Fix CI until green

For a staging/production failure that is not a PR check — especially with no local repro — use debug-production (and hotfix-and-rollback if service is down) instead of this workflow.

1. Watch — ci-watcher agent reports failing checks with links and a category guess.
2. Classify & fix — fixing-ci: env/infra → retry or escalate; flake → triaging-flaky-tests; real → reproduce locally and fix the root cause (debugging-an-issue).
3. Re-run the exact failing command locally, then push.
4. Loop until the gate holds. Never disable a test to get green.

Run the fix cycle as a verified loop (agentic-loop) with a hard iteration budget. Hand-off: ci-watcher surfaces failures; fixing-ci routes each to the right skill; the state file tracks which checks remain.