bench: Go benchmarks for store, code index, survey, grammar, parse, memory

[jmylchreest]

Benchmark suite for the store / code-index / survey hot paths

Adds 18 benchmarks across 6 packages, measuring the scan / commit / retrieval paths. All construct their own stores in b.TempDir() with in-memory synthetic fixtures (mock CodeGrapher/CodeSearcher reused from existing tests) — zero real-data access, zero network. No existing files modified.

Files (all new)

File	Benchmarks
pkg/store/store_bench_test.go	BoltStoreAddMemories, SearchMemories{,WithScore}, ListMemories
pkg/store/code_bench_test.go	CodeIndexFileBatch, AddSymbol, SearchSymbols{,References}, GetFileSymbols, ClearFile
pkg/survey/survey_bench_test.go	RunChurn, RunEntrypoints, BuildCallGraph, ComputeFreshness
pkg/grammar/grammar_bench_test.go	ScanProject, ScanDetail, NormaliseLang
pkg/code/code_bench_test.go	TokenizeFile{Go,TS}, ParseFile{Go,TS}
pkg/memory/memory_bench_test.go	ScoreMemory{,Detailed}

Key findings (devbox, -benchmem -count=3)

• Writes are fsync-bound (~9–25 ms per bbolt commit), not CPU-bound. IndexFileBatch is ~27× cheaper per record than AddSymbol (~0.7 ms vs ~19 ms per symbol) — empirically validates the batching design.
• Retrieval is index-bound: code symbol search ~2.7 ms over 3k symbols; reference lookup ~10 µs (bbolt prefix scan); GetFileSymbols ~250 µs (JSON decode dominated).
• Parsing is the scan CPU hotspot (~0.65 ms/file) but storage gates indexing (~22 ms commit ≫ parse).
• Language detection ~97 ns and memory scoring ~330 ns, both alloc-free.
• Caveat: absolute store numbers reflect this VM's high-fsync-latency disk; the relative ratios are the durable signal.

Verification

• go vet — clean, exit 0
• Smoke test (-benchtime=1x -count=1) — all 18 benchmarks PASS
• Every benchmark builds its own store in b.TempDir() (local /tmp, not NFS)

Co-authored-by: Hermes jmylchreest+hermes@gmail.com

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!