[gabe4coding]

Summary

This PR fixes two findings that shall review found in findings.py. The code came from review.py, so the problems were there before.

JTOOL-0002.6: only a verified finding blocks

Before, Finding.blocking returned corpus_blocking when there was no verdict. Thus --no-verify, review.verify: false and shall validate gave blocking findings that no verifier saw. The exit code, the GitHub review and the Stop hook read blocking, so all three were affected.

• Finding.blocking is true only when the corpus marks the finding as blocking and the verdict is confirmed.
• needs_verification reads corpus_blocking. Before verification, blocking is now false for every finding.
• sort_key sorts by corpus_blocking second, so the comment cap keeps the findings that can block.
• A finding without a verdict shows "Not verified — not blocking."
• shall review --fail-on-blocking with verification off prints a warning, because it cannot fail.
• Breaking: shall validate no longer accepts --fail-on-blocking. The flag could never fire. The gate is shall verify --fail-on-blocking.

JTOOL-0004.2: one reviewer prompt

chunk_prompt repeated rules 1, 2 and 4 of agents/shall-reviewer.md (only violations, an empty list, line anchors). Now it sends only data: <chunk>, <known_effects> and <standards>. The bundle header tells the steps (reviewer, validate, verify) and gives no review rules. A new test fails if text gets added outside the data tags.

shall-reviewer.md did not change. The plugin version is 0.4.2, because a skill reference changed.

Eval: old prompt and new prompt

Haiku (ClaudeJudge through claude -p, no Jev) selected the statements for the 11 eval/cases. Both prompts used the same selection. The reviewer was Sonnet, with 2 runs per prompt and no cache. The findings were compared with the violations labels in eval/labels.yaml.

Run	Recall	Precision	Dropped by validation
old #1 / #2	0.868 / 0.868	0.825 / 0.786	1 / 1
new #1 / #2	0.868 / 0.868	0.805 / 0.846	1 / 0

The difference is inside the noise between runs. Without the anchor text, the reviewer did not give more invalid line numbers. Cost: $1.19. make eval-realpr did not run.

Test plan

• make test (202 passed), make check, make self-check
• /simplify and /code-review, with their fixes applied
• /verify with cli on case 08 and a Haiku selection:
  • bundle: the new header, and only data tags in each chunk section.
  • validate: exit 0, the finding is advisory.
  • validate --fail-on-blocking: exit 1, usage error.
  • verify --fail-on-blocking: the verdict is confirmed, the finding blocks, exit 2.
• Not driven: review --no-verify. review needs Jev for selection.

🤖 Generated with Claude Code