perf(context): share project-wide aggregates across processes

[gtarpenning]

Summary

getDominantFile() and getStats() aggregate the whole graph, and neither depends on the query. The prompt hook is a fresh process per prompt, so the in-process memo from #1864 never helps it: every prompt recomputes both (about 1.3 s warm, more cold). This keeps the results where a new process can read them, and refreshes them when the graph changes.

Independent of the trigram PR; either can merge first. #2115 (an in-process getStats cache) and #2260 touch nearby code in queries.ts; this stays minimal around them.

What changed

• graph_epoch is a project_metadata row replaced whenever the graph changes: indexAll, sync (unless it changed nothing), indexFiles, resolveReferences, resolveReferencesBatched and clear. No schema change.
• Sidecar memo (<db>.memo.json) holds both results, tagged with the epoch. A tag mismatch, a missing file or unreadable JSON recomputes and rewrites it. It is a file, not a metadata row, because the hook's connection would otherwise need the database write lock the daemon's writer holds.
• Warm after writes: indexAll recomputes inline; other writers do it after 15 s of quiet on an unref'd timer, so a burst of edits pays once.
• indexFiles now awaits orchestrator.indexFiles inside its try, so the file lock is held until indexing finishes instead of being released as the promise is created.
• Reads are skipped inside a transaction and on an index with no epoch yet; any sync seeds one.

Measured

Prompt hook on a large TS/Go/Python monorepo index (about 1.5 GB, 350k nodes, 1.2M edges), same JSON on stdin:

	Wall time
Before	3.2-3.3 s
With this PR	1.6 s
With this PR and the trigram PR	about 1.1 s

The first run after a real change pays the recompute once. Hook stdout is byte-identical to the unmodified build on every run.

Compatibility

• A writer without this change (an older install on a shared index) does not bump the epoch, so the memo can be stale until a newer build writes.
• One-shot codegraph sync exits before the deferred warm; the next hook recomputes once.

Testing

• npx tsc --noEmit
• npx vitest run __tests__/persisted-memo.test.ts __tests__/dominant-file-cache.test.ts __tests__/sync.test.ts: a stale tag and a corrupt sidecar both recompute and repair; a fresh process sees the new graph after indexFiles and clear; the deferred warm populates the sidecar.
• npx vitest run: 5747 passed, 23 failed. 20 are in daemon-older-version, daemon-registry and index-command and fail the same way on unmodified main; 3 are git-index-currency timeouts under load that pass in isolation on both.

🤖 Generated with Claude Code