Auto-link reliability: punctuation-blind name matching and an event-driven transcript wait

[nicw]

Closes two auto-link gaps, one per commit.

1. Punctuation-blind meeting-name matching. Filename sanitization drops / and quotes, so a calendar subject like "Platform / Data Weekly Sync" can never equal its transcript/note title under the current normalizer (whitespace collapse only) — those meetings sit in the unlinked list forever unless a session guid links them. normalizeMeetingName moves to utils/meetingName.ts and now strips all non-alphanumerics (Unicode-aware, so non-ASCII names don't collapse below the too-generic length guard). This still matters with guid matching in place: the name matcher is the only recovery for un-stamped/legacy transcripts and for guid-stamped notes that were deleted and recreated.

2. Event-driven transcript wait. The recording service's post-processing can land the final export minutes after capture ends. Today the post-poll file wait is 15×1 s, and the status-poll give-up paths return "Transcript not ready" without ever checking disk again — the session then only recovers on the next reload (persistence) or unlinked-list refresh. This PR adds a TranscriptWait module: the file wait re-arms on metadataCache "changed" events (30-min cap, 30-s belt recheck, cancelled immediately on plugin unload). The session-guid match ladder is preserved intact as the wait's findReady probe (rung order unchanged; the newest-file rung additionally rejects candidates already linked to another note), the status poll is skipped when a finalized transcript is already on disk (a reconcile/watch hand-off can arrive after /status describes a different meeting), and the give-up paths now fall through to the event wait. The terminal "failed" path still removes persistence and does not fall through.

Deliberately not ported from my local branch, because upstream already covers them (as well or better): watch-loop /status failure tolerance, the reconcileRecordingLock hand-off, and a cross-link guard superseded by the namePrefix scan restriction + guid rejection.

Tests: 16 (9 name normalization, 7 transcript wait), plus an npm test script (no new deps).

Notes for review:

• Behavioral change to weigh: the give-up paths falling through extend a session tail's worst case from ~5 min to ~35 min (cancelable on unload). If you'd rather keep the guid-stall path as a hard return — the 2 h transcribe backstop already covers honest long jobs — that's a one-line revert on the legacy path only.
• Pairs with Defer speaker tagging until Tome finalizes the transcript body #2: once both land, findReady should also gate acceptance on hasLiveLegLabels. Kept out of this PR to avoid overlapping diffs.
• Sibling PRs Add opt-in setting to strip Teams/Zoom join blocks from new note bodies #1 and Defer speaker tagging until Tome finalizes the transcript body #2 also add test files; whichever lands later gets a trivial package.json test-list union conflict — I'll rebase.

Tested on this branch against v0.8.7: npm run build, npm run lint, npm test (16/16).

🤖 Generated with Claude Code