feat(resummarize): re-run the AI pass on edited meeting/memo transcripts

[rymalia]

Implements #523 — re-run the AI pass on an edited meeting/memo transcript,
per the maintainer-approved plan (docs/plans/rerun-ai-pass-resummarize-2026-07-21.md,
maintainer notes M1–M5 folded in).

What this adds

• minutes-core::resummarize::resummarize_meeting — re-runs the
  summarize stage + local derivations (action items, decisions, intents,
  entity links) against the current transcript text and splices the
  regenerated AI-owned sections back in under a strict safety contract:
  • Never destroy on failure. Engine none, provider errors, empty
    output, scaffolding-only output, and unstructured-regression output are
    hard no-write failures.
  • Concurrent-edit guard. Full-content snapshot before inference,
    re-compared before the backup copy and again immediately before the
    atomic rename.
  • Splice, don't rewrite. Only AI-owned body sections
    (## Summary … ## Commitments) and derived frontmatter change;
    ## Notes, ## Transcript, speaker_map, capture/consent metadata are
    preserved byte-for-byte / field-for-field. Fence-aware, fail-closed
    section parser (duplicate headings reject the write).
  • Status-preserving merge (M2). Exact normalized identity only — never
    fuzzy. User-curated state (status: done, due dates, decision
    authority/supersession) carries forward; everything ambiguous, unmatched,
    or dropped is surfaced in the preview, never silently resolved.
  • Timestamped backups. Each apply writes
    .<name>.pre-resummarize.<unix-secs>.bak (mode-preserving); the newest
    3 per artifact are kept.
• CLI: minutes resummarize <meeting|memo|search-term> — preview by
  default (the preview does invoke the model, stated in the output per
  M3); --apply, --engine, --template, --json (errors also emit a JSON
  envelope). Serves minutes import text archives (feat(import): generic importer for existing text transcript archives #516) as their first AI
  pass.
• Run record (M4): summarization: health block shaped like the existing
  speaker_mapping block (model, template, duration_ms, last_run).
• Engine choke point (M1): engine resolution happens only inside
  summarize_with_template — resummarize passes config through and can never
  bypass per-meeting egress policy. Note for the in-flight privacy work:
  this adds a second caller of that choke point.
• Scope per M5: text-only, CLI + core. Desktop button and MCP tool are
  follow-ups (plan Phases 3–4).

Review findings intentionally declined (with rationale)

Two findings from the Codex review rounds were declined rather than fixed:

1. Assignee-only action items count as "pristine" (droppable).
   Assignee and due are extractor-derived on every fresh pass (@name:
   prefixes), so treating assignee as user-curated state would make items
   effectively undroppable. Only status != open and an explicit due date
   count as user state — and every drop is surfaced in the preview
   regardless.
2. Unknown YAML keys are lost on frontmatter reserialize. Identical to
   the in-tree update_frontmatter precedent that
   redo-speaker-mapping already ships with; fixing it is an orthogonal
   schema-preservation project, not part of this feature.

Clarification: sensitivity: restricted artifacts

Restricted artifacts are not blanket-rejected on purpose: generated
sensitive artifacts carry capture: none, and that gate already rejects them
(their "transcript" may be a placeholder plus human-authored debrief).
A restricted artifact with a real transcript is a legitimate resummarize
target; on cloud engines the preview/apply output states that the model was
invoked, and the M1 choke point is where per-meeting egress policy will be
enforced.

Notes

• The core fence parser tracks fence char + opener run length
  (CommonMark closing semantics) — a strict improvement over the historical
  CLI transcript extractor, which tracked the char only.
• The tmp file's mode is enforced with set_permissions on the open handle
  (fchmod), not OpenOptions::mode alone, which is umask-masked — preserves
  0640 Team-visibility artifacts under restrictive umasks. A
  umask-manipulation test was deliberately skipped (process-global, not
  parallel-safe); a 0640 mode-preservation test is included.
• Backup retention count (3) is cheap to change if you'd prefer another
  number.
• Redaction caveat documented in README: transcript edits never touch the
  retained WAV — true audio redaction is minutes cleanup/retention
  territory.

Tests

1158 core unit + 8 integration + 80 CLI tests pass (--no-default-features);
workspace clippy -D warnings and cargo fmt --check clean. New coverage
includes: the fence-aware section parser (incl. fence-length and
closing-syntax cases), splice, the status-preserving merge (densest coverage,
per M2), concurrent-edit guard + backup isolation on conflict, 0640
mode preservation, scaffolding and unstructured-output rejection, timestamped
backup naming + prune-to-3 retention.

