feat(mcp): index structured meeting-summary fields (decisions/action items/open questions)

[r3dbars]

What

Makes the structured summary fields (Decisions / Action Items / Open Questions) queryable from the Transcripted MCP index. Today the summarizer extracts these into each meeting's summary, but the MCP index (TranscriptIndex) only builds FTS over raw utterances + dictation — it never parses the summary. There was no decisions/action_items/open_questions table anywhere.

This is the indexing foundation only. The cross-meeting tools (list_action_items, etc.) are a separate thread that builds on listSummaryItems.

How

Parser (TranscriptedCaptureKit) — new CaptureSummaryParser → ParsedMeetingSummary. The app-target preview parser (RecentMeetingSummaryPreviewParser in Sources/UI/Shared/RecentCaptureScanners.swift) can't be imported across the module boundary, so the proven section logic is ported into the shared kit that the standalone tools already depend on. Handles both shapes:

• inline transcript summary (local_summary_version frontmatter + marker-bounded ## Local Summary body block with ### subsections, plus pipe-joined local_summary_* frontmatter fallback)
• legacy generated meeting_summary sidecar (# sections)

Action-item owner is extracted conservatively: a leading Title-Case name/team before ": " (≤3 words, ≤24 chars, no lowercase function words), with placeholders (unassigned, tbd, team, …) normalized to unassigned. Discuss the new API: … stays plain text.

Index (TranscriptedMCP)

• One normalized meeting_summary_items table (kind discriminator + owner + FTS5 + AI/AD triggers, mirroring the utterances pattern). A single table with a kind column rolls up trivially across meetings — designed so the future tools thread can GROUP BY owner / filter by kind without a per-category table.
• Parse + insert in indexMeeting (inside the existing transaction); delete in removeFromIndex.
• listSummaryItems(kind:owner:dateFrom:dateTo:) is the cross-meeting query foundation (owner: "" selects unassigned).
• Exclude *.summary.md sidecars from meeting enumeration — read as a summary fallback instead of being indexed as empty junk meetings.
• Schema user_version gate (v2): existing on-disk indexes rebuild once on upgrade, since unchanged transcript mtimes would otherwise skip backfill. (The live path writes the summary into the transcript, so re-summarizing bumps its mtime and reconcile catches it; the sidecar-only staleness is a documented legacy-path limitation.)

Tests

• CaptureSummaryParserTests (9) — inline/sidecar parsing, owner extraction + sentence-colon guard, placeholder normalization, None-found → empty, frontmatter fallback, nil cases.
• SummaryItemIndexTests (8) — parse→index→query, owner + unassigned rollup, date filter, newest-first ordering, reindex-replaces, delete-clears, sidecar-not-a-meeting, older-schema-forces-rebuild.

Verification

Per .agents/test-matrix.yml (touched TranscriptedCaptureKit/** + MCP):

• swift test --package-path Tools/TranscriptedCaptureKit → 34 pass
• swift test --package-path Tools/TranscriptedCLI → 45 pass
• swift test --package-path Tools/TranscriptedMCP → 82 pass
• bash run-e2e-smoke.sh → OK

Independent codex review: PASS, no P1. The two P2s it raised (backfill on upgrade; sidecar-only staleness) are addressed by the schema-version gate and documented, respectively.

Unblocks

Cross-meeting tools thread (list_action_items, open-questions roll-up) builds on meeting_summary_items + listSummaryItems.

🤖 Generated with Claude Code

[r3dbars]

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