perf: cache session scan metadata

[sorafujitani]

Summary

• add an mtime/size-validated on-disk cache for Claude session scan metadata
• cache unusable transcripts as empty records so dead files are not reparsed every launch
• keep cache failures non-fatal and fall back to parsing
• isolate scan cache dirs in tests and add cold/warm scan benchmarks

Fixes #99

Benchmark

Baseline was measured on origin/main with the same benchmark body added temporarily and no scan cache implementation.

Commands:

go test -bench=BenchmarkScanManySessions -benchmem ./internal/session -count=10

benchstat /tmp/ccsession-99-before.txt /tmp/ccsession-99-after.txt:

ScanManySessions/warm-cache-12  21.86m ± 7%   11.63m ± 15%  -46.81% (p=0.000 n=10)
ScanManySessions/cold-cache-12                 134.3m ± 10%

ScanManySessions/warm-cache-12  19.275Mi ± 0%   3.694Mi ± 0%  -80.83% (p=0.000 n=10)
ScanManySessions/cold-cache-12                  22.82Mi ± 0%

ScanManySessions/warm-cache-12  57.06k ± 0%   31.05k ± 0%  -45.59% (p=0.000 n=10)
ScanManySessions/cold-cache-12                 87.14k ± 0%

The cold-cache sub-benchmark includes cache writes, so it is expected to be slower than the old no-cache scan. The user-facing win is repeated picker/list reloads once the metadata cache is warm.

Verification

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