Step 4 rewritten: /Library/Refresh is a silent no-op on this build,
must POST to /ScheduledTasks/Running/<scan-task-id> directly. Old
endpoint moved to known-broken table.
Step 5 rewritten: /Items/Counts is scope-cached and stays stale
even after items are indexed. Use /Shows/<id>/Episodes?Season=<NN>
as authoritative verify with provider + image-tag checks.
Both bugs surfaced in archer-s02-2009 run. LibraryMonitor inotify
auto-fire also confirmed broken (failed on lilo-stitch-2002 and
archer-s02-2009 runs).
Replaced user@192.168.0.100 (LAN IP, RFC1918 — flagged by
gitleaks lan-ip-rfc1918) with user@nullstone throughout. SSH config
already aliases nullstone -> 192.168.0.100. Aligns with CLAUDE.md
two-file doc rule: IPs belong in SYSTEM.md, not operational docs.
audit-coverage.py DEFAULT_OUT still pointed at processes/subtitles/
COVERAGE.md after the processes/->playbooks/ rename. Updated to
playbooks/subtitles/COVERAGE.md.
Refreshed COVERAGE.md to current state (230 episodes total, was 197):
+ Archer (10) — OK-EMBED
+ Maul Shadow Lord (10 new 2160p) — OK-EMBED
+ Maul [Before Upscale] (10 split) — OK-EMBED
+ Lilo & Stitch movie — sidecar OK
+ 9 American Dad gaps closed by owner since last audit
+ Big Lez / Donny & Clarence / Mike Nolan now 100%
Strict-sidecar shows: Sassy, Big Lez, Donny & Clarence, Mike Nolan,
American Dad (84% partial — 9 gaps remain). Other shows OK-EMBED
which is playable but not STYLE.md-compliant — fetch optional.
13 eps imported, all matched on TVDB/IMDb. Run notes flag two
new playbook bugs found: /Library/Refresh is a silent no-op (must
trigger /ScheduledTasks/Running/<id> directly) and /Items/Counts
is scope-cached (use /Shows/<id>/Episodes?Season=N for verify).
Pending follow-ups recorded for v1.1 playbook revision.
10 episodes, 1080p AI x265 10-bit Joy release, HE-AAC 5.1 ENG +
3x embedded DVDsub (ENG/SPA/FRE) per episode. Per README.md:41
quality bar — AI-upscale acceptable when source doesn't support 4K.
Items/Counts: Series 11->12, Episodes 207->217. Series item
9d22c409d531...
Run log notes 3rd consecutive LibraryMonitor failure — playbook
v1.1 should make manual /Library/Refresh MANDATORY. Same TMDb
auto-match miss as Maul; flagged for owner UI Identify step or
[tmdbid-NNNN] folder token.
Saved variant 1 "Netflix-cinema" detail-page redesign to
web-overrides/skins/detail-variant-01-netflix-cinema.html for
future reference. Not applied to dev/prod.
Imported Star Wars: Maul - Shadow Lord S01 2160p WEB-DL HEVC SDR
(10 episodes, ~21 GB) per playbooks/import-media/ v1.0. First
"replace-with-comparison" run:
- Existing 1080p upscale renamed in-place to
"Star Wars - Maul - Shadow Lord [Before Upscale] (2026)/"
with tvshow.nfo <lockdata>true</lockdata> to suppress Jellyfin
TMDb auto-merge with the new canonical.
- New 2160p staged on onyx, rsync'd to nullstone canonical path
/home/user/media/tv/Star Wars - Maul - Shadow Lord (2026)/Season 01/
- Both series live as separate items; Items/Counts bumped
Series 10->11, Episodes 197->207.
Run log at playbooks/import-media/runs/star-wars-maul-shadow-lord-
2026-2160p.md flags v1.1 follow-ups: document the
[Before Upscale] pattern, rsync resume idempotency, and
[tmdbid-NNNN] folder token for titles that fail auto-match.
ffprobe confirmed E01 = HEVC Main 8-bit 3840x2160 SDR @ 12 Mbps,
EAC3 5.1 ENG Atmos + ITA dub, 4x embedded subrip per episode.
Direct-play on capable clients.
TMDb match failed auto on both folders (recent Disney+ release);
operator to manually identify via UI.
Adds web-overrides/popup-designs/ with the 4-up preview (preview.html) and
three standalone candidate designs (a-strip.html / b-terminal.html /
c-minimal.html) so we can revisit the alternates later without re-running
the design generator. Owner picked A.
