Round 2 mutation-checked round 1's own fix and found the one line it left
uncovered: every table row was an integer, so `Math.floor(sec)` in the new
seconds tier could be deleted or swapped for round/ceil with the suite still
green — exactly the "later cleanup removes it as redundant" failure the fix's
comment claims to prevent. Two fractional rows kill all three mutants.
Epic-end deep review, round 1. Six findings survived verification; five of them
were one defect wearing different hats.
formatRuntime floored 1-59s to "0m". Harmless while it only ever formatted a
Plex movie's length, but 12D repurposed it for an aggregate, and a total is the
only input that can land under a minute: a playlist of Shorts reported "0m"
while the rail beside it read "0:45". Fixed in the function, not behind a
`>= 60` guard at each of the seven call sites — the invariant belongs where it
cannot be forgotten. Its contract comment claimed "Plex movie/episode lengths"
and was two epics stale.
The rail rendered the same playlist total as formatDuration (1:23:45) while the
new header rendered formatRuntime (1h 23m) — two shapes of one number, side by
side on one screen. The rail now uses formatRuntime too; formatTotalHours had
already recorded that H:MM:SS is wrong for an aggregate.
The rename input kept min-w-0 without flex-1 in the compacted row, so it was
the only shrinkable child beside a shrink-0 action cluster. The Plex half of
this same epic already had the fix; this restores the symmetry.
plex.browse.items had no plural forms, so a one-item Plex playlist read
"1 items". Pre-existing, but the new header is the call site where count=1 is
ordinary. Both languages get both forms, no key drift.
formatRuntime had no unit tests at all across seven call sites. Added a tier
table plus a sweep asserting no non-zero second ever renders as "0m".
Dropped after 3/3 adversarial refutations: the meta-line segment ordering
(the filtered count is labelled, and count-then-runtime is the convention) and
the null-duration lower bound (the server's own SUM has skipped NULLs since
long before this branch; changing only the client would split them).
Parity with the YouTube playlist header (E9 12D): the Plex playlist detail
now carries a meta line under the title — "N items · Xh Ym", summed from the
items' duration_seconds. No backend change: leaf_card already serializes
duration_seconds for both movies and episodes, and no new i18n key either —
the count reuses plex.browse.items, the same key the collection header uses.
The title/rename input moved into a min-w-0 flex-1 column so the meta line
sits under both; the line is hidden on an empty playlist.
The header portals into the shell's FIXED toolbar band, so its height is
list height the user never gets back — and it was spending four stacked
blocks on it: title, meta line, an action row, then a separate view-control
row under the cover. Option B collapses that into two tight rows beside a
smaller cover: title + rename/YouTube-link on the first, with Play-all,
push/revert/delete and the unsynced badge pushed right; the meta line
(count · playtime · filtered) and the view controls (group-by, reset view,
save order, view status, undo/redo) sharing the second.
Every control and its gating is unchanged — the view-control group still
appears only for `items.length > 1 || viewActive`, so a one-item playlist
shows just the meta line.
Measured on a 32-item playlist: the band goes 170px -> 88px at rest
(122px while a view is active, where the status sentence wraps), which is
~80px more visible list.
The playlist content wrapper carried `pt-3`, so its sticky column header sat
12px below the scroller top at rest and jumped up to `top:0` on the first
scroll. Drop the top padding so the header is flush from rest and never
shifts — the fixed toolbar band above supplies the breathing room, matching
the channel table's already-correct pattern.
Surface the playlist's total playtime beside the item count in the detail
header (12D) — sums every item's duration (formatRuntime "h m", matching the
approved mockup), hidden when nothing sums. Data is already client-side; no
backend change.
Pass fadeTop:false for the playlists page (12B) so the shell's top edge-fade
no longer sits over the already-sticky column header (tableHeader.tsx:257),
matching the channels/audit pages.
The bodies are text/plain only, so whether the URL in a verification or
password-reset mail is a link at all was the client's decision. Gmail, Outlook and
Apple Mail linkify plain text; lighter web clients render the body verbatim and leave
the link as dead text — in a message whose entire purpose is that one link.
