fix(briefing,hooks): derive where the work stands; stop storing it

[kevintseng]

What prompted this

This session opened with memesh telling me:

Where "memesh" was left off (8 days ago):
- Just finished: v4.6.0 已發佈
- Next: UX-2：Search/Browse/Manage/Lessons 四面合一

38 PRs had merged since v4.6.0 and npm was serving 4.7.3. UX-2 was neither done nor being worked on.

Nothing was corrupted. Every field was true when it was written.

The category error

	memory	status
what it is	true once written	true right now
goes stale	never	silently
who writes it	someone said it	the world changes
when wrong	one fewer memory	actively misleads

task_state is stored like the first and read like the second, so it can only ever report the moment someone last spoke.

src/core/task-state.ts already carried a comment saying the age belongs in the heading rather than being decoration — the staleness was anticipated, and (8 days ago) was right there in the line that misled me. An annotation cannot fix a sentence that asserts.

The change

src/core/repo-state.ts (new) derives, never stores: branch, uncommitted count, last tag, commits since it, and whether the declared version has a tag. Read from local git on every call, so staleness is impossible by construction rather than by discipline.

The task-state heading attributes instead of asserting: Stated about "X" 8 days ago, and not revisited since. done becomes Had just finished.

Local git only — open-PR state would need auth, add latency to a hook budgeted in seconds, and fail offline. Best-effort throughout: no repository, no git, a timeout → null, and the block simply does not appear.

Three things the tests caught that my plan had wrong

1. "Reuse runGit from why.ts" was wrong. It made repo-state un-mirrorable — generate-hook-core.mjs flattens paths as it copies, so a mirrored module can only import mirrored ones. The hook would have derived status differently from the CLI: exactly the divergence briefing.test.ts's parity case exists to forbid. repo-state owns its eight lines and is a leaf.

2. Prepending the facts unconditionally broke "say nothing when there is nothing to say." Inside a git repository every call returned a fence whose entire content was a branch name — the one thing the agent can already see. The facts prefix a briefing; they never constitute one.

3. assembleBriefing(project) takes a NAME, not a path. process.cwd() is the right repository for a CLI run and can be an unrelated one over MCP. Reporting this repo's branch under another project's heading would have been a new way of saying something false. Derivation is skipped when the process is not standing in the project being asked about.

Acceptance

Same stored data, on this repository:

Where the repository actually stands (read just now):
- branch main · 20 uncommitted
- at tag v4.7.3

Stated about "memesh" 8 days ago, and not revisited since:
- Had just finished: v4.6.0 已發佈

It no longer claims to be the present.

Evidence

node scripts/run-tests-isolated.mjs   exit=0   190 files, 2607 tests
npm run verify:release                exit=0   all detectors new=0
npm run typecheck                     exit=0
npm run build                         exit=0

Break-tests, each restored byte-identical:

Mutation	Result
derivation stubbed to a constant	exit=1 — the case comparing against git's own count fails
best-effort catch removed	exit=1 — 9 failed

Deliberately not done

No coordination primitives (claims, leases, ownership). git branches, PR state and CI already carry those, and they need the opposite consistency guarantees from the best-effort, lossy, boundary-sampled design that is correct for memory.

Strings stay English, consistent with config.ts's existing contract — the injected block is scaffolding for an agent, not UI. The dashboard's 11 locales don't translate it (dashboard/src/lib/i18n.ts matches 0), and field values are whatever the user wrote; verified rendering with Japanese, German, Arabic (RTL) and emoji.