fix(playlist): reconcile fallback lifecycle exactly once

[nicoechaniz]

Outcome

Hardens the already-deployed playlist fallback decision with snapshot-based, exact-once lifecycle reconciliation.

• derives availability from the complete current LiveKit publication snapshot;
• emits a transition only when availability actually changes;
• makes duplicate publish/unmute and mute/unpublish events idempotent;
• preserves suppression when an obsolete publication disappears but another live audio publication remains;
• ignores a stale ParticipantDisconnected when replacement beacon01 already has live audio;
• reconstructs both sides of the decision after reconnect.

Audio boundary

This PR does not change codec, bitrate, sample rate, channel count, PCM, gain, fade duration/equation, buffers, source selection or routing. It remains draft and requires Nico's explicit audio review before merge. No deploy is requested.

Evidence

• npm test --prefix services/playlist-bot: 13/13 pass.
• npm run build --prefix services/playlist-bot: pass.
• git diff --check: pass.
• Full repository CI rerunning on head 1922042; the prior head passed browser E2E and frozen-audio.

Remaining #39 acceptance

This does not close #39. A controlled LiveKit-native rehearsal and a truthful signal for “neither source is producing audio” remain outstanding.

Risk / rollback

The runtime path still calls the existing setBeaconAudioAvailable and onBeaconAvailabilityChanged; the helper only suppresses duplicate/stale transitions. Revert the PR merge commit to roll back.

Refs #39.

[nicoechaniz]

Adversarial follow-up 1922042: ParticipantDisconnected no longer forces playlist availability to false from a stale event. It reconciles the current LiveKit snapshot, so a replaced beacon01 with a live unmuted audio publication cannot be covered by a transient fallback fade-in. Added the exact replacement test; playlist-bot now passes 13/13 tests and TypeScript build locally. PR remains draft + audio-touching; no merge/deploy without explicit Nico review.

[nicoechaniz]

LiveKit-native evidence added in 9ef9b28 without changing runtime or acoustic behavior.

The opt-in integration test connects real @livekit/rtc-node 0.13.32 participants to disposable livekit/livekit-server:v1.13.4 and verifies the actual remote snapshot across: identity present without track → publish → server mute → unmute → unpublish → same-identity replacement. Result: 14/14 playlist-bot tests pass (0 skipped under the opt-in environment); TypeScript build passes.

The first controlled attempt also confirmed the production-safe LiveKit default: remote unmute is rejected. The successful lab rerun enabled room.enable_remote_unmute only on the disposable server so the SDK's TrackUnmuted representation could be exercised; this does not propose enabling remote unmute in production.

The local container was stopped and auto-removed. PR remains draft/audio-guarded: no merge or deploy requested. Remaining #39 gap is the explicit health signal for “neither source produces frames” plus audible rehearsal/approval.

[nicoechaniz]

Follow-up f731485: fixes the sole CI lint finding (prefer-const) in the opt-in native test. The behavioral test evidence is unchanged; fresh gates are running.

[nicoechaniz]

Fresh head f731485 is fully green: frozen-audio 17s, lint/build 1m24, unit gate 1m48 and complete browser E2E 10m35. The PR remains draft/audio-guarded pending explicit acoustic approval and the remaining frame-health/audible rehearsal; no merge or deploy.

[nicoechaniz]

Fast Forward checkpoint policy: the newly triggered full E2E workflow was intentionally canceled while this remains an audio-guarded draft. Focused playlist lifecycle tests/build/diff-check are green and the earlier unchanged native LiveKit lifecycle evidence is preserved; unit/lint-build/audio-boundary continue. Full E2E must be explicitly rerun and pass at the approval checkpoint before merge.

[nicoechaniz]

Adversarial re-review on current head 99a2a4e: no new blocker found. Every relevant LiveKit event now derives from the room publication snapshot, the final state transition remains idempotent at both helper and setter boundaries, stale disconnect/unpublish cannot override a live replacement, and reconnect reconstructs rather than replays event history. The important limitation remains explicit: an unmuted publication is not proof that non-silent frames are flowing. That acceptance belongs to health + audible rehearsal, so the draft must not be promoted or merged yet.