Design A is wired into Jellyfin's stock .upNextDialog by overriding its
CSS to a full-bleed bottom 26 % strip with white 'Start Now' CTA and a
custom SVG countdown ring that mirrors .upNextDialog-countdownText. The
DOM stays intact so Jellyfin's own countdown timer and click handlers on
.btnStartNow / .btnHide keep working untouched.
Shim is bracketed by NEXT-EP-POPUP-BEGIN / NEXT-EP-POPUP-END markers
inside the existing ARRFLIX-SHIM block in web-overrides/index-dev.html.
Only deployed to dev (dev.arrflix.s8n.ru) for spot-check; promote to
prod once verified by editing prod's index.html the same way and
redeploying via the nsenter trick.
Adds bin/revert-next-ep-popup.sh — sed-deletes between markers, defaults
to dev with --prod flag for prod target. Saves a timestamped backup and
prints the redeploy command.
Adds lib/audit-coverage.py: queries Jellyfin live for every series, every
episode, and every movie; classifies each by whether the English subtitle
comes from a sidecar, embedded stream, or doesn't exist; renders a
Markdown report with one-char-per-episode bars for visual scanning. Output
file is processes/subtitles/COVERAGE.md, regenerated on demand.
v2 sub-rest-fetch.py and v3 sub-a7d-fetch.py now invoke the audit at end
of a successful run, so the committed coverage file stays in sync with
library state without manual intervention. v3.5 yt-fetch path skips the
auto-call since it doesn't speak to Jellyfin directly; run audit manually
after copying YT sidecars to nullstone.
README.md surfaces the audit at the top so anyone landing in the recipe
folder sees current state before starting a run.
Owner accepted Sassy the Sasquatch S01 v3.5 YouTube-auto-CC subs as
'85 % acceptable, fine, not great' but flagged them for v4 WhisperX
rebuild. Adds a single worklist file (STOPGAP-SUBS.md) so every show
that ships via the v3.5 path gets logged for the eventual v4 sweep
instead of being silently forgotten.
Sassy run log gets a STOP-GAP banner at the top pointing to the new
worklist. README.md gets a stop-gap-exception note alongside the
STYLE.md hard-prereq paragraph. ROADMAP H5 now points at the worklist
file as the canonical source of which shows v4 needs to regenerate.
Adds lib/sub-yt-fetch.sh (yt-dlp wrapper) and lib/yt-clean.py (collapses
YouTube's rolling-window auto-caption VTT into a flat SRT). For shows
distributed YouTube-first that have no community subs anywhere -- verified
via three parallel research agents covering OpenSubtitles REST, OS legacy,
Addic7ed, SubDL, SubSource, and Podnapisi for the 5 niche shows in the
library, plus a price-vs-coverage analysis of OpenSubtitles VIP.
Findings: OS VIP would not have helped on the niche shows (it is
download-cap relief, not coverage unlock; same catalog as free). All 4
Jarrad Wright shows in the library (Sassy, Big Lez Saga, Donny &
Clarence, Mike Nolan) live on the same channel and have only YouTube
auto-CC available. v3.5 ships those, explicitly violating STYLE.md
'best quality' as a tracked stop-gap.
Sassy the Sasquatch S01 5/5 episodes subbed with cleaned auto-CC. Mike
Nolan special-case noted: a 'COMPLETE SEASON | SUBTITLES' YT upload from
Oct 2025 carries hand-typed CCs and should be preferred over per-episode
auto-CC when subbing that show.
ROADMAP H5 added: v4 WhisperX large-v3 on the friend RTX 4080 node will
regenerate the v3.5 stop-gap with proper-noun-prompted transcription
(~4-6%% WER vs ~12%% YT auto-CC) and restore the STYLE.md quality bar.
H1 OpenSubtitles credentials marked done (was completed 2026-05-09).
Stock Jellyfin paints .listItem.selected and .focused in cyan
(#00a4dc) on actionSheet/selectionList/dialogContainer dropdowns.
Clashes with Cineplex red on audio + subtitle pickers.
Pick: variant 04 "Hairline ring" — 1px #E50914 outline (offset -1px
inside row) + dark near-black bg + white text. Architectural,
quietly on-brand. Applied via body.arrflix-themed scope.
Saved alternative: web-overrides/skins/selector-variant-02-red-
underline.css — variant 02 "Red underline" (matches search-input
focus). Future skin/swap option per owner. Not currently active.