_send now attaches an HTML part generated from the SAME plain body, rather than
per-template HTML that would have to be kept in sync with it: one converter, all
eleven templates, and text/plain stays the primary part and the source of truth.
Escaping runs before linkification — bodies interpolate untrusted values (a
requester's address, the operator contact) — so no caller can inject markup, and a
query string's `&` is already `&` when it reaches the href, which is what an HTML
attribute wants. Sentence-final punctuation is trimmed out of the href while
url-safe token characters (`-`, `_`) are left alone. Inline styles only, no images,
no external CSS, no tracking.
Review found the backfill's fourth predicate `AND email_verified = false` strands
exactly the subset that needs it most. The revoke does not invalidate the account's
outstanding verification token, so clicking the emailed link is the natural next
thing such a user does — and it sets email_verified=True while password_hash stays
NULL. That row then has no route back at all: reset ineligible, resend refuses an
already-verified address, set-password needs a session the account cannot obtain,
and nothing re-stamps it because the revoke itself is guarded on being unverified.
The remaining three predicates already pin the victim set exactly: registration and
the setup wizard always set a hash, every Google row carries a google_sub (adoption
assigns it before revoking), and nothing in the codebase ever clears google_sub — so
a local, non-demo, passwordless row can only have come from the revoke. Widening is
therefore safe, and it is the argument the docstring was already making while the
SQL contradicted it.
Also corrects that docstring: set-password is unreachable because the account cannot
authenticate at all, not because the epoch bump killed a session — an unverified
account is refused at login before a session is ever minted, so it never had one.
Two independent bugs made an email+password registration unrecoverable in prod.
The verification link did nothing. urlWithParams/clearParams rebuilt the URL as
pathname+query and dropped the fragment — but the one-time auth tokens ride the
fragment (#verify=/#reset=, SB3), and the providers that mirror nav params
(useUrlMirror in FeedFilters/Playlists/Plex/Channels) sit ABOVE the lazily-loaded
<App> in main.tsx. Their mount effect replaceState'd the token away before the auth
card's chunk had loaded, so the card read an empty hash: the link opened a bare
sign-in page, no request, no banner. Both helpers now carry the fragment over.
The account was then locked out for good. SEC-10 nulls password_hash when an
unverified account is attested (here: admin approval landing before the user could
verify — which the bug above made impossible). That is indistinguishable from a
Google-only account, and reset eligibility keyed off password_hash alone, so no
reset link was ever issued: no password to sign in with, and no way to get one.
users.password_revoked_at records the revocation, reset eligibility honours it, and
setting a password again (reset-confirm or set-password) clears it. An account that
never had a password stays ineligible on purpose — its owner adds one from Settings.
/auth/verify now also answers password_reset_required so the card can say why the
registered password stopped working, instead of leaving the user at the uniform
"Invalid email or password". Telling the holder of a valid verify token leaks
nothing: they have already proved they own the mailbox.
The 0078 migration backfills the stamp for rows carrying the exact lockout signature
(local, non-demo, no password, unverified), so existing victims recover unaided.
Round 4's only finding: the file header and one comment still said Deny stops a
refused person re-asking. It does not — upsert_pending_invite reopens a denied
row as pending — and the delete_invite docstring in the same branch already says
so. The real difference is that Deny records an auditable decision and mails the
requester; Delete forgets the address.
Round 3's correctness/security pass came back clean; what was left said more than
the code does.
- emailOff promised the access-request list is where sign-ups show up. A sign-up
on an address that is already approved or already pending adds nothing to it
(roadmap UX-25), and "only" was falsified by this release's own inbox notice.
- deleteConfirm claimed the admin was dropping "an approval you already gave" —
it is the confirmation for pending and denied rows too.
- sendFailed blamed the mail server for rejecting a message; the reason covers
connect, DNS, TLS and login failures alike, so it now describes the symptom and
leaves the diagnosis to the log.
