fix(radio): Keep recovery from reporting aborted pending loads

Merged
FC
15d ago
+127-5

Summary

Recovery's watchdog reload calls loadAttachedSource without bumping generation. That cancelLoad()-aborts a still-pending waitUntilReady, and the same-generation AbortError: Load aborted was reported as an unexpected PLAY_ERROR (RADIO-27 on radio@0.7.2, Firefox, single-mode Icecast).

 recovery reload while load() still pending
-  cancelLoad → AbortError → onError / PLAY_ERROR
+  attachment epoch supersedes the pending waiter
+  settle without onError; recovery owns resume
 real media errors during pending load
-  unchanged: still report through onError

Fixes RADIO-27
https://avoidquest.sentry.io/issues/RADIO-27

Evidence

  • Root cause: scheduleReload → reload → loadAttachedSource → attachment.load → cancelLoad rejects the prior same-generation waitUntilReady with AbortError: Load aborted. MediaElementPlaybackSource.load() treated that as terminal because generation was unchanged.
  • Not covered by #457 / RADIO-23/24 on main (7f7ed3b): #457 ignores queued media-element abort events and scopes intentional play() supersession for pause/stop/cleanup/load(). It does not treat recovery's controller abort of a pending same-generation attachment as expected cancellation.
  • Red → green (TDD):
    • Red (before source change): recovery reload canceling a pending load does not report AbortError failed because onError received AbortError / Load aborted.
    • Green (after): that test plus a real media error during a pending load still reports onError and the existing recovery / RADIO-23 suites pass (73 tests across playback-source + audio-manager).
  • Validation: bun run check, bun run typecheck, bun run --filter @avoid.quest/radio build, and focused bun test on media-element-playback-source.test.ts + audio-manager.test.ts pass.
  • Not verified: physical Firefox against a live Icecast stream; production deploy; Sentry resolution (no Sentry mutation).

Overlap / composability

PRScopeShared filesConflict risk
#457 (merged)Queued media abort + intentional play supersession (RADIO-23/24 station-switch path)same playback files historicallyAlready on base 7f7ed3b
#475 (open)Recovery reattach supersedes in-flight play intent; resume via playForIntent (Chrome play() interrupted by new load)media-element-playback-source.ts / .test.tsLow / two-way door
This PRRecovery cancel of pending load / waitUntilReady AbortError reporting (RADIO-27)same two filesLow / two-way door

#475 does not change pending-load cancellation or onError reporting. This PR does not add supersedeInFlightPlay, does not re-route recovery resume through playForIntent, and does not duplicate that play-intent fix. The only adjacent play() change here is: after a superseded pending load() settles, if recovery is still reloading, leave resume to recovery instead of racing a second play(). That is load-generation ownership, not play-intent invalidation.

Expected merge: either order works. Text conflicts are likely around loadAttachedSource (this PR) vs recovery reload/resume wrappers (#475) and separately added tests in the same describe block.

Merge Danger

Door: two-way

Code-only playback ownership change; no migration or deploy.

Blast Radius: playback

Shared media-element recovery for Single, Node, and DJ. Genuine pending-load media errors remain visible; only same-generation recovery cancellations of superseded attachments are settled.

Open in Web Open in Cursor 

Summary by cubic

Fixes recovery's reload canceling a still-pending load so it no longer reports a false PLAY_ERROR (RADIO-27). Tracks an attachment epoch so a superseded pending load settles silently and recovery owns resume, while genuine media errors during a pending load still reach onError.