Reconciled prod drift: prod overlay had been externally modified
since last snapshot (md5 2da61583 -> c62898). Pulled prod's current
overlay as new repo baseline + snapshot. Then applied variant 04 on
top -> dev md5 0ab8b258 (= prod + v4). Prod still c62898 untouched.
Codifies the bar every fetch must hit:
- one plain English .srt per episode, named <base>.eng.srt
- drop SDH / HearingImpaired / MachineTranslated / AiTranslated / Forced
- prefer fps-matched, then highest download count
- UI label collapsed to just "English" via web-overrides shim
README.md now points at STYLE.md as a hard prereq before any fetch run.
The picker logic in v1/v2/v3 helpers already encodes these rules; STYLE.md
is the canonical source of truth that they should be checked against if
anyone (or future-me) is tempted to relax them.
Stock Jellyfin renders subtitle stream entries as
"<lang> - <CODEC> - (External|Internal|Embedded)[ - flag]". For ARRFLIX,
which has at most one subtitle format per language, the codec and source
suffix are noise. Add a small JS shim to web-overrides/index.html that
matches that exact shape and collapses the label to "<lang>" (with
"(Forced)" / "(SDH)" / "(Hearing Impaired)" if those flags are
present; "Default" is dropped since it's redundant when there's only one
stream of that language).
Audio labels like "5.1Ch Surround Sound - English - AAC - Stereo - Default"
have a different number of segments and don't match, so they pass through
untouched.
The shim runs after DOMContentLoaded and re-walks any nodes added later
via MutationObserver (covers actionsheet dropdowns that mount lazily).
Bracketed by /* SUB-LABEL-SHIM-BEGIN */ and /* SUB-LABEL-SHIM-END */
markers; bin/revert-sub-label-shim.sh deletes between the markers in one
sed pass and saves a timestamped backup. No container restart needed
(index.html is bind-mounted).
Adds lib/sub-a7d-fetch.py: free, no-daily-cap path via subliminal's
addic7ed provider (anonymous). Uses OpenSubtitles REST search-only (no
quota cost) to translate library S/E to the show's primary catalogue
numbering, then drives subliminal to download from Addic7ed and writes
sidecars direct to nullstone via SSH.
Picker quirks: subliminal series-name matcher is broken by '!' in the
title, so the script strips it before building the synthetic
Video.fromname() string. OS feature_details S/E happens to align with
Addic7ed's indexing for the test show (American Dad).
Recipe README now reflects three paths in cheapest-first order: v3
Addic7ed, v2 OS REST (20/day), v1 plugin. American Dad run log updated
to 49/58 (S01 7/7 v1, S02 16/16 mixed v2/v3, S03 16/19 v3, S04 10/16
v3). 9 misses identified, deferred to next OS REST quota window.
Adds lib/sub-rest-fetch.py: direct OpenSubtitles REST, looks up subs by
per-episode IMDB id (e.g. tt0511631) instead of the plugin's
(parent_imdb_id, season, episode) combo path. This sidesteps shows where
library numbering diverges from OpenSubtitles' catalogued numbering --
American Dad uses Hulu S1=7 eps; OS uses Fox S1=23 eps; the plugin path
returns 0 hits past S01E07 even though every per-episode IMDB id is
correct.
Recipe README updated to surface the two paths (v1 plugin / v2 REST) and
recommend v2 by default. American Dad run log now shows 19/58 episodes
subbed (S01 7/7 via v1, S02E01-E12 via v2). S02E13-S04 (39 eps) deferred
to next 20/day quota windows.
Quirk fixed in v2: OpenSubtitles /download endpoint consistently returns
HTTP 503 to Python urllib.request despite identical headers/body via curl.
_curl() shim routes all OS API calls through curl. Each 503 still
consumes a download slot, so urllib path was unsafe to retry on.
Adds processes/ umbrella for repeatable acquisition workflows. First child
is subtitles/, with recipe README (executable by Claude Code), CHANGELOG,
per-show run logs, and a tested helper at lib/sub-fetch.sh.
Run on American Dad: S01 (7 eps) passed, S02-S04 (51 eps) broke. Library
uses Hulu/DSP season ordering; OpenSubtitles indexes by Fox airing order;
plugin queries by (parent_imdb_id, season, episode) so library S02E01
returns 0 hits. v2 design = direct OpenSubtitles REST with per-episode
imdb_id lookup; pending API-key registration.