- ADMIN_EMAILS re-promotion happens in the Google callback only, not on any
sign-in.
Tests: the decided-list copy of the delete affordance was never rendered (both
new cases fail if that branch is dropped), and the throttle release — a
deliberate behaviour the notes advertise — was asserted by nothing; a drop with
no admin yet, or with a failing write, must not silence the following hour.
Round 2 on the feature diff (15 → 12):
- still_allowed_by_env now rides the invite LIST as well as the delete response,
so the confirmation warns while the admin can still change their mind. It
mattered most for a DENIED row on an env-allow-listed address: deleting it
silently converted a refusal into a grant, and the only warning arrived after
the irreversible click.
- Both inbox bodies claimed to name "the most recent" dropped message. The
throttle records the FIRST of each hour and never updates it, so they now say
"one of them", and the not-configured body finally names where to fix it.
- settings.invites.emailOff told the admin they "get no notice when someone
signs up" — rendered directly above the queue that shows exactly that. Only
the email is lost, and it now says so.
- The disabled delete buttons put their reason out of reach: `disabled` drops a
control from the tab order and suppresses its title. Switched to
aria-disabled + preventDefault, the pattern Playlists.tsx already established
for the same reason.
- deleteInvite is quiet, so the 409 the UI handles with a translated toast no
longer also raises the server's English detail as a modal.
- After a completed password reset, Back landed on the reset form again with a
consumed token — the history change made that state reachable. The token is
cleared on success so the reset view stops being the card's root.
- The hu deleteHasAccount named a tab that renders under a different label.
Tests: a dom test for the Access-requests panel covering both has_account
variants and all three delete outcomes, /api/me flag cases, an ADMIN_EMAILS
case, and the env-allowlist test now patches the raw setting so the parsing
stays under test.
Round 1 on the feature diff found 15; the substantive ones:
- The delete control VANISHED on any invite whose address had registered, which
is the ordinary sign-up shape (register writes the User and the pending invite
together). The guard stays broad — that row is the account's approval gate, and
orphaning it strands someone who can then neither sign in nor be approved — but
the control is now a disabled button carrying the reason, on both lists, so an
admin never sees a row silently missing a button its neighbours have.
- Deleting an entry is not always a revoke: is_allowed falls back to the env
ALLOWED_EMAILS / ADMIN_EMAILS lists once no row exists, and an ADMIN_EMAILS
address is force-promoted on its next Google sign-in. The route now reports
still_allowed_by_env and the UI says so instead of claiming access was removed.
- "Deny keeps the row so they can't immediately re-ask" was simply false —
upsert_pending_invite reopens a denied row. Corrected everywhere it was stated.
- The Configuration warning reads me.email_configured, which save() did not
invalidate, so it kept warning about fields the admin had just filled in.
- The read-only variant was an aria-hidden span explained only by a Tooltip that
the hints toggle can switch off; now a disabled button with title + aria-label.
- The delete error path threw away the 409's reason (the stale-list case) and
left the stale row on screen.
- The throttle window was spent even when no notice was written, silencing the
next hour on the path whose job is to stop silent drops.
- email_enabled() opened a second session and decrypted the SMTP password on
every /api/me poll; it now takes the caller's session.
- welcomeMode.dom.test.ts never rendered the card, so reverting the history
wiring to useState left it green. A render test now drives the real links and
fails on that revert (verified).
Plus /api/me flag tests, the registration-shaped pending case, the env-allowlist
case, and a release note that matches what the code does.
Three operator gaps, all found the same way — a person could not sign up and the
app showed nobody anything.
- DELETE /api/admin/invites/{id} with a delete control on both the pending
requests and the previously-decided list. Deny keeps the row so the person
cannot immediately re-ask; delete forgets the address. Refused with 409 while
a User row exists for it, because is_allowed reads that row and removing it
would revoke a live user's access with no other sign — the list ships
has_account so the UI withholds the button instead of offering one that fails.
