fix(meeting): gate transcript-idle close on ASR partials, not just finals

[WingedGuardian]

Why

The #38 transcript-idle close keyed on the committed-turn count. Day-one real use dropped ~15 min of a live meeting: on quiet/far-field audio Speechmatics can go many minutes without committing a FINAL, so a live meeting read as "idle", the session closed, and the dormant state (correctly refusing to reopen on room noise) kept it closed until a quiet gap finally re-armed it.

What

Evidence-based idle signal, verified against the live Speechmatics API:

• ambient_bridge/active_session.py (additive; ambient path unaffected — nothing there reads the new surface): last_activity (monotonic) stamps ASR speech evidence. Two live-API behaviors shape the predicate, both observed with a direct probe:
  • Speechmatics re-emits the same trailing partial for as long as audio flows → a partial counts only when non-empty AND changed vs the previous text.
  • Speechmatics emits empty finals every ~2s (commit heartbeat) while audio flows → a final counts only when it actually commits a turn (len(acc.committed) grew).
• meeting_bridge/server.py: the idle close now tracks last_evidence_ts fed by last_activity (primary) with the turns count as fallback for backends without it; both absent ⇒ idle-close disabled for that backend. Seeded at session open; dormant/re-arm machinery unchanged.
• MeetingSession protocol documents the optional liveness attributes.
• Tests: 3 ambient evidence tests (frozen partials ≠ evidence, empty finals ≠ evidence, committing final = evidence + tracker reset), meeting lifecycle test for partial-evidence keep-open; 50 meeting + 13 ambient pass.
• Docs: config comment, README, meeting.env.example, CONTRACTS §1c.

Verification

E2E against real Speechmatics on the edge (isolated instance, synthesized-speech PCM):

• Leg A (the fixed failure): 30s of real speech (2.5× a 12s idle window) → 0 idle closes; transcript verbatim-correct.
• Leg B (the fix(meeting): close a session on transcript idle, not just mic silence #38 win intact): loud tone (above-threshold energy, no speech) → idle-close at 15.3s, then dormant: continued tone opened no new session.
• Leg C: 5s silence re-armed; speech opened session fix(s2s): age cached context from the last real turn, not last reconnect #2.

Direct-probe transcript of the live API (empty finals every ~2s during tone; frozen partial re-emission) is what drove the predicate — both behaviors are pinned by unit tests.

Reviewed twice: full review (DONE, no findings) + focused delta re-review of the post-E2E predicate changes.

Rollout

Edge currently runs with MEETING_TRANSCRIPT_IDLE_CLOSE_S=0 (idle-close disabled as the stop-the-bleed mitigation). After this deploys, the env override is removed to re-enable at the 300s default on the new signal.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0b7b77c6b5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Count merged finals as ASR activity

When Speechmatics sends another final for the same speaker, TranscriptAccumulator.add_final() merges it into the existing committed run instead of appending a new entry, so len(self._acc.committed) does not change even though new transcript text was committed. In that single-speaker/merged-turn case, neither last_activity nor the turns fallback advances, and the meeting bridge can hit MEETING_TRANSCRIPT_IDLE_CLOSE_S and finalize/dormant a still-active session if partials are absent or not changing. Track whether add_final actually added text/results rather than using only the committed-list length.

Useful? React with 👍 / 👎.