Image-12 incident: I'd set <video> z-index:9999, which covers the OSD
scrubber + buttons (Jellyfin's stock OSD controls live at z-index
1100-2000, above the 1000 of .videoPlayerContainer but BELOW our
9999). Drop the lift entirely. Stock z-index hierarchy already has
controls floating on top. The fix for black-screen was always
transparent ancestor backgrounds (L1+L2), never z-index.
Reorganized inject-middle-theme.py CSS string from one-line dense
concat into a triple-quoted multi-line block with header comments
explaining each section + the layer model + DO-NOT rules. Same
output bytes (verified md5 deterministic). Added long-form comment
header to JS too.
Doc 31 (new): "Theme layer model + edit guide" — comprehensive
checklist for any future CSS edit. Covers:
- Stacking order layer 0..7 with stock Jellyfin z-indexes
- The two body classes (.arrflix-themed, .arrflix-video-active)
- Specificity tiers + cascade order (L1 vs L2)
- CSS load order (inline < bundle < branding.xml)
- Recurring bug list (6 incidents now, all same anti-pattern)
- DO NOT DO foot-gun list
- 4-step smoke verify procedure
- CI gates still TODO
Snapshot bumped to md5 2da61583. Prod+dev byte-identical.
Two bugs in prior video-isolation fix (452ce68):
1. .htmlVideoPlayer wrong class — Jellyfin actually uses
.videoPlayerContainer + <video class="htmlvideoplayer"> (lowercase).
:has(.htmlVideoPlayer) never matched, so L1 :not(:has) was always
true — body never excluded from #000 paint, L2 (lower specificity
0,2,1) lost to L1 (0,3,1). Replaced :has() gate with :not(.arrflix
-video-active) — JS body class is reliable signal.
2. html background-color stayed transparent on details/video pages
despite 5 stylesheet rules saying #000 !important. Headless Chrome
reported computed bg as rgba(0,0,0,0) — root canvas propagation
bug or cascade quirk. Fix: JS shim sets inline style on
<html> with !important. Inline styles win against any stylesheet.
Result: html paints #000 (so letterbox bars are black not white),
body transparent during playback (L2 wins on equal specificity via
source order), videoPlayerContainer + <video> transparent → frames
visible. Off-video: body opaque #000 (L1 fires).
Verified DOM probe: html bg rgb(0,0,0), body rgba(0,0,0,0) when
arrflix-video-active. Both flip when class removed.
Two-layer defense for the recurring "black screen during playback"
bug class (5+ occurrences in 24h per doc 26/28/30):
L1 (prevention): scope every black-bg rule with
:not(:has(.htmlVideoPlayer)):not(:has(#videoOsdPage)) so the rules
self-disable while a player is in the DOM. Covers body,
#reactRoot, .skinBody, .backgroundContainer, .mainAnimatedPage,
.mainAnimatedPages, .pageContainer.
L2 (override): when JS-toggled body.arrflix-video-active is set,
high-specificity (0,3,2 + tag) transparent rule wins against any
ancestor opaque-bg rule (including future regressions someone adds
without scoping). Covers all known wrappers, the
videoPlayerContainer + videoPlayerContainer-onTop, #videoOsdPage,
.libraryPage, .htmlVideoPlayer.
L3 (z-index lift): force .htmlVideoPlayer + child <video> to
z-index:9999 + isolation:isolate during playback so it floats above
any opaque ancestor that still leaks through.
Verified in playwright: with arrflix-video-active + .htmlVideoPlayer
mounted, all 7 ancestors return rgba(0,0,0,0). Without — all 7
return rgb(0,0,0). Self-disabling works.
Lesson reinforced (doc 30 roadmap open): add darkPct assertion to
bin/headless-test-v2.py + xmllint CI gate. Five recurrences without
those gates says we keep relearning this. TODO next.
Grey #101010 stripe at bottom of pages: jellyfin-web theme.css:44-50
sets .backgroundContainer + html to #101010, neither Cineplex nor
prior overlay overrode it. Added rule forcing #000 across html, body,
.backgroundContainer, .skinBody, .mainAnimatedPages, .pageContainer,
#reactRoot under body.arrflix-themed.
Search input: stock cyan focus ring (#00a4dc, theme.css:262-272)
swapped for borderless slab + 2px red bottom on focus + soft red
box-shadow halo. Cineplex/Netflix-faithful.