- Email is fail-soft by design, which had come to mean invisible. _send() now
records an email_undelivered notification for every admin when a message is
skipped or rejected, naming the message that was lost, throttled to one an
hour. /api/me carries email_configured, driving a standing warning on
Configuration → Email / SMTP and above the access-request list.
- The sign-in card's Create account / Forgot password / Didn't get the email
views are history entries (useHistorySubview), so Back returns to the previous
view instead of only moving the address bar.
Also: the per-test rate-limiter reset sweeps every loaded app.* module rather
than a hand-kept list of three — this change added a limiter outside it.
Round 6's two remaining nits. The operator note said "Configuration → Email"
but the group renders as "Email / SMTP" in both locales, and
test_an_admin_can_activate_a_pre_approved_registrant requested make_user without
using it — a leftover from the reverted rounds that made the test look like it
seeded an admin when it did not.
Round 5's correctness/security pass came back clean; what was left was text that
said more than the code does.
- The register-path notify comment (and its test) claimed the admin "is still
told". The call fires, but on a no-SMTP instance _send drops it — which is the
open half of the problem, not something this branch fixed. Both now say so.
- The release notes tell admins what to actually do: an account whose address was
whitelisted in advance has no pending request to approve, so the whitelist box
is what activates it. That instruction only existed in the roadmap and in
chores, which the dialog does not render.
- "That one line is the whole production change" now accounts for the test-only
RateLimiter.reset() that also ships.
- The rate-limiter test claimed to pin that blocked attempts are not counted, on
an assertion that held either way. Replaced with one that drives the clock
across the window boundary and fails if hammering extends it.
Round 4 showed the `inv is not None` gate added in b5a1002 removed a signal that
matters: for an address the admin had already approved, that email is their only
notice that someone has now created an account on it — the invite is not pending
so nothing lands in the Access-requests queue either. Suppressing it left the
registrant at "awaiting admin approval" with the admin unaware they existed. The
duplicate it avoided (Request access, then sign up) is mild noise by comparison.
The production change is now one db.commit(), which is all the reported bug ever
needed. Release notes corrected again: a previously-denied address was also
unaffected, the pending-request case never lost its queue entry (only the
account behind it), and the test inventory was missing a guard.
Round 3 of the review killed the reopen from b5a1002: flipping an approved
invite back to pending let an unauthenticated POST /auth/register REVOKE a
standing approval for any whitelisted-but-unclaimed address, which then fails
is_allowed() and locks the real owner out of Google sign-in — the primary way
people get in here. It also re-armed the pre-emption window it was meant to
close, because with no SMTP email_verified is set True and the SEC-10 guard the
design leaned on never fires.
Two attempts at fixing the stranded-account UX in the same change both came back
worse than the wart. What remains is the narrow, real bug: the account was
silently rolled back and is now committed. An admin still binds the approval to
the account, exactly as before this branch. UX-25 in the roadmap records the
follow-up and both dead ends so they are not retried.
Also from round 3: the upsert docstring no longer claims the admin UI cannot
reach these accounts (re-entering the address in the same panel's whitelist box
does, and a test now pins it), the release notes are scoped to what the diff
actually changes with the operator note moved where the dialog renders it, and
RateLimiter has unit tests so reset() is not production code only a fixture
calls.
Replaces the activation added in ea931dc, which round 2 of the review showed was
a real weakening: on a no-SMTP instance it made a standing whitelist entry a
bearer grant — the first anonymous party to register an approved-but-unclaimed
address got an immediately usable account under that identity, with no
attestation, no admin action and no signal anywhere. The verify-time activation
had the same shape one step removed, and neither site ran the SEC-10 preset-
password guard or left an audit trail.
Keep a human in the loop instead. An approval granted before any account existed
was a decision about the ADDRESS, so registration reopens the invite
(upsert_pending_invite gains revive_approved, used only by the register worker).
The request lands in the Access-requests queue, approving it runs
_activate_user_by_email — and with it revoke_preset_password_if_unverified — and
the account activates. One mechanism, already audited, no new grant path, and
the "awaiting admin approval" copy is honest again.
