[sjkim1127]

[AUTOMATED]

Scope

Reuse immutable positive FID candidate records at LazyFidDatabase. This follows #246; no new dependencies, CLI/schema/adapter/archive change, changed worker/function/instruction budgets or matching policy.

• Cache parsed records by archive-local full hash, preserving order/duplicates/decoder fields.
• Recompute specific-hash, force/auto flags, relation contexts, thresholds, library lookup and final matches on every query; returned names are owned copies.
• Parse/read outside the cache mutex; positive-only insertion, retry misses/malformed-empty/errors, poison fallback. Concurrent cold reads may duplicate work. Records are retained per immutable database; this is not a fixed memory-cap cache.

Measured evidence

Frozen C++ qualification revision 07dd95dbe6214b8a78fc1e4a55961ab515986489, 45 cells / 250 stripped functions. All inputs rehashed; same resources/cwd/addresses/limits; fresh processes, --no-db, DECBENCH_NO_CACHE=1, release LTO, native M4 Pro/24 GiB. No own builds/tests/profilers during controls; OS caches/scheduling uncontrolled.

Large six-function input, three alternating pairs: wall 14.6610 -> 13.8580 s (-5.5%), CPU -9.0%, median RSS +2.7%. Small five-pair changes are only 0.2–1.3%; no broad short-task gain or memory-cap claim.

Fresh 250/250 outputs, zero changed non-time function records: code/layers, variables/types/PCs/maps, errors/fallbacks/evidence/counters all compared; only seven existing durations excluded. Focused 22-function complete records match three repetitions in both worker lanes and fresh replay. Whole-population two-run determinism, execution/recompilation and rescoring were not measured.

Pinned Kuna 1.808, both requested jobs/render workers 1 and 14: choosing each tool's faster observed median independently still leaves Fission 2.73–6.10x slower. No semantic/readability/score/ranking or Kuna-parity claim.

Validation

Five focused tests; 1,952 related nextest passed, six skipped; crate checks, scoped CI Clippy, release, workspace fmt and owner-boundary scan pass. Differential reference covers all flags, specific hashes, confidence boundaries, relation contexts, ties, malformed rows, resource ownership, concurrency/poison fallback and returned ownership.

Full contracts, source hashes, CPU/RSS ranges, both worker lanes, controls and reproduction: docs/validation/fid-candidate-performance.{md,json}; proposal docs/proposals/2026-10-15-fid-candidate-cache.md. Full local artifacts: benchmark/artifacts/performance/2026-10-15-fid-candidates/.

Initial frontend preparation, misses and completed-function hot-cache growth remain. Primary checkout, drafts #203/#204, DecBench #103 and seven unrelated dirty main files are preserved. Remote CI on this exact head remains required before merge.

Additional combined controls

A separate three-pair direct comparison was run after all profiler processes exited, using immutable pre-#246 executable vs this final candidate on the same six-function large batch. Median wall 26.2682 -> 14.3305 s (-45.4%), CPU +2.9%, RSS +3.5% (2,327,576,576 -> 2,409,152,512 bytes). All six complete non-time records match every pair/repeat. These are direct combined controls, not compounded percentages from earlier experiments; their raw commands/JSON/stderr and ranges are retained in benchmark/artifacts/performance/2026-10-15-fid-candidates/combined/ and COMBINED-REPORT.md. The narrower cache-only table above remains independently reported.