Movies/Series nav: variant E "cinematic glow" picked from preview.
Active state = red text + text-shadow halo + font-weight 700. JS
hash matcher toggles .arrflix-nav.active when location.hash matches
#/movies.html or #/tv.html (and short forms). hashchange listener +
existing setInterval keep state in sync.
My Media row: .homePage .homeSectionsContainer .verticalSection.section0
hidden — matches prod which had it hidden via different mechanism.
Snapshot at snapshots/2026-05-09-v6-stable/index.html bumped to
md5 2c8f5d5f7c99611fa93d15c34fbe35d1; same as prod and dev.
Stock Jellyfin shows .headerBackButton on library pages
(Movies/Series). With our nav links already in headerLeft, the back
arrow is redundant clutter and confuses the home button intent.
Add .headerBackButton to the existing display:none rule under
body.arrflix-themed. Verified visually on dev (md5 c99aca0f), then
shipped to prod with overlay swap (md5 364cc890 -> c99aca0f). Both
sides byte-identical.
Snapshot at snapshots/2026-05-09-v6-stable/index.html updated.
Owner pronounced "near perfect". Save current state as the rollback
target. Replace older 2026-05-08-pre-elegantfin snapshot.
Snapshot md5 364cc890c58f02d07cf50b43b31a48f0 — matches both prod
and dev deployed overlay.
Doc 30 lists every file/path-of-record + rollback procedure +
remaining roadmap items.
Tag this commit v6-stable-2026-05-09 after push.
Favicon: prod's older lockFavicon() shim was clobbering our injected
A-logo <link> tags every head mutation. Tag our links with
data-arrflix-icon="A" + add a hijack IIFE that re-pins the A URL on
matching tags AND removes any other large data:image/png link tags.
Tonemap: encoding.xml flipped EnableTonemapping false to true on dev
+ prod (server-side, not in repo). Doc 21 documented this fix
2026-05-08; prod was still grey-washing HDR10 sources because
setparams was relabeling PQ pixels as bt709 without zscale + tonemap
conversion. API now reports EnableTonemapping=True. Next HDR10
transcode gets the proper zscale -> tonemap -> format ffmpeg chain.
Both verified on dev first then promoted. Prod overlay md5 c6c85076
to 364cc890. dev and prod overlay byte-identical.
assets/screenshots/01-search.png — search + suggestions list
assets/screenshots/02-detail-mandalorian.png — Mandalorian detail w/ backdrop
assets/screenshots/03-playback-sassy.png — Sassy the Sasquatch playback
Embedded in README.md above 'What you get' so the repo landing page on
git.s8n.ru reads as a finished product rather than text only.
Add ARRFLIX wordmark center, Movies/Series nav left, search right,
favicon=A-mark, auth gate so login stays stock, hide on video page.
Side-effect of branding.xml escape (<video> → <video>): prod's
CustomCss block now actually loads, so the INC7 transparent-video
rule reaches the browser. /Branding/Css.css 0 B → 36 256 B; doc-28
black-screen issue closed at the delivery layer.
Markers: ARRFLIX-MIDDLE-THEME-BEGIN/END (style + script) and
ARRFLIX-FAVICON-BEGIN/END (link). Idempotent.
See docs/29 for design + deploy procedure + recovery quirk.
Agent 6 applied SW-pin fix and marked verified via element state
(currentTime advancing, videoWidth=1920, readyState=4). Headless pixel
histogram still showed darkPct=100% — element decoded fine but CSS
overlay covered it.
Real cause: branding.xml BLACK-PASS paints .libraryPage with
#000 !important. Jellyfin OSD page renders <div id=videoOsdPage
class=libraryPage>; class match -> opaque black div above <video>.
Fix: extend transparent-scope using :has(.htmlVideoPlayer) +
#videoOsdPage selector. Post-fix darkPct=9.8% (was 100%), MNS S1E4
video frame visually paints.
Removed INC6 clear-cache-only middleware (no longer needed, was
burning HTTP cache every visit).
bin/apply-26-incident-fixes.sh extended with INC7 patch (idempotent
re-apply if branding.xml ever drifts back).
Lesson: video-element state alone is insufficient verification.
Always sample pixel histogram + canvas drawImage on the painted
viewport.