The explicit commit from the previous commit still matters: with an invite
already pending, upsert returns without committing and _issue_token is gated on
SMTP, so a no-SMTP registration was still being rolled back.
Also from round 2: gate the admin notice on a fresh decision (matching the two
other upsert callers) so a retry no longer re-pings every admin, and reset the
module-level rate limiters between tests — the suite pooled every login into one
300s window and was two logins short of failing an unrelated test with a 429.
Follow-up to the commit fix, from the deep-review. Persisting the account was
only half the story: when the admin had approved the address BEFORE the person
registered, the account landed is_active=False against an invite that is no
longer pending — absent from the Access-requests queue, and approve_invite 409s
on an already-decided invite. The sign-up worked and still could not sign in.
Derive activation from is_allowed() at the moment ownership is attested:
- verify_email_confirm activates a pre-approved account when the emailed link is
confirmed (only the mailbox owner can follow it), so SEC-10 stays intact;
- _register_account activates only when there is no SMTP to verify with, where
the admin whitelist is by design the sole gate — the same one add_invite
already activates on.
Tests: the pre-approved no-SMTP case and the verify-activates case are real
regression tests (both fail without this change); guards cover the unknown-email
and denied->pending branches, verification of an unapproved account, and the
first end-to-end coverage of the /auth/register route and its
allow_registration gate. Release notes rewritten — the previous entry miscounted
its regression tests and advertised an email fix this release has no code for.
_register_account() flushed the new User but relied on someone else to commit:
upsert_pending_invite() commits only when it actually writes an invite row, and
_issue_token() runs only when SMTP is enabled. For an email an admin had ALREADY
approved, on an instance where email is unavailable, neither fired — the session
closed and the account was rolled back. The would-be user got the uniform "ok"
and no account, could not sign in, and re-registering was a silent no-op.
Commit explicitly after the invite upsert (a no-op when upsert already committed
the same transaction). Three DB-lane regression tests cover the pre-approved
no-SMTP path (fails without the fix), the unknown-email path that always worked,
and the pre-approved path with SMTP on.
hls.js startPosition:0 so the transcode/remux HLS session starts VOD-style
at the requested position instead of chasing the live edge of the growing
EVENT playlist.
Our HLS is a growing EVENT playlist with no ENDLIST, which hls.js reads as
LIVE. Its default startPosition (-1) then starts at the live edge and hugs it
(no forward buffer → rebuffer spinners), and on a resume/seek session the
segments' absolute PTS (-copyts, offset by the server's -ss) don't line up
with the live-edge timeline → 0 fragments load → permanent stall. The
seek-restart model makes fragment 0 the requested content position, so start
VOD-style at 0 and buffer forward. Fixes the transcode playback stall/rebuffer
found in 0.78.2 UAT.
Deep-review round 1 on bug/download-mediakind-probe found a real HIGH the prior
review missed:
- resolve_media_facts used a per-field `probe or info` fallback, so a probe that
RAN and deliberately found no video track (cover-art-only audio: attached_pic
skipped; a still image: format-guarded) had its None video overridden by
yt-dlp's stale vcodec. An audio-mode download from a progressive source (info
carries the source's video codec, embed-thumbnail on) was then classified VIDEO
-> a Play button + a broken in-app <video> with no video track. Only the
download path regressed (upload calls probe_media directly).
Fix: gate precedence on whether ffprobe RAN, not per-field. If the probe read
the file, its codec verdict wins outright (a None means 'no such track'); only a
full ffprobe failure falls back to info. Dimensions still gap-fill from info.
Refactored the parsing into _facts_from_probe so resolve_media_facts can see the
raw probe-ran signal; probe_media keeps its 5-tuple (upload + TestProbeClassifier
unchanged).
- tests: the cover-art-audio and still-image resurrection cases, a differing
populated-info precedence case (the prior-flagged gap), and the ffprobe-failed
info fallback — now stubbing run_ffprobe_json so the real parsing + precedence
run together.
