Recover web audio when a stream drops mid-round - #2404

Problem

A web listener's <audio> element had no recovery path. An error, a truncated body, or a silent stall meant silence for the rest of the round, with no affordance for the viewer.

This was reported live from a prod game. nginx access logs showed a complete body delivered for every round of that session (and no ffmpeg failures), so playback died browser-side after the bytes arrived — exactly the case nothing handled.

Change

Re-fetching a stream URL is always safe: the server seeks each GET to the live position. So every failure mode now takes the same path — re-request and resume in sync with the room.

Recovery triggers:

• error
• ended landing earlier than the playback should have finished (truncated body)
• the stalled event
• a decode-progress watchdog — nothing decoded for 5s while the element believes it is playing (this is the case that fires no event at all)
• online, and the tab becoming visible again

Bounded so a genuinely dead playback can't hammer the server: retries stop at the playback's own deadline (playbackDurationSec is now passed into handleRoundAudio), after 4 backed-off attempts, or after 6 total fetches per playback. Past that, stalled surfaces and the sound widget offers a manual "Sound stopped — tap to resume" (activity.web.sound.retry, filled in for all locales). Retries append ?a=<n> so the media cache can't replay the dead response.

Also fixes a misclassification: any play() rejection was treated as an autoplay block, showing an "Enable sound" pill that wouldn't help and blocking auto-recovery. Only NotAllowedError is a permission problem now — everything else is a dropout.

Server side, stream starts (with seek / remaining), 404 and 410 lookups, and clients disconnecting mid-body are now logged. Previously nothing was logged unless ffmpeg itself failed, which is why the original report had no server-side trail.

Verification

Driven in headless Chrome against a fake audio server, in real wall-clock time:

• Stream hangs mid-body (3s of audio, socket held open, no error/ended — the reported symptom): previously permanent silence; now re-fetches at +3.9s and plays through to the end.
• Permanently broken stream: stops after 6 fetches and raises stalled, so the manual pill appears.
• Healthy stream: single fetch, plays to the end, no spurious retries.

tsc (both projects), prettier, lint-ci, lint-i18n (0 warnings), vite build, and the 71 web unit tests all pass.