[vercel]

@rymalia is attempting to deploy a commit to the evil genius laboratory Team on Vercel.

A member of the Team first needs to authorize it.

[rymalia]

Note on CI: the four MCP jobs (Lockfile ×3, MCP Server) fail because crates/mcp/package.json pins minutes-sdk@0.23.0 (from fc9e4ee), which isn't published to npm yet — the same failure is currently on main itself, so it's inherited from the mid-release state rather than caused by this diff. They should turn green once minutes-sdk@0.23.0 is published (or on a re-run after that).

The two agent-docs sync failures were real (the new resummarize error variants weren't in the generated error catalog) and are fixed by the latest push, which regenerates site/public/docs/errors.md + site/app/docs/errors/data.json.

Rust build/test jobs pass locally on the merged tree: 1168 core + 8 integration + 80 CLI tests, clippy -D warnings clean.

[silverstein]

Heads up on the red checks before you spend any time on them: the MCP Lockfile / MCP Server / CI Gate failures are not yours. Main is mid-release for v0.23.0 - the release pin commit references minutes-sdk@0.23.0, which only appears on npm when the release tag publishes it, so every PR's merge ref fails npm ci in that window. Those checks will heal on a re-run once v0.23.0 is out (shortly).

On the PR itself: impressive turnaround, and the safety contract reads exactly like the plan we agreed - the declined-findings section with rationale is exactly how I like disagreements documented. As noted on #523 this lands in the release after 0.23.0, so I will do the full review right after the release ships. Two early notes meanwhile: (1) the CLAUDE.md command-count and commands.rs test-struct updates are fine to keep in scope; (2) please rebase on main once 0.23.0 is tagged so the generated site constants regenerate against the released tree - that has bitten two PRs this week.

[silverstein]

Full review done: approved. This is the most disciplined feature contribution this repo has received, and the safety contract is not just described - it is implemented and tested.

Verified independently (fresh worktree, own test runs, plus reading every safety-critical path):

Merge semantics (the part I flagged as hardest): correct. Exact normalized identity with uniqueness required on BOTH sides before anything matches - duplicates go to Ambiguous rather than being guessed at, user-stated items survive unmatched, conflicts keep the user's value and emit an explicit note, and every drop is surfaced. That is the wrong-is-worse-than-none doctrine implemented, not just cited.

Splice: fail-closed as specified. All AI-owned ranges resolved strictly before any mutation; duplicate headings abort; covered by tests.

Engine choke point (M1): honored. No engine resolution anywhere in resummarize - everything flows through summarize_with_template, with an injectable seam that makes the failure/concurrency/splice contract testable without a live engine. This composes correctly with the in-flight per-meeting egress policy work.

Write safety: better than the plan asked for. Unique tmp names, mode-correct from the first byte, post-write frontmatter validation, the concurrent-edit guard compared twice (before backup and again before rename), failed runs cleaning up their own backups, newest-3 retention. The residual compare-to-rename window is inherent without file locking and acceptable for a user-edited file.

Both declined findings: agreed with your rationale. Assignee/due are extractor-derived each pass, so treating them as user state would make items undroppable; and the unknown-YAML-keys limitation matches the existing update_frontmatter precedent - an orthogonal project, correctly kept out of scope.

My runs: 34/34 resummarize + 81/81 CLI green, scoped clippy -D warnings clean, fmt clean. Your branch CI already covered the full three-platform matrix; the only reds were our release-window npm ci failures, now healed since v0.23.0 published.

One mechanical step before merge: rebase onto current main (v0.23.0 shipped, so the site constants need regenerating against the released tree - node scripts/generate_llms_txt.mjs && node scripts/sync_site_release_version.mjs, commit site/) and push. Green CI after that and this merges as the headline of the next release.

[silverstein]

I pushed a mechanical update to your branch (maintainer edit) to unblock the two red checks — no resummarize logic touched:

1. Merged current main into the branch. Your PR predated the v0.23.0 release, and the release moved the generated version constants (site/lib/release.ts, llms.txt, error + MCP catalogs). Merge was clean — no conflicts with your code. I regenerated those docs so counts reflect the released tree (58 commands, 1660 tests including your 34 new resummarize tests). Regen is idempotent.
2. That also re-rolls the one macOS failure, which was not your code: it was the pre-existing flaky copilot::relay reconnect test (broken pipe), tracked in Flaky test: copilot::relay repeated_reconnects_replay_each_cursor_advance fails with broken pipe on macOS CI #529. 1,167 other tests passed.

Once CI is green on the new head I'll squash-merge. This lands as the headline of the next release. Thanks again for an exceptionally clean feature.