Five sibling agents converged on root cause:
jellyfin-asset-immutable Traefik router (priority 90) was matching
/web/serviceworker.js (Jellyfin PWA's actual SW filename), pinning it
with Cache-Control: public, max-age=31536000, immutable. The
priority-100 jellyfin-html-nocache router only excluded the literal
path /web/sw.js, missing serviceworker.js.
Stale SWs from earlier ARRFLIX iterations intercepted /Videos/* and
/web/* fetch events, returning cached/empty bytes. Result:
MediaSource appendBuffer got bad data -> black <video>. INC6's
Clear-Site-Data: "cache" couldn't fix it (per MDN spec, "cache"
excludes SW registrations; "storage" would have worked).
Fix: added jellyfin-sw-nocache router at priority 250 in
/opt/docker/traefik/config/dynamic.yml on nullstone, forcing
cache-no-store@file on /web/serviceworker.js + /web/sw.js. Hot-reload
via Traefik file provider, no docker restart.
Verified at the wire (curl -I /web/serviceworker.js now returns
no-cache, no-store, must-revalidate; main.jellyfin.bundle.js still
immutable as intended) and via headless Chromium probe of MNS S1E4
(33s of currentTime advance, readyState 4, videoWidth 1920x1080,
no errors, both s8n admin and USER-F user).
bin/prod-vs-dev-compare.py also lands as a one-shot diff helper used
during the investigation.
INC5 fmp4-disable shim required browser hard-reload to fire. Owner's
MNS S1E4 re-test still black-screened because cached index.html ran
old shim + fmp4-HLS bug recurred. Add Traefik response header
'Clear-Site-Data: "cache"' on /web/index.html. Cache-only is safe
(no cookie/storage wipe -> auth + localStorage preserved). One fresh
visit refetches index.html with new shim. Remove middleware after
owner confirms working, otherwise every revisit re-flushes cache.
Symptom + root cause + fix + lesson for the scrollbar grey strip at
very bottom of page. Patch already in bin/apply-26-incident-fixes.sh
and live in branding.xml on prod from earlier commit; this is the
documentation subsection that was missing dedicated coverage.
Two parallel fixes for MNS S1E4 black-video bug. Belt+braces.
INC5 fmp4-disable (this agent):
- Add localStorage.setItem('enableHlsFmp4', 'false') shim to
web-overrides/index.html (idempotent, marker INC5 fmp4=false 2026-05-09)
- Forces TS segments instead of fMP4 for all HLS transcodes,
works around upstream black-video bug with HEVC+fMP4
(jellyfin-webos#126, jellyfin#16612)
- Browser localStorage verified false via headless playwright;
server confirmed emitting -hls_segment_type fmp4 before fix
- Repo + deployed file md5 match: 5b212d7d60b8a2b910a2f47dd0470a09
INC5 AV1 force-transcode (parallel agent):
- Re-encoded MNS S1E1-5 AV1->H.264 in container; PlaybackInfo
now returns DirectStream/DirectPlay=true on S1E4
- Doc additions covering the AV1 work included here since
same file (already authored by parallel agent, not yet committed)
Two regressions slipped through INC1-3:
INC4a -- BLACK BAND behind every detail-page carousel
Pre-existing 2026-05-08 home-page rule painted .emby-scroller {bg:#000
!important} UNSCOPED. Hits every carousel inside .itemDetailPage incl
admin-only More from Season N, More Like This. INC1-3 transparent-scope
list missed .emby-scroller / .verticalSection / .padded-top-focusscale.
Fixed by extending scope.
INC4b -- VIDEO 'BLACK SCREEN' on play
Not actually black-screen. CPU-only nullstone cannot sustain real-time
4K HEVC HDR tonemap+x264 transcode -- 0.5x realtime, ffmpeg takes ~6s
per 3s segment. With user resume seeks adding restart overhead, total
wait ~18s before browser readyState rises. User saw black, gave up.
Fix: disable EnableTonemapping (R&M fake HDR per doc 21) + cap
RemoteClientBitrateLimit=20Mbps on every user (1080p target, no 4K
scale). Headless v2 test confirms HEVC + AV1 episodes now hit
readyState=3/4 within wait window; 4K HDR R&M still slow (heaviest).
INC4 testing methodology audit -- bin/headless-test-v2.py
v1 only logged in as USER-F and never clicked Play. v2 runs both admin
and USER-F, walks 3 codec-tagged items per role (HEVC/AV1/H.264),
clicks Play, captures <video> state, sweeps DOM for opaque bgs over
backdrop layer. False positives: off-viewport #reactRoot + collapsed
.mainDrawer (negative coords). Allowlist refinement TODO.
