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    pl = _own_playlist(db, user, playlist_id)
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     state = aliased(VideoState)
     rows = db.execute(
         select(*feed_columns(user, state))
@@ -423,7 +456,8 @@ def get_playlist(
         .join(Channel, Channel.id == Video.channel_id)
         .outerjoin(state, and_(state.video_id == Video.id, state.user_id == user.id))
         .where(PlaylistItem.playlist_id == pl.id)
-        .order_by(PlaylistItem.position)
+        # `id` breaks a position tie the same way on every read (the renumbering add does too).
+        .order_by(PlaylistItem.position, PlaylistItem.id)
     ).all()
     return {
         "id": pl.id,
@@ -447,7 +481,8 @@ def rename_playlist(
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    pl = _own_playlist(db, user, playlist_id)
+    # Locked: the new name moves `dirty`, which the scheduled mirror refresh reads under this lock.
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist", lock=True)
     if "name" in body.model_fields_set:
         name = (body.name or "").strip()
         if not name:
@@ -484,37 +519,70 @@ def delete_playlist(
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    pl = _own_playlist(db, user, playlist_id)
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     if pl.kind == "watch_later":
         raise _refused(400, "watch_later_undeletable")
-    if on_youtube:
+    # Not in the middle of a sync of it (`sync_claim`): a push that read the list before the delete
-        if not pl.yt_playlist_id:
+    # and wrote YouTube after it planned from no items at all, so it emptied the YouTube playlist a
-            raise _refused(400, "not_on_youtube")
+    # local-only delete means to keep. The push checks for a vanished list too (`PlaylistGone`);
-        if not has_write_scope(user):
+    # this refuses the delete while a sync runs rather than race it.
-            raise _refused(403, "youtube_editing_off")
+    with sync_claim(db, pl.id) as free:
-        try:
+        if not free:
-            with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_DELETE):
+            raise _refused(409, "sync_busy")
-                with YouTubeClient(db, user) as yt:
+        if on_youtube:
-                    yt.delete_playlist(pl.yt_playlist_id)
+            if not pl.yt_playlist_id:
-        except YouTubeError:
+                raise _refused(400, "not_on_youtube")
-            raise _refused(422, "youtube_failed")
+            if not has_write_scope(user):
-    db.delete(pl)  # items cascade
+                raise _refused(403, "youtube_editing_off")
-    db.commit()
+            try:
+                with quota.attribute(user.id, quota.QuotaAction.PLAYLISTS_DELETE):
+                    with YouTubeClient(db, user) as yt:
+                        yt.delete_playlist(pl.yt_playlist_id)
+            except YouTubeError as exc:
+                raise youtube_refusal(exc, playlist=True) from exc
+        # The playlist first, as every item write: the cascade deletes the items before the list row.
+        lock_playlists(db, pl.id)
+        db.delete(pl)  # items cascade
+        db.commit()
     return {"deleted": playlist_id}
 
 
+class AddItemIn(VideoRefIn):
+    # Where the video lands among the list's items, in their stored order (0 = first). Omitted, it
+    # is appended. The list page's "Undo" of a removal puts the row back where it was with this —
+    # ONE request, because append-then-reorder leaves it at the end whenever the second call fails.
+    position: int | None = Field(default=None, ge=0)
+
+
 @router.post("/{playlist_id}/items")
 def add_item(
     playlist_id: int,
-    body: VideoRefIn,
+    body: AddItemIn,
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    pl = _own_playlist(db, user, playlist_id)
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     video_id = body.video_id
     get_or_404(db, Video, video_id, what="video")
-    _add_item(db, pl, video_id, mark_dirty=True)
+    added = _add_item(db, pl, video_id, mark_dirty=False)
-    return {"playlist_id": pl.id, "video_id": video_id}
+    if added and body.position is not None:
+        # Renumber the whole list with the new row at `position` (clamped to the end): positions
+        # can have gaps after removals, so "index N" is only meaningful over the ordered rows.
+        rows = list(
+            db.execute(
+                select(PlaylistItem)
+                .where(PlaylistItem.playlist_id == pl.id)
+                .order_by(PlaylistItem.position, PlaylistItem.id)
+            ).scalars()
+        )
+        new = next(r for r in rows if r.video_id == video_id)
+        rows.remove(new)
+        rows.insert(min(body.position, len(rows)), new)
+        for pos, row in enumerate(rows):
+            row.position = pos
+    recompute_dirty(db, pl)
+    db.commit()
+    return {"ok": True, "playlist_id": pl.id, "video_id": video_id}
 
