Restore speaker review samples from retained meeting audio

[r3dbars]

Why

A speaker-naming failure deletes temporary clips, but Speakers review only looked for saved profile clips. A meeting with retained audio could therefore still show disabled sample buttons after failure. This adds playable recovery for those already-saved meetings.

The 1.1.55 stable-ID resolver also had a confirmed defect: decoding a strict 2 KB UTF-8 prefix could fail when the read split a multibyte character, even with the correct ID earlier in the file. #1686 already fixed that resolver, and subsequent changes on main restored pending-review retranscription and durable failure retries. The sanitized customer evidence cannot establish the exact original byte boundary. This PR preserves those fixes and covers the real save-to-finalization path instead of duplicating them.

Fixes #1681.

Product Impact

• Affects: meetings and speaker review
• Lane: meeting reliability
• Existing failed reviews can play short samples from retained audio without naming a profile or running transcription again.

What changed

• Resolve missing review samples by transcript speaker identity, timestamp, and audio channel, for both raw and styled transcripts.
• Stream at most eight seconds with AVPlayer, bounded by the next source turn and saved duration. Keep playback identity separate for voices sharing one recording; clear stale, stopped, and failed playback.
• Keep existing profile clips preferred. Recovery does not write global speaker samples, transcripts, or retained audio, and transcript rows remain static.
• Add a production-save regression with nine speakers, 441 system utterances, large Unicode frontmatter and a renamed transcript. Extend scanner recovery tests and exercise real muted WAV playback.

How I checked it

• Preflight, build source-list validation, shell syntax and diff checks
• bash build-deps.sh --force
• bash build.sh --no-open — isolated launch smoke and bundle performance gate passed
• bash run-tests.sh — 12,794 assertions passed
• bash run-integration-smoke.sh
• TRANSCRIPTED_DISABLE_FILE_LOGGER=1 swift test — 1,085 tests, 13 opt-in skips, zero failures
• Independent GPT-6 Astra / xhigh review of the full diff against origin/main (7c483f8c): no findings; no findings rejected. The additional beyond-EOF playback check passed on a real muted WAV.

Native UI click/screenshot verification is incomplete because the automation host lacks Accessibility/Event Posting permission. See review evidence. Synthetic tests do not reproduce the customer's recording or prove its diarization/audio quality.

Risk Review

• Read-only playback fallback; no storage migration, telemetry payload, release, or update-path change.
• No customer recordings or transcript contents are included.
• Draft PR for human review; no merge or release performed.

Agent handoff

COORD_DONE: GREEN | PR linked to #1681 | retained-audio speaker samples and regression coverage | no GitHub cleanup | human review; native UI click proof remains incomplete | mapped checks passed | lanes used: Codex=implementation plus independent Astra review; Claude=skipped, risk covered by independent review; Local=skipped, bounded Swift work; Windows=skipped, macOS playback task | review the draft PR