[itercharles]

Summary

Closes the loophole noted in #454: a CR whose triage_result says rejected skipped the code and impact_analysis checks, and any agent can write that field, so an agent could reject its own CR and change requirements or code under it.

• design_validation.impact_findings now decides once:
  • rejected, and the branch changes only the CR → skipped as before (only the dependents check runs, as it always did);
  • rejected, and the branch changes another item or any code → a triage_result finding ("a rejected CR changes nothing", naming up to five changes) and every impact check, as for an accepted CR.
• The rejection test is removed from _unreviewed_code and _check_impact_analysis. triage_rejected, cr_triage_rejection and build plan's early stop on a rejection are unchanged.
• docs/interface.md: one sentence in the verify impact paragraph. CHANGELOG ### Fixed under Unreleased.

Measure

New seeded omission rejected_to_skip_the_checks: the CR marks itself rejected while the branch also changes src/export.py, which it never accounts for.

• Before this change: passed every gate.
• After: stopped by verify impact.

Seeded omissions stopped: 9 of 12 (10 passed, 3 xfailed).

Tests

• test_verify_impact.py: the "rejected is not checked" test now uses a branch that changes only the CR; new test that a rejected CR whose branch changes items fails with triage_result: plus the normal findings.
• test_a_changed_file_reaches_its_module.py: new test that a rejected CR whose branch changes only a code file is stopped.
• Full suite: 1949 passed, 9 skipped, 3 xfailed (before the rebase onto main; the three affected test files re-run after it).

Manual review

• Drafted by a Haiku subagent. The seed case it was given did not exercise the loophole (it was already stopped by the dependents check); it was replaced and verified before and after the fix.

🤖 Generated with Claude Code