 
 @router.delete("/{playlist_id}/items/{video_id}")
@@ -524,9 +592,10 @@ def remove_item(
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    pl = _own_playlist(db, user, playlist_id)
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     _remove_item(db, pl, video_id, mark_dirty=True)
-    return {"playlist_id": pl.id, "video_id": video_id}
+    db.commit()
+    return {"ok": True, "playlist_id": pl.id, "video_id": video_id}
 
 
 class MoveItemIn(BaseModel):
@@ -548,7 +617,7 @@ def move_item(
     exactly the kind of half-applied state this codebase treats as a bug rather than an edge
     case. Idempotent: if the video is already in the target, the source row is simply dropped
     (moving into a list that already has it is a removal, not a duplicate)."""
-    src = _own_playlist(db, user, playlist_id)
+    src = owned_or_404(db, Playlist, playlist_id, user, what="playlist")
     # Its own reason: "this playlist doesn't exist any more" about a TARGET that is gone was read as
     # the list on screen — which is still there.
     dst = owned_or_404(db, Playlist, body.target_playlist_id, user, what="move_target")
@@ -556,6 +625,8 @@ def move_item(
         raise HTTPException(
             status_code=400, detail={"code": "playlist_move", "reason": "same_playlist"}
         )
+    # Both lists, in id order: a move A→B beside a move B→A would otherwise lock crosswise.
+    lock_playlists(db, src.id, dst.id)
     item = db.scalar(
         select(PlaylistItem).where(
             PlaylistItem.playlist_id == src.id, PlaylistItem.video_id == video_id
@@ -571,19 +642,24 @@ def move_item(
             PlaylistItem.playlist_id == dst.id, PlaylistItem.video_id == video_id
         )
     )
-    if already is not None:
+    dst_changed = False
-        db.delete(item)  # the uq_playlist_video constraint would reject a second row anyway
+    if already is None:
-    else:
         # Re-parent the existing row rather than delete+insert: `added_at` ("when I saved this")
-        # survives the move, and there is no window where the video belongs to neither list.
+        # survives the move, and there is no window where the video belongs to neither list. The
+        # check above holds: both lists are locked, and an insert into the target (another tab's
+        # add, a YouTube merge) waits for the lock — its foreign key needs the target row.
         item.playlist_id = dst.id
         item.position = _next_item_position(db, dst.id)
-    # Both sides changed, so both get the dirty recompute and the timestamp — the asymmetric
+        dst_changed = True
-    # version of this (touching only the source) is the recurring smell CLAUDE.md warns about.
+    if not dst_changed:
+        db.delete(item)
+    # The source always changed; the target only when the row landed in it. `_touch` on a target
+    # that already had the video promoted an untouched list in the rail's "recently changed" sort.
     recompute_dirty(db, src)
-    recompute_dirty(db, dst)
     _touch(src)
-    _touch(dst)
+    if dst_changed:
+        recompute_dirty(db, dst)
+        _touch(dst)
     db.commit()
     return {"ok": True, "from_playlist_id": src.id, "to_playlist_id": dst.id, "video_id": video_id}
 
@@ -599,7 +675,8 @@ def reorder_items(
     user: User = Depends(current_user),
     db: Session = Depends(get_db),
 ) -> dict:
-    pl = _own_playlist(db, user, playlist_id)
+    # Locked: a concurrent add between reading the items and writing the order would collide.
+    pl = owned_or_404(db, Playlist, playlist_id, user, what="playlist", lock=True)
     order = body.video_ids
     items = {
         it.video_id: it
@@ -625,4 +702,4 @@ def reorder_items(
     recompute_dirty(db, pl)
     _touch(pl)
     db.commit()
-    return {"playlist_id": pl.id, "count": pos}
+    return {"ok": True, "playlist_id": pl.id, "count": pos}