Backend gate green (ruff + full pytest).
Deep-review round 2 on the round-1 fixes (all low):
- part_key churn (state): round 1's part-key reset re-probed EVERY sync for a
multi-version item (an original + a Plex 'Optimized' copy), because
_media_facts reads media[0]/part[0] whose order Plex doesn't guarantee stable,
oscillating transcode_hw between the two versions. Gate the part-key signal on
a single-version item (len(Media)<=1); codec-change reset still applies to all.
A multi-version same-codec re-encode not self-healing is an acceptable rare
edge; a per-sync re-probe loop is not.
- dead i18n key: this diff's playableClass() collapses to transcode_hw/sw, so the
bare plex.playable.transcode key it orphaned is now unused (knip can't see
dynamic keys). Dropped from both locales, parity kept.
- test: cross-facet narrowing by an active playable class (a co-facet's year
range collapses to the selected subset) — the flip side of self-exclusion,
previously unpinned.
The aria-label+title double-announcement on the card dot is a deliberate accepted
trade-off (reliable SR name + mouse tooltip). Backend + FE gates green.
Deep-review round 1 on feature/plex-playback-class-filter:
- Sticky watermark (medium): _apply_item reset the bit-depth probe watermark
only on a codec-string change, so a same-codec re-encode (8-bit->10-bit HEVC)
kept a stale video_bit_depth/transcode_hw forever (the backfill only re-probes
media_probed_at IS NULL rows). Also reset when the Plex Part key changes (the
file was replaced), not just the codec.
- i18n parity (medium): the HU playback-class labels were mixed-language (only
'direct' translated). Translated remux/transcode_hw/transcode_sw and the
tooltips consistently to Hungarian (Kozvetlen streameles / Atkodolas).
- Layering (low): _PLAYBACK_CLASSES was module-private yet imported across into
routes.plex.browse; renamed to public PLAYBACK_CLASSES (it is the shared
source of truth for facet order + zero-fill).
- a11y (low): the browse-card playback-class dot conveyed meaning only via a
mouse-hover title; added role=img + aria-label so SR/touch users get the class.
- tests: the NULL/unprobed transcode_hw -> SW fold (filter + facet), facet
self-exclusion under an active playable filter, the migration codec-set ==
HW_DECODABLE_CODECS drift guard, the probe-batch bound + oldest-first order,
the transient probe-failure retry path, and the same-codec re-encode re-probe.
Backend gate green (ruff + full pytest); FE gate green (tsc/lint/knip/prettier/
vitest). NOTE: the facet-vs-runtime disagreement for 10-bit HEVC (C1) is the
known merge-order constraint — this branch's runtime is codec-only; the facet
becomes truthful once BUG1/tonemap's bit-depth-aware routing lands with it. Left
as a sequencing constraint, not code-changed here.
Deep-review round 3: the only surviving finding — a comment-drift nit. The
fallback's cull-yield comment named the route's 'Not available yet' path, but a
culled-key segment request actually returns 'No active playback session' (the
substantive invariant — never a None that 404s as file-not-found — is unchanged).
Reworded to match the route. Concurrency + correctness came back clean.
Deep-review round 2 on the round-1 fix:
- Race (medium): the round-1 identity guard did the supersede check and the
respawn in TWO separate _lock sections. A newer same-key session installed in
the gap was still torn down by _spawn's unconditional pop+kill (the round-1
comment even had the direction backwards). Fold the check into _spawn via a
new replace_only= param: the identity test and the pop+replace now share ONE
lock, so the respawn is a true no-op when superseded/culled.
- Honest states (low): the superseded branch returned current_session(key),
which is None when the key was cap-culled during our out-of-lock wait; the
route turns None into a dishonest 'Media file not found on disk' 404 for a
file that exists. Yield 'current_session(key) or s' so a culled session
degrades to the route's retryable 'not available yet' path instead.
- Hygiene: _want_hw docstring corrected (it is a test-only shim now; production
decides via _select_encode) and the two stale _want_hw references refreshed
(stream.py header, config.py plex_transcode_hw).
