Fix loop watch writer fallbacks and stale-pending self-heal

[Jonnyton]

Summary

• Treat a recent successful auto-fix-bug.yml workflow_run as productive fallback evidence when the required scheduled writer run is stale.
• Add a regression for the observed live state: stale schedule at 2026-05-08T03:55:48Z plus recent workflow_run success at 2026-05-08T04:28:48Z should downgrade writer health to yellow instead of red.

Why

auto-fix-bug.yml intentionally runs from workflow_run events after Wiki bug sync / deploy completions. The watch already downgraded stale schedule reds for workflow_dispatch and issues successes, but missed the workflow's own backfill path.

This does not claim the schedule is fresh. It keeps that as yellow while preventing a false red when the writer has proven productive recently.

Verification

• python -m pytest tests/test_community_loop_watch.py -q - 28 passed
• python -m ruff check scripts/community_loop_watch.py tests/test_community_loop_watch.py - passed
• python scripts/community_loop_watch.py --json with this patch - overall yellow; Writer workflow yellow with fallback_event: workflow_run

Gate

Writer family: Codex/OpenAI. Needs Claude/Cowork checker review and explicit host merge key before merge.

[Jonnyton]

Codex update after the 2026-05-08 loop-health pass:

• Kept the original fix: stale auto-fix schedules are downgraded when a recent successful workflow_run/dispatch/issues fallback proves the writer path is alive.
• Added a second loop-health fix from live evidence: when the watch run sees stale writer schedules plus pending backlog, it dispatches auto-fix-bug.yml and then self-dispatches one follow-up watch run via workflow_dispatch input self_heal_followup=true.
• Added regression coverage for in-progress fallback runs so the watch reports yellow while a valid writer fallback is still running.
• Live evidence: manual stale-dependency recovery moved issue [WIKI-DOCS] auto_hydrate_then_invoke_v1 — Stage A bridge convention; loop hydrates source_bundle_refs before invoking user-authored classifiers #684 out of pending and opened [auto-change] WIKI-DOCS: auto_hydrate_then_invoke_v1 — Stage A bridge convention; loop hydrates source_bundle_refs before invoking user-authored classifiers #686; local community_loop_watch.py is yellow, with intake/writer/canary green and only review/backlog warnings remaining.

Verification on this branch:

• python -m pytest tests/test_community_loop_watch.py -q -> 29 passed
• python -m ruff check scripts/community_loop_watch.py tests/test_community_loop_watch.py -> passed
• PyYAML parse of .github/workflows/community-loop-watch.yml -> passed
• GitHub checks on Fix loop watch writer fallbacks and stale-pending self-heal #682: policy and actionlint pass

Writer family remains Codex. Required checker remains Claude/Cowork before any host merge key can matter.

[Jonnyton]

Codex follow-up from the duplicate-PR loop-health signal:

Live evidence after #684/#686: auto-fix-bug.yml later selected issue #483 again even though #483 already had open PR artifacts (#678/#679/#681/#685), and it opened another duplicate, #688. Root cause is discovery retrying auto-fix-stale-gate issues without first checking for linked open PRs.

New patch in commit 177e9d0:

• auto-fix discovery now paginates open PRs once per run and skips an issue when a linked open PR already exists by branch ref/body/title issue mention.
• added static workflow regression coverage so stale-gate priority cannot bypass duplicate-open-PR detection.

Additional verification after this patch:

• python -m pytest tests/test_auto_fix_workflow.py -q -> 68 passed
• python -m pytest tests/test_community_loop_watch.py -q -> 29 passed
• python -m ruff check scripts/community_loop_watch.py tests/test_community_loop_watch.py tests/test_auto_fix_workflow.py -> passed
• PyYAML parse of community-loop-watch.yml and auto-fix-bug.yml -> passed

This keeps #682 as a Codex-written loop-health PR; it still needs Claude/Cowork checker before any host merge key can matter.

[Jonnyton]

Codex writer-side rebase complete after PR #617 merge. New head 3ee5654; merge state CLEAN; policy + build-smoke + lint success. Current diff is the loop-watch code/workflow/test slice only; stale #617 baggage is gone. Still needs Cowork/Claude checker review; no host key requested.