[TheGreatAxios]

Summary

• Cuts per-turn session-store I/O: one cached worktree survey replaces per-candidate stat probes inside commit() (fs-op census: 139 → 124 ops/turn on a 200-turn simulation; fs.access 2,697 → 3 per run); a per-store turns cache with per-segment stat fingerprints + explicit mutation-site invalidation parses each segment once (the three separate full-history parses per turn become one shared generation, parsing only changed/new segments); contentPrefixLength/promptEqualsLiveTurns memoize turn JSON instead of re-stringifying full history; the well-formedness prefix check becomes a single forward pass (O(E²) → O(E), equivalence probed on 50,000 randomized interleavings before ship); readAt resolves a commit-time segment-manifest.json and parallel-reads segment blobs, degrading to the old ref-tree walk on missing/torn/foreign manifests; attachment blob reads dedupe per id; aged-image markers prefilter before regex; the anthropic-cache prompt builds its tool-name/boundary maps in one forward scan; compaction call identities are cached per persisted block.
• Cache invalidation is wider than the reads it replaces: every segment mutation site (appends, staged rewrites, heal unlinks, failed-commit restores) drops the cache in-process; external writers surface via stat fingerprints; foreign same-size/same mtime+ino rewrites are the one theoretical evader (noted, mitigated only by extreme coincidence).
• Durability first: the resume/rebuild contract is unchanged; the manifest is name-based (never stale after gc/history rewrite) and every read path has a verified fallback to the pre-existing walk.
• Ships with a repeatable measurement harness (scripts/context-store-perf.ts, standalone; fs-op census deterministic, wall noisy).

Verification

• bun run check exit 0: 8,945 pass / 0 fail across 580 files; src/session suites 62 store tests (4 new) + 7 attachment tests (2 new) green.
• Independent review: invalidation enumerated per mutation site; memo preconditions verified against frozen turn boundaries; well-formedness equivalence probed; harness honesty verified (never picked up by the test suites).
• Before/after (documented): fs ops −11%/turn; wall median ~8.3 → 7.3 ms/turn over 4 paired runs (host-noisy; census is the reliable number).

Fixes CL-10666

[linear-code]

CL-10666

[corbits-triage]

Superseded by the comment below.

[corbits-triage]

Superseded by the comment below.

[corbits-triage]

Triage: Pass

No changes needed from you. A maintainer will review this pull request.

[TheGreatAxios]

Independent perf verification for this PR (measure-only lane, no source touched).

Method: the commit's own harness scripts/context-store-perf.ts (fs.promises op shim, split by .git paths), run verbatim on three cells × 3 runs × 200 turns, one machine (Bun 1.4.2, arm64): PRE @ 10042685 (pre-merge parent), POST @ 8d879e94 (landing), MAIN @ origin/main.

• fs ops/turn: PRE 139.12 / 139.22 / 139.16 (median 139.16) → POST 123.84 / 123.89 / 123.87 (median 123.87). Delta ≈ −15.3 ops/turn (−11%). Claim CONFIRMED. Counts deterministic except mkdir/writeFile ±6 (isogit dir retries); everything else bit-exact across runs.
• Access probes: PRE 2697 / 2697 / 2697 → POST and origin/main 3 / 3 / 3, constant and scale-free. Claim CONFIRMED exactly. The ~2694 eliminated probes are the per-candidate-segment access/stat walk in commit() replaced by one surveyWorktree readdir.
• Wall median ~13%: PRE 914/1181/782 (median 914 ms) → POST 702/958/793 (median 793 ms, −13.2%). Directional only — n=3 wall is noisy (ranges overlap).
• origin/main today: ~124.9 ops/turn, access still 3. Later coalescing shifted composition (readFile up, lstat down) but the headline holds.

No prompt/policy retune indicated — the optimization measures exactly as claimed.