Add literal search for archived observations

[aalvsz]

Summary

• Add optional literal UTF-8 search to obs_recall for archived large tool results.
• Return exact byte spans, exclusive ends, LF-based line numbers, bounded UTF-8-safe source context, and continuation metadata.
• Preserve the existing paged-recall path when query is omitted.
• Add validation for malformed Unicode, query size, output limits, cancellation, file guards, overlaps, and continuation.
• Add a reproducible local comparison script.

Validation

• npm run check — 17 test files, 142 tests, typecheck, and package dry run passed.
• npm audit --audit-level=high — passed at the requested severity; two pre-existing moderate Vitest advisories remain.
• node scripts/check-pi-compat.mjs — passed.
• npx --no-install vitest run tests/all-mechanisms.test.ts — 4 tests passed.
• Independent full-buffer oracle — 23 cases / 57 search calls passed.

Real-output benchmark

Five real rg commands from the repository produced 30–166 KiB archived outputs. In a 50-case targeted benchmark (eight observed literals with at most 20 occurrences plus two misses per artifact):

metric	paged recall	literal search	change
tool calls	330	53	83.94% fewer
returned bytes	4,900,500	204,085	95.84% fewer

All 50 cases were source-oracle verified, including offsets, line numbers, contexts, and continuation. A separate dense stress run is retained: queries matching nearly every line required more output than paging, which documents the feature's workload boundary.

These are retrieval measurements over real command output. They do not claim improved model-answer accuracy, provider-token usage, or billed cost. Timings are descriptive local wall times.

[gaoanze888]

Reviewed the literal-search semantics, UTF-8/chunk boundaries, overlapping matches, continuation, output limits, cancellation, and file guards. I ran npm run check locally: typecheck, 17 test files / 142 tests, and package dry-run all passed.

I also integration-merged this with #13 because both touch Observation Pack; the merge was clean and the combined suite passed 17 files / 144 tests. The PR size is justified by the streaming search boundary work plus substantial tests and a reproducible benchmark, rather than unrelated production scope. I did not find a blocking correctness issue. Nice discipline around reporting returned bytes/calls without overclaiming model accuracy or billed cost.