feat(bench): re-grade stored runs when a scenario's checks change

[e54-bot]

Fixes #614 (SPEC-414 REQ-007).

Summary

A scenario's check change turns its grader hash over, but stored runs kept their old verdict — report --regrade only flagged unstable graders and never wrote grades back (and left regrade-1/regrade-2 scratch dirs inside each trial, 162 in one report).

• New agent_bench.py regrade <dirs> --wright <binary> --out <dir>: copies every trial into <out>/<run-dir>/<trial path> — input directories are never modified — and re-grades only trials whose recorded grader.hash differs from the current scenario's.
• A re-graded result.json keeps everything the agent produced (transcript, trace, usage, identity) and records the earlier hash as grader.regradedFrom. Trials at the current hash are copied without invoking the grader.
• A stored invalid caused solely by grader unavailability clears when the new grade has every grader available; agent-side invalid reasons are preserved.
• The report prints each re-graded scenario's hash transition and the count of changed usable verdicts.
• report --regrade now grades into a TemporaryDirectory — no scratch directories inside the run directories on either path. regrade grades into <out>/<trial>/grading, matching run_trial's layout.
• --out must be disjoint from every input directory and must be empty or absent.
• docs/agent-benchmark-howto.md gained a "Re-grade runs after a scenario's checks change" section: when to re-grade and where the output goes.

Verification

• python3 -m unittest test_agent_bench — 103 tests; 2 pre-existing environmental failures (test_scenarios_are_solvable_and_not_vacuous needs the unpinned oracle install; test_matrix_skips_... fails identically on the clean checkout), 3 oracle-skipped, all others pass including 3 new RegradeTest cases:
  • test_a_changed_check_regrades_the_copy_and_preserves_the_original — check change → new verdict + regradedFrom + byte-identical original.
  • test_a_current_hash_is_copied_without_grading — matching hash → copy, no grading/ directory.
  • test_report_regrade_leaves_no_scratch_inside_the_run_directory — regrade_notes leaves nothing behind.
• End-to-end smoke: agent_bench.py regrade on a fabricated run with a stale hash re-graded the repair-runaway-loop seed (usable true→false: no-unbounded-loop + root-cause-named fail), recorded stale-hash-forced → d7eb57b9…, and left zero files in the input tree.
• git diff --check clean.

[e54-bot]

Review fixes in 652de1e:

• Pre-hash archives no longer crash the transition summary — regradedFrom: null is preserved in the copied result and renders <none> in the report; new test test_a_pre_hash_archive_regrades_and_reports_the_transition covers the grader-less path.
• Unused tempfile import removed; run_status + INFRA_EXIT moved to bench_report so agent_bench and bench_regrade share one definition instead of a duplicated status_of with a hardcoded 75.

Note: test_scenarios_are_solvable_and_not_vacuous and the matrix interruption test fail locally — every .opy scenario's upstream-valid check fails because the pinned OverPy oracle toolchain isn't installed in this environment (ana-paintball-ws, which needs no oracle, passes), and the matrix test hits local adapter limits (pi-m missing, provider outages). Both are environmental — the grading/status logic the diff touches is unchanged.