[Nawid3333]

Final review of the whole repo (2026-10-10): the findings that held up. Each fix has a test that fails on the old code.

Player

• A page that removed a player after it had linked up (a re-render) kept its own video hidden and paused until FastStream was turned off or the page navigated. The page now gets its video back (content.js; e2e in content-cleanup).
• Video delay: a resync that gave up left the spare audio player playing muted; and a speed changed while the delay was 0 never reached the kept audio players, so the sound drifted for good once the delay was set again.
• Subtitle sync: removing the track being synced, while another fine-time panel was on top, brought the sync timeline back for the removed track.
• Sources: the Play button stayed on "Loading" after a source failed.
• Audio track list: a video without audio levels kept the previous video's list.
• Bilibili: a video without sound (audio null) was never detected.
• DownloadEntry: a watcher removed while watchers were being told made the next one miss the news; a throwing progress handler stopped the rest.
• CSS: the shake animation's rotate(0eg) and font-weight: medium were invalid.

Background

• Toolbar cycle on an mpv-allowlisted page (MPV -> Off -> On): an On picked by hand went off again on the first page of the site the allowlist does not name; it now stays for the tab, like the plain toggle (your call). The MPV the list started still ends when the tab leaves the list (unit tests for both; toolbar-state/off-race/recover e2e pass; toolbar-cycle-mpv runs on CI's Windows shard).
• A blocked-header rule given up after failed removals left its id in Firefox, and every later rule with that id failed to add (remove + add in one call now).
• A tab closed right as the event page woke came back as a phantom tab state.
• StreamLengths' queue let a woken waiter race a new caller (5 probes instead of 4).
• A custom source pattern with an unknown type ("hsl") was accepted silently and matched nothing; it is reported now.
• application/ogg streams went to the MP4 player; they play directly now.
• Facebook: a batch whose first manifest was already reported dropped a new video after it.
• content.js: a refused fullscreen request left the fullscreen state on.

Cleanup and docs

• Removed the mixer/clip/channel helpers and ReferenceTypes.ANALYZER, left over from the audio settings window and the speech detector (no caller anywhere).
• README: "RAM and disk" FAQ (what each setting does, incl. the new disk switch); the old "browsers offload RAM to the SSD" sentence was wrong.
• CI: the MP4 fixture cache save may fail like the ffmpeg one; on Linux a failed pactl set-default-sink stops the job with an error instead of testing sound on the wrong sink; amo-signing-failsafe reports a failed upload as an error, not pending.
• content-overlays e2e: the fullscreen bar case (failed twice on main's Windows shard, then passed on a re-run) now says what held the bar up.

Not changed on purpose: FASTSTREAM_MPV_DEBUG=0 turns the host log on. A fix means a host change, which must raise the host version, and that tells every user to reinstall the host. It can go with the next real host change.

Checks: lint, tsc, vitest (2233 passed), build; e2e: content-cleanup (28), content-overlays (8), sources-browser-ui, video-delay, subtitle-sync.

🤖 Generated with Claude Code