perf: move transcript restyle + saved-transcript analytics off the main actor

[r3dbars]

Why

Every transcript save ran MeetingTranscriptStyler.restyleTranscript synchronously on the main actor — a full transcript read, rewrite, and file + audio-dir rename — and savedTranscriptAnalyticsProperties re-read the entire transcript again on the main actor just to bucket analytics. Multi-hour meetings produce multi-MB markdown, so the UI hitched exactly at "transcript saved". (Audit-verified finding.)

Product Impact

• Affects: meetings
• Lane: meeting reliability
• Why this matters: removes a save-time UI freeze that scales with meeting length, and stops a parse regression from being able to silently destroy transcript bodies.

What changed

• MeetingSessionController: the taskManager.$lastSavedTranscriptURL subscription now runs the restyle on a serialized detached task (restyleSavedTranscriptInBackground); only the published state updates, live-sidecar attach, and diagnostics hop back to the main actor. Restyles are chained so two passes never touch the same artifacts concurrently (matters for the speaker-review re-save), and a stale-result guard drops superseded updates while retained-audio compression still runs for every save.
• savedTranscriptAnalyticsProperties is now nonisolated static and runs detached. It awaits the in-flight restyle so it reads the renamed artifact — previously it read the pre-rename path, which the synchronous restyle had often already renamed away, silently producing empty analytics buckets — and it uses the bounded TranscriptFrontmatter.readValues(from:) chunk reader instead of loading the whole file. The AnalyticsReporter.track call itself stays on the main actor (the reporter's lazy distinctID is not thread-safe).
• MeetingTranscriptStyler: persisting now fails closed when the body carries transcript text the entry parser cannot understand — no rewrite, no rename, logged as meeting_transcript_restyle_skipped — instead of destructively replacing the body with _No transcript captured._. Genuinely empty transcripts (empty block, or only ---/*Generated by... remnants, matching the real saver's empty shape) still get the placeholder rewrite.
• Tests: two new fast tests pin the fail-closed and genuinely-empty behaviors; the source-slice anchor in RepoCommandContractTests follows the renamed analytics function. Sources/Meeting/CLAUDE.md flow step 12 updated.

How I checked it

• Selected checks from .agents/test-matrix.yml for the files changed (Sources/Meeting/** rule)
• bash build-deps.sh --force
• bash build.sh --no-open
• bash run-tests.sh — 4660/4660 passed, including MeetingTranscriptStylerTests and the repo-contract suite
• bash run-integration-smoke.sh
• Manual check: n/a — behavior covered by fast tests; no UI surface changed

Risk Review

• Privacy / local-first behavior reviewed — same meeting_transcript_saved event and allowlisted bucketed properties; only timing/threading changed
• Storage path or migration impact reviewed — none; restyle output and rename behavior unchanged for parseable transcripts
• No private transcripts, audio, tokens, personal paths, or customer data are included

Notes

Known trade-off: if a second save lands while the first save's restyle is still in flight, the first meeting's live-Codex final-transcript attach can be skipped — but the attach guard (lastSavedTranscriptTaskId == awaitedJobID) would have rejected it at that point anyway, and the activeCount-driven attach path provides a second chance before that window closes.

Side effect worth knowing: because the restyle no longer runs inside the @Published willSet, populateSavedMetadata in TranscriptionTaskManager now reads the transcript before any rename, so lastSavedTitle/lastSavedDuration/lastSavedSpeakerCount populate correctly in cases where they previously read a just-renamed-away path.

🤖 Generated with Claude Code