Open: 4K HDR sources still slow even post-fix. Real fix path = pre-
transcode masters to 1080p H.264 SDR via separate batch, OR migrate to
10.11.8 with vaapi/qsv driver fixed.
After INC1 fixed the Abspielen + first-fold backdrop, owner reported black
band hiding artwork in More from Season 1 / below-fold sections. Two more
patches required:
INC2 — pin .backdropContainer + .backgroundContainer position:fixed; height
100vh so backdrop persists during scroll. Added vertical fade ::after.
INC3 — extend transparent-scope to ALL detail-page sub-sections
(.detailVerticalSection, .scrollSlider, .padded-bottom-page,
.itemsContainer etc) so section wrappers don't paint over the pinned
backdrop section by section.
bin/headless-test.py now takes top + scrolled viewport screenshots.
full_page=True hides position:fixed regressions, dual-screenshot exposes
them. Use both to bisect.
bin/apply-26-incident-fixes.sh updated with INC2+INC3.
Open: AV1+Opus playback (Mike Nolan Show) still tracked for 10.11.8
migration. .detailLogo regression possible — test in actual browser.
Symptoms: Page Unresponsive on poster grid, posters missing then black
backdrops, 'Abspielen' German Play button surviving Traefik+force-english
chases, video black-screen on play.
Root causes (different from initial guesses):
- Browser hangs: deployed index.html drifted ahead of repo; uncommitted
forceEnglishUI() text-walker MutationObserver froze main thread on poster
lazy-load. Reverted to repo HEAD.
- 'Abspielen': Cineplex theme HARDCODES German via 'content:' ::after rule
-- not a Jellyfin locale issue. Doc 25 already proved per-user UICulture
is theatre. Override CSS with content: 'Play'.
- Backdrops black: BLACK-PASS CustomCss block paints #000 !important on
.layout-desktop / .pageContainer -- occludes backdrop layer (z-index:-1).
Existing transparent-scope rule used body.itemDetailPage selector that
doesn't match in 10.10.3 (body class is libraryDocument). Replaced with
:has(.itemDetailPage) ancestor scoping.
- HLS 499: encoding.xml had EnableThrottling+EnableSegmentDeletion=true,
segments reaped before browser re-request. Disabled both.
Verified via new bin/headless-test.py (playwright Chromium login + screenshot
+ computed-style probe). Fixes idempotent and re-runnable via new
bin/apply-26-incident-fixes.sh.
Open: AV1+Opus items still black-screen in Chrome due to DirectStream
codec-tag mislabel bug. Tracked for 10.11.8 migration.
AudioLanguagePreference=eng, SubtitleLanguagePreference=eng,
SubtitleMode=Default, PlayDefaultAudioTrack=true, UICulture=en-US.
Per-user Configuration POST applied to all 9 existing users + wrapper
updated for future creations.
Server-runtime focus; supplements doc 13. Headline: 4 concurrent ffmpeg
processes for ONE viewer all transcoding 1080p->2160p with PGS subtitle
burn-in, on uncapped jellyfin container sharing 12-core host with
uncapped Forgejo BlueBuild CI runner (88-99 % CPU). Load avg 15.4 on 12
cores. Throttling+SegmentDeletion still off (doc 13 finding 03 now
non-optional). Top quick-win: enable transcode throttling + segment
deletion + cap RemoteClientBitrateLimit.
Edge audit complementing doc 13 (server-side perf). Confirms cold-load
"feels slow" perception is dominated by:
- no HTTP compression at Traefik (2.74 MiB raw JS bundles per cold load)
- no Cache-Control on hashed-asset URLs (28 conditional GETs per warm load)
- first-fetch poster image transcode ~385 ms (server-side, doc 13 #02)
TLS, MTU, HTTP/2, cert chain, middleware chain, Pi-hole hairpin all
audited and clean. Pi-hole missing local DNS rewrite for arrflix.s8n.ru
(LAN clients hairpin via WAN unless /etc/hosts pin in place).
Top quick win: add `compress@file` middleware in
/opt/docker/traefik/config/dynamic.yml + reference from Jellyfin router
label. ~70 % cold-load wire-size reduction (2.74 MiB to ~0.82 MiB
gzip / ~0.69 MiB brotli). One file edit, no architectural change.
No fixes applied. No state mutated. No Traefik reload.