- tests: the cap-cull no-respawn/no-None sub-case; the _opencl_present except
OSError branch (glob raising, not just empty); tightened a tautological -vf
assertion. Backend gate green (ruff + full pytest).
Deep-review round 1 on feature/plex-hdr-hw-tonemap:
- F1 (correctness): _select_encode gated tonemap on bit-depth alone, so a
10-bit *SDR* (bt709) source got the HDR hable tone curve applied, visibly
altering an SDR picture where only a 10->8-bit downconvert was needed. Split
the ffprobe gate into a 3-way classifier (_probe_video_kind: hwsafe / hdr /
sdr10); only true-HDR (PQ/HLG) takes the OpenCL tonemap graph, 10-bit SDR
drops to software libx264 (an exact downconvert). Applies to both auto and on.
- F2 (race): the runtime HW->SW fallback popped+killed _sessions[key] with no
identity check. Because wait_for blocks OUTSIDE _lock, a concurrent same-key
_spawn (rapid seek-restart) or _enforce_cap cull could kill this session; the
fallback then respawned software over the NEWER session and breached the
transcode cap. Guard the respawn: only fall back if _sessions[key] is still
this session, else yield the currently-registered one.
- tests: 3-way classifier (incl. HLG + empty-probe), real _opencl_present glob,
real wait_for proc fast-fail branch, tonemap-variant fallback, the F1
regressions (auto + on), the supersede/no-respawn race, and the on/opencl
matrix cell. Backend gate green (ruff + full pytest).
Upgrade mode-3's 10-bit/HDR branch from "route to software libx264" (BUG 1's
behaviour) to a full-GPU pipeline: VAAPI decode → OpenCL tonemap (hable,
HDR10/HLG → bt709 SDR nv12) → h264_vaapi. Gen 9.5 (UHD 630) can't tonemap in
the VAAPI VPP (tonemap_vaapi fails), so OpenCL is the only HW route — proven on
the prod iGPU at ~4.2x realtime, ~1 CPU core.
- stream.py: `_select_encode` returns (hw, tonemap); a 10-bit/HDR HW-decodable
source picks the tonemap graph when the OpenCL runtime is present, else drops
to software. `_opencl_present` gates on a registered ICD. `_venc_opts` /
`_hwaccel_in_opts` gain the tonemap variant (named vaapi+opencl devices,
`-filter_hw_device ocl`, hwmap→tonemap_opencl→hwmap reverse). `_want_hw` kept
as the hw-half wrapper. The runtime HW→SW fallback still backstops a session
that emits no segment.
- Dockerfile: add the Intel NEO OpenCL runtime (intel-opencl-icd). It was removed
from Debian trixie, so it's pulled from bookworm (22.43 still supports Gen 9.5;
the .deb runs on the trixie base — verified live). Harmless without /dev/dri.
- route + FE: thread `tonemap` through the session response; the stats overlay
mode row reads "transcode · GPU · HDR→SDR" for the tonemap path.
- tests: tonemap command branch, the (hw, tonemap) selection matrix (auto/on ×
8-bit/10-bit × opencl), start_session tonemap wiring, and the fmtMode tag.
Adds a movie-scoped "Playback" filter to the Plex feed so a user can isolate items by how
they play — the groundwork for collecting HW-transcodable (or no-transcode) versions.
Four classes derived from the stored `playable` + a new `transcode_hw` facet:
- Direct Play (direct), Direct Stream (remux)
- Transcode · HW — playable=transcode AND the iGPU can hardware-transcode it
- Transcode · SW — playable=transcode, software libx264 only
The HW/SW split is precise, not codec-only: a 10-bit/HDR HEVC is HW-decodable yet breaks the
VAAPI nv12 pipeline (the same case plex/stream special-cases), so it must count as SW.
Backend:
- New PlexItem columns (migration 0077): transcode_hw, video_bit_depth, media_probed_at.
Migration backfills transcode_hw codec-only so the filter works immediately.
