[JacobStephens2]

Summary

CI watch now owns observation and both per-head re-run allowances for one Delivery.

Delivery creates one CI watch
  Repair loop: merge Base → push → assess CI → choose Repair / Base fix / Self-merge
    completed CI-fix Repair → notify CI watch
  CI watch: observe checks + reconcile attempts + remember full-head allowances

Closes #955.

Evidence

• Before: the first assessment test failed to compile because Watch did not exist; the Repair loop selected three raw CI operations and tracked allowances.
  After: 52 CI watch tests and 47 Repair loop tests pass through the narrowed interface. The per-head policy matrix now uses scripted check/job observations; the loop tests retain ordering, stages, budgets and Base fix routing.
• The CLI regression makes an unchanged head red after a refused Self-merge: one CI-fix Repair, one Check re-run, then the same Declined CI fix in CI watch. The targeted test and direct GitHub cancellation test pass.
• Local cargo check, cargo fmt --check and cargo clippy --all-targets -- -D warnings pass. The full Rust suite passed in CI on exact head 4eb7e7f: 2,249 passed, 51 ignored. CI also passed site, Console and the macOS gate.
• /usr/bin/time -p scripts/affected-tests main selected the full Rust suite using nextest. The redundant local run was stopped after 1,182.07s, including compilation, once exact-head CI passed; its successful result is reused per docs/testing.md.

Test seams: CI watch assessment with the existing in-memory GitHub adapter; Repair loop assessment/completed-CI-fix notification; CLI Delivery with real Git and fake GitHub, including captured pull request gates.

Independent Standards and Spec reviews read all seven changed files. No generated exclusions or unread changed files remain.

Merge Danger

Door: two-way

An internal ownership refactor with no configuration, persistence or CI policy change.

Blast Radius: Delivery

Runs and Spec PRs share the CI assessment path.

Unaddressed findings

• Standards: None (0 findings).
• Spec: None (0 findings).
• Security: None (0 findings).

Built with codex · gpt-6.1-sol · xhigh

Security review outcome

No unaddressed introduced findings.