feat(meeting): always-on cheap field extraction at save time

[r3dbars]

What & why

From docs/NEXT_WORK.md #5. The heavy local meeting summarizer (LocalMeetingSummarizer.swift — the ~12GB Gemma / Apple beta path) is the only thing that writes the structured fields Decisions / Action Items / Open Questions. It only runs when a user opts into the beta, so the "ask my history" moat (the search index) only ever covered a subset of meetings.

This adds an always-on, cheap, dependency-free extraction at save time so the index can cover 100% of meetings (live + imported), not just the beta opt-in. The heavy summarizer stays the high-quality path; this is the baseline that guarantees coverage.

Method

A precision-leaning heuristic pass (no model, no 12GB dependency):

• MeetingQuickSummaryExtractor parses the already-styled transcript into speaker turns, splits into sentences, and classifies each into at most one bucket (Decisions › Action Items › Open Questions) via curated cue lists, with small-talk filtering, dedupe, and per-section caps. Action items are owner-prefixed from the speaker label. Produces a LocalMeetingSummarySections (same shape the heavy path uses).
• MeetingQuickSummaryWriter writes those into the saved transcript's YAML frontmatter under a parallel auto_summary_* namespace. It's idempotent (skips once auto_summary_version is set), skips non-meetings, and is frontmatter-only (transcript body preserved verbatim).

Why a separate auto_summary_* namespace (not local_summary_*)

The heavy summarizer's local_summary_* keys also drive Home UI gating — the inline summary card and the "Run AI summary" affordance key off local_summary_version. Reusing them would make the cheap heuristic masquerade as the AI summary and hide the upgrade affordance. The parallel namespace means: the index gets the same logical fields for every meeting, the heavy summarizer overwrites nothing, and existing Home/Settings UI is untouched.

Wiring

Runs in MeetingSessionController.restyleSavedTranscriptInBackground, immediately after restyleTranscript on the same chained detached task — so it's off the main actor, never races the next restyle, and the body is already in canonical styled form. Covers both live captures and imported audio (both flow through taskManager.lastSavedTranscriptURL).

Sequence vs Moat #1 (index summary fields)

Moat #1 is not merged in this branch. This PR builds to the same field shape so #1's indexer can consume it directly. Field keys + value format (bullet lines flattened with " | ", mirroring local_summary_*):

auto_summary_version, auto_summary_generated_at, auto_summary_method (heuristic-v1), auto_summary_participants, auto_summary, auto_summary_decisions, auto_summary_action_items, auto_summary_open_questions, auto_summary_risks_or_followups, auto_summary_accuracy_notes.

Index precedence the #1 reader should adopt: prefer local_summary_* when present (heavy, higher quality), else fall back to auto_summary_*. Until #1 (or #4's cross-meeting tools) reads these keys, the fields are written but not yet queried — that's the intended ordering, not a regression.

Tests

• New Tests/MeetingQuickSummaryExtractorTests.swift (registered in Tests/FastTests.manifest + run-tests.sh APP_SOURCES): owner-prefixed action items, decision/action de-conflation, substantive vs rhetorical questions, tiny/empty input, inline transcript form, the frontmatter writer (keys present, body preserved, no local_summary_* leakage), and idempotency / non-meeting skip.
• bash build.sh --no-open ✅
• bash run-tests.sh ✅ (10177/10177)
• bash run-integration-smoke.sh ✅
• Independent codex-reviewer pass: PASS, no P1/P2 in the diff (only flagged the expected Moat Comprehensive Bug Fixes: State Machine, Error Handling, and Resilience #1 reader gap).

🤖 Generated with Claude Code

[r3dbars]

Closing as superseded by #1331, which merged the combined ask-meeting-history path covering this draft's scope.