- `plex.sync.transcode_hw_capable` is the single classifier; the HW-decodable codec set moves
to sync as `HW_DECODABLE_CODECS` and plex.stream imports it, so the catalog facet and the
runtime encoder choice can't disagree by codec.
- Bit-depth backfill (`_probe_media_batch`, wired into sync): probes only HW-decodable transcode
items' video Stream for bit-depth + HDR and refines a 10-bit/HDR HEVC to software. Bounded
per run, watermarked by media_probed_at. A codec change resets the watermark (stale probe).
- Filter + self-excluded facet on a shared `playback_class_expr` CASE (movie-leg only — a show
is an aggregate with no single class). transcode_hw threaded onto the feed/detail/card JSON.
Frontend:
- Multi-select "Playback" Combobox in the sidebar (movie scope, facet-empty drop-out), active
chips, 4-way corner-dot badge (green/blue/amber/red) + tooltip, and an info-page playback chip
with a codec · bit-depth subtitle. New `bg-info` (blue) token, i18n HU/EN parity.
- Fix a regression where the card's playback dot lost its hover tooltip: the full-card reveal
overlay swallowed the cursor; the dot now sits above it (z-20).
Gate: backend full lane + frontend typecheck/lint/knip/prettier/vitest all green. Browser-UAT'd
on localdev (filter narrows to the SW set, facet counts, badge colour + tooltip, info chip).
A 10-bit/HDR HEVC source (e.g. an HDR10 x265 file) died at scale_vaapi=format=nv12
on the UHD 630 (Gen 9.5 VPP: "VAProfile not supported"); ffmpeg produced 0 segments
and the item never played, with no fallback — a retry just re-picked the failing HW
path. An 8-bit HEVC and the same file on libx264 both work.
Two defenses:
- `_want_hw` auto mode now also gates on an 8-bit-SDR source via a new
`_probe_video_hwsafe` ffprobe (bit depth from pix_fmt/bits_per_raw_sample, HDR from
color_transfer) — a 10-bit/HDR source takes the software libx264 path, which
downconverts to yuv420p cleanly. The Plex mirror stores no bit-depth signal, so this
is probed at session start. "on" still trusts the admin and skips the gate.
- `start_session` adds a runtime HW→SW fallback: a HW session that emits no first
segment respawns on libx264 and reports hw=False (the route/stats overlay honestly
show CPU). `wait_for` gained an optional `proc` so a dead ffmpeg fails fast instead
of burning the 25s budget before the fallback fires.
Tests: TestProbeVideoHwsafe (parsing + probe-failure-assumes-safe), the 10-bit auto
gate + "on" skip in TestWantHw, and the no-segment HW→SW respawn in
TestTranscodeStartSession. Backend lane green.
A downloaded video was silently demoted to a download-only "file" (no
in-app play/edit) whenever yt-dlp's info dict omitted vcodec/acodec — which
it does for Facebook and other sources. Finalize derived media_kind from
those nulls (`derive(None, None)` -> file), so a genuine H.264/AAC MP4
became non-playable in the library.
Fix: probe the finished file on disk (the ground truth) at finalize and
derive from that, falling back to the info dict only for a field the probe
couldn't read. This is the same ffprobe the upload ingest already used, now
extracted to a shared downloads/probe.py so the download worker and the
upload path share one classifier (and can't drift). Also backfills the
stored vcodec/acodec/width/height from the real file.
Tests: TestResolveMediaFacts pins probe-wins-over-info (incl. the Facebook
no-codec case and yt-dlp's literal "none"); upload's probe tests repoint to
the shared module.
The runtime image ships ffmpeg with h264_vaapi/libva2 but no VA driver, so
h264_vaapi can't initialise and a Plex transcode silently falls back to
libx264. Add the iHD driver (intel-media-va-driver) so the UHD 630 iGPU can
hardware-encode once /dev/dri is passed into LXC 118 (host + compose side,
tracked in siftlode-ops). Harmless where no render node is present.
peter