perf: consolidate grep cache index

[sorafujitani]

Summary

• replace per-transcript grep cache files with one per-cache-dir index.json
• store v2 cache records as a joined text payload plus fragment count
• preserve fragment-boundary matching by splitting cached text before TextsContain
• treat v1/no-version records as misses and tolerate corrupt indexes by falling back to extraction
• expand the repeated-query benchmark to 512 sessions

Fixes #102

Benchmark

Baseline was measured on origin/main with only the benchmark session count changed to 512.

Command:

go test -bench=BenchmarkFilterRepeatedQuery -benchmem ./internal/grep -count=10

benchstat /tmp/ccsession-102-before-512.txt /tmp/ccsession-102-after.txt:

FilterRepeatedQuery/direct-read-12  462.83m ± 178%  18.35m ± 4%  -96.04% (p=0.000 n=10)
FilterRepeatedQuery/warm-cache-12    28.046m ± 777%  1.294m ± 5%  -95.39% (p=0.000 n=10)

FilterRepeatedQuery/direct-read-12  41.67Mi ± 0%  41.27Mi ± 0%   -0.97% (p=0.000 n=10)
FilterRepeatedQuery/warm-cache-12   6490.5Ki ± 0% 378.0Ki ± 0%  -94.18% (p=0.000 n=10)

FilterRepeatedQuery/direct-read-12  19.52k ± 0%  15.43k ± 0%  -20.96% (p=0.000 n=10)
FilterRepeatedQuery/warm-cache-12   13.367k ± 0%  3.636k ± 0%  -72.80% (p=0.000 n=10)

The direct-read timing is noisy because the benchmark deliberately points the cache dir at a file to disable caching. The main target is warm-cache grep reloads, which now read one index instead of one cache file per session.

Verification

go test ./...
go vet ./...
golangci-lint run
gofmt -l $(git ls-files '*.go')
git diff --check