The code cited its plan instead of saying why: `E28 review J4`, `(R6 S3b)`, `S1 justified…`,
`deep-review round 5`. That plan lives in the private ops repo, so for every reader of this one the
citation was noise, and a sentence whose subject was the plan step ("round 2 showed…") hid its
lesson behind an id. Each one is gone and the "why" kept: a bare citation is deleted, a sentence
carrying a lesson is reworded around it ("a later attempt showed…").
Measured (`scripts/audit_metrics.py`, product code): 652 lines in 270 files carried an epic code by
the definition the next commit holds (507 / 241 by the old, two-digit-only one) -> 0. Across all
code, tests, scripts and configs: ~1100 epic-code lines plus ~320 sprint/round ids.
Behaviour unchanged, checked mechanically: every changed .py file has the same AST once docstrings
are blanked; css/sh/toml/yml/json files are identical once comments are stripped; the only
non-comment changes in TS/JS are test titles, one description string in a contract table, and five
ESLint messages, whose text is all the gates snapshot diff holds. Not touched: Markdown (docs and
the changelog keep their history), Alembic migrations, TV episode notation (`S2 · E5`), DESIGN.md
ids (`C1`, `B4`, `S3h`), tracker ids (`SEC-16`, `DX-17`).
One stale claim fixed on the way: test_hls_reaper's docstring said the player checkpoints "every
10 s whether or not it is playing", which plex/stream.py calls wrong; it now says what the code does.
POST /auth/request-access had no caller left: the login page's form was
gone, and registration and a refused Google sign-in already land in the
same admin approval. A third public write path with no protection of
its own (owner decision, E34).
Its one guard moves to where every path passes: upsert_pending_invite no
longer opens a pending row for an address the env already lets in. The
registration path lacked it — registering b.ob@gmail.com beside an
env-granted bob@gmail.com opened a pending row, and is_allowed answers
from the invite rows once any exist, so the real owner was locked out.
The tests that drove the endpoint now drive registration or the shared
function.
The 7 baselined Optional accesses in plex/stream.py were one fact the type did not carry: the
initial spawn (replace_only=None) always supersedes, so only the fallback respawn can get None.
An assert at the call site says so; the pyright baseline drops 195 -> 188.
test_abuse_limits: the public-URL and YouTube enqueue tests stopped at the profile lookup
(400 no_profile) and passed on that unrelated refusal. They now send an inline spec and assert
the job is queued. The same shape — a request whose response is discarded — sat at 53 more
places across the suite (all returned what they should, measured with a TestClient probe);
each now asserts its status, and test_response_checked keeps it that way (seen red).
Three low findings, all in what the review rounds themselves wrote.
The comment on `test_approving_an_unverified_account_keeps_its_password` still asserted the argument
rounds 1 and 2 refuted — that dropping the approval-time revoke was safe because `password_login`
reads `email_verified` first — while the docstrings on the two functions it describes now say the
opposite. Two contradictory "why" records for one rule is how the round-1 takeover primitive got
reasoned into existence in the first place, so the test comment now carries the corrected account
and names SEC-16.
`test_a_later_registration_leaves_a_link_already_in_flight_alive` asserted a subset relation that is
vacuously true on an empty set, and the guard that made it non-empty had gone out with the revert.
It now pins both halves: the old link survives, and exactly one new one appears.
Round 2's own fix — activation moved to after a successful send — was unpinned; putting it back
above the send left the suite green. The failed-invitation test now also asserts the account was not
activated by a request the UI reported as a failure.
Round 1 fixed the address pre-emption chain by letting a later registration replace the password on
an unverified account and delete its outstanding verify tokens. Round 2 read that code and showed it
was worse than what it fixed, so it goes back out.
It closes nothing. The rule is symmetric and an attacker can simply register last — on a timer,
inside the five-per-five-minutes budget — while the owner registers once. What it does add is two
primitives an unauthenticated caller did not have: overwrite a real registrant's chosen password,
and delete the confirmation link already sitting in their inbox (the approval mail's 7-day token is
an ordinary verify token, so the blanket delete took that too). Worst of all it fired on the rows
SEC-10 left stripped — password_hash NULL, password_revoked_at set, active — installing an
attacker's password where the entire state means "there is none", and clearing the stamp that is the
only reason password-reset still accepts them. That is a straight takeover which did not exist
before, in the code written to prevent one.
