bench: evaluate agent continuation with --json-envelope

[yearth]

Summary

• add a deterministic 120-record HTML fixture with ground truth kept outside the agent workspace
• add conformance and guided/ablated adoption prompts with a fixed multi-page budget
• normalize Claude Code stream-json, correlate standalone ax calls through a benchmark-local shim, and grade answer, continuation protocol, access isolation, and run metrics
• add a guarded 3+3+3 matrix runner and publish sanitized Sonnet 5 results in bench/RESULTS.md

Results

Condition	Exact	Protocol pass	Envelope adoption
Conformance	2/3	3/3	—
Adoption guided	0/3	2/3	2/3
Adoption ablated	0/3	0/3	0/3

The report keeps answer correctness, protocol correctness, and active envelope adoption separate. It also discloses three earlier stopped matrix attempts and the resulting methodology changes, so these results are presented as adaptive and exploratory rather than confirmatory.

Raw session records, telemetry, generated fixtures, and ground truth remain untracked. No src/ files are changed.

Validation

• bun test — 255 pass, 0 fail
• bunx tsc --noEmit
• bun run format
• bun run build
• replay grading of all nine retained runs remained gradable with access checks passing and published scores unchanged
• independent final code review: ready to merge

Closes #53

[yearth]

@yusukebe When you have a moment, could you take a look and let me know if you would prefer any changes to the benchmark design, results reporting, or scope before review? The current CI is green, and the PR remains limited to bench/ with no product-code changes. Thank you!

[yusukebe]

@yearth

Okay. I was feeling bad these days, but I'm back. I'll check it!

[yearth]

Glad to hear you're feeling better. Please take care of yourself, and there's no rush at all from my side. Thank you for taking a look when you have time!

[yusukebe]

LGTM!

[yusukebe]

Hey @yearth

Super interesting! This is an ideaful and well-implemented. Looks good. Thanks!