feat: turn caps (--max-turns) and raw per-run results (--raw-out)

[queso]

Fixes #33, fixes #34. Built for the AST repo-map spike, where the primary metric is exploration tokens and a turn-capped run must score as a localization failure rather than a cheap success.

Turn caps (#33)

• maxTurns scenario field (per-case override allowed) + --max-turns on run, compare, and measure, passed through to claude-p as --max-turns. Completion runners are single-turn and ignore it.
• A run that stops at the cap — result subtype: "error_max_turns", on any exit code — is returned as a measured outcome flagged exhaustedTurns instead of thrown as a crash, so one capped run no longer aborts a whole comparison.
• The engine scores a capped run as a failed run without grading: partial output that happens to satisfy a grader must not count as a pass, and a judge grader would bill a model call for a result that cannot stand either way. Summary line reads hit the N-turn cap before finishing.
• The effective cap joins the baseline cache key. stableStringify drops undefined entries, so scenarios without a cap keep their existing cache keys — no spurious cache busts.

Raw per-run results (#34)

• --raw-out <dir> on compare and measure writes each run's full runner result JSON (claude-p: usage token categories, modelUsage, subtype) as <scenario>_<arm>_<n>.json (_measure_ for measure), sanitized the same way receipts sanitize scenario names.
• Files are written the moment each run finishes, so an invocation that dies midway still leaves the completed runs' records.
• Cache-served baseline arms ran in an earlier invocation and have no raw results; the progress output says so explicitly instead of leaving a silent gap.
• Opt-in; summaries, receipts, and the ndjson report are unchanged.

Verification

• bun run check green: 113 tests (6 new), typecheck clean.
• Live smoke test against claude CLI 2.x on macOS: --max-turns is accepted in -p mode (registered in the binary despite being absent from --help), and a capped run returns exit-relevant JSON with subtype: "error_max_turns", result: null — handled by the null-tolerant normalizer.
• New tests: arg construction with/without the cap; a fake-binary integration test for the non-zero-exit capped path; the crash path still throws; config parsing/validation at both levels plus CLI override precedence; engine test that a capped run fails even when its output would satisfy the grader; raw-out file contents for both compare arms and measure, and that nothing extra is written without the flag.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EqktLpWgVX5orcGAzx7EK1