feat(review-fix-loop): add the cross-cutting, result-blind evaluation corpus

[shaug]

Summary

• Add a reusable, result-blind evaluation corpus for the standalone
  review-fix-loop skill (skills/review-fix-loop/scripts/evals/):
  corpus.py (twenty scenarios), grader.py (independent Git-evidence
  diffing), helpers.py (shared repo/reviewer/decide/apply_fix fixtures,
  importing the identical subset already owned by
  scripts/tests/helpers.py rather than duplicating it), and runner.py
  (the CLI entry point).
• Cover convergence, repeated findings, invalid/incomplete reviews, declined
  findings, budget exhaustion, interruption and recovery, validation failure,
  reviewer mutation, and publication races across both local_commit and
  update_pr, plus the fresh-subagent default and the explicit in-agent
  override.
• Grade every scenario against Git evidence computed independently of the
  returned terminal-result document (a real commit count, a real file's
  content at a real commit, a real remote ref, a real object's reachability,
  a real dirty-worktree listing) rather than the implementation's own claims.
• Add scripts/tests/test_evals.py, run under just test: asserts the full
  corpus passes, asserts corpus shape/scope coverage, and demonstrates the
  grader rejecting both a fabricated convergence claim and a fixture that
  cannot actually converge (this ticket's required seeded-fault
  demonstration).
• Add just test-review-fix-loop and just eval-review-fix-loop, and
  document the corpus in skills/review-fix-loop/evals/README.md and a new
  "Evaluation" section in SKILL.md.

Why

Issue #101 (child of epic #95): establish the standalone review-fix-loop
skill's behavioral contract across both execution modes with deterministic,
result-blind evaluation coverage, complementing (never duplicating) the
capability-owned unit suites test_local_commit.py/test_update_pr.py from
#99/#100.

Acceptance criteria

• Evaluators are result-blind and cannot pass solely because the
  implementation asserts success — grader.grade_case diffs each
  scenario's own checks mapping against Git evidence independently
  queried via rev_parse/ls_remote/show_file/object_exists/
  rev_list_count/worktree_is_clean/untracked_paths, not by
  re-parsing the terminal-result document. test_evals.py's
  SeededFaultDemonstrationTests demonstrates this directly with two
  tests: one fabricates a "converged" claim with no supporting Git
  evidence and shows it's rejected; one runs the real engine with a
  deliberately unfixable apply_fix fixture and shows a wrong
  "converged" expectation is rejected.
• The corpus exercises both local_commit and update_pr — 15 lc_*
  scenarios and 5 up_* scenarios; asserted by
  test_corpus_exercises_both_publication_policies.
• Negative cases prove that incomplete review, reviewer mutation,
  premature publication, lost commits, and stale targets fail closed —
  see the scenario list in skills/review-fix-loop/evals/README.md.
• Evaluation output identifies the exact fixture, observed evidence, and
  reason for failure — grade_case emits
  f"{case_id}: {name}: expected {expected!r}, observed {observed!r}".
• The suite is deterministic enough for repository CI — no subprocess
  boundary, no model call, no network; the full corpus ran repeatedly
  with identical results (one genuine timing/content-collision flake was
  found in the two-clone publication-race scenario during development
  and fixed by giving the two clones distinct commit messages).

Validation

• just format — all checks passed.
• just lint — all checks passed (all 8 skills validated via skills-ref,
  plugin packaging validated).
• just test — 721 tests pass, including 273 in
  skills/review-fix-loop/scripts/tests (8 new in test_evals.py) and 318
  in review-suite/scripts/tests.
• just eval-review-fix-loop (standalone corpus entry point) — 20/20
  scenarios pass, re-run repeatedly with identical results.
• Two independent fresh-context review-code-change passes: the first found
  one strong_recommendation code-simplicity finding (duplicate fixtures in
  scripts/evals/helpers.py vs. the sibling scripts/tests/helpers.py); the
  fix (importing the identical subset instead of duplicating it) was applied
  and the second pass came back fully clean (solution-simplicity,
  correctness, code-simplicity).

Fixes #101

🤖 Generated with Claude Code