The owner's recourse needed no new primitive and already worked: password-reset accepts a
pre-empted account, proves the mailbox, rewrites the hash and bumps the session epoch. What remains
is the bounded window between a confirmation click and that reset, which is SEC-16 — logged with
this attempt written down, so the next person does not reach for it again.
Two smaller round-2 findings go with it. `upsert_pending_invite`'s new exact-spelling tie-break ran
before any status preference, so it could deterministically pick a DENIED sibling over a pending one
and reopen it — the very outcome its comment claimed to prevent; the status preference now comes
first. And `send_invitation` activated the account before the send it might have to report as
failed, so a request the UI called a failure had quietly let someone in; activation now follows a
mail that actually left. The verify token minted to render that mail still commits either way — it
is single-use, expires on its own and was mailed nowhere, and the note says so rather than leaving
the next reader to wonder whether it was missed.
Eleven findings survived adversarial verification; these are the fixes.
**Address pre-emption.** S1 dropped the approval-time password revoke, arguing `password_login`
refuses an unverified account before it reads `is_active`. Two verifiers showed that holds only
while the address stays unverified — and S6's approval mail exists to get the real owner to end that
state, at which point the squatter's password works. While an address is unverified the NEWEST
registration now owns it: the password is replaced and every outstanding verify token deleted, so
the owner takes the row back simply by registering, and a link already in flight cannot confirm it
for anyone else. The residual (an owner who confirms and neither registers nor resets) is logged as
SEC-16 beside SEC-15, where the re-attestation path it needs already lives. The comments that
asserted the old, false invariant now say what actually guards this.
**The entry URL.** S2 moved `clearEntryUrl()` out of the signed-out auth card into a hook App mounts
unconditionally, and it blanket-stripped the whole query string, the fragment and `history.state` on
the FIRST commit. That broke first-run setup outright — the wizard mounts only after its query
resolves and then finds no `?token=` — and silently ate `?link=`/`?oauth=` (no toast, no refetch
after connecting Google), `?v=`/`?chan=` deep links, and scroll restoration. It now clears only the
fragment and the two one-time flags it owns, and preserves `history.state`.
**The env allow-list.** S3 canonicalised the Invite branch of `is_allowed` and left the
ALLOWED_EMAILS/ADMIN_EMAILS fallback matching literally — the 2026-09-03 lockout with a different
source of truth, plus a shadowing trick where an unauthenticated request under a sibling spelling
permanently masked an env grant. Canonical on both sides now, and the same for the delete dialog's
warning and the ADMIN_EMAILS role bootstrap.
**Honest states.** `send_invitation` stamped "I already wrote to them" before a fail-soft background
send; it sends inline and records only what left. The admin-side resend answered `{ok: true}` for a
dropped mail; it now says so. Both also activate/serialize consistently with the other decide paths.
**Correctness sweep.** The 7-day approval link said "expires in 168 hours" — durations render in
days once there is more than one. `_register_account`, `upsert_pending_invite` and `add_invite`
picked among canonical siblings with no ORDER BY and no exact-spelling tie-break, so which account
got the re-send, and whether a denied row was reopened over a pending one, was up to Postgres.
`purge_user` deleted every canonically-equal invite, revoking a surviving sibling account's access.
`ConfirmProvider` wrapped its ReactNode message in a `<p>`. Approving no longer leaves the
stuck-sign-ups list stale. The refusal panels announce and take focus.
**Tests.** Three mailers rewritten this epic never rendered in any test; every `send_*` now renders
in both languages. The orphan-key guard could not fail (it scanned the catalog that defines the
keys) and had no reverse direction. Newly pinned: canonical activation, the surviving sibling's
grant, a failed invitation and resend, an expired reset link, a throttled login's silence, the
per-language admin fan-out, and that the entry URL keeps what it does not own.
peter