fix(pr-swarm): react to first failing CI check instead of full matrix

[Mathews-Tom]

Summary

pr-swarm lanes previously treated CI/feedback state too passively:

• On launch, a lane only checked state/mergeable/mergeStateStatus — it never looked at whether the PR already had failing checks or open review feedback, and nothing told it to act on that immediately.
• The watch loop was a flat "push, wait ~10 minutes, re-check everything" cycle keyed off the run-level aggregate, so a lane sat idle even after an individual check had already failed.
• Neither the skill nor its lane prompt said anything about gh run rerun requiring repo admin rights — a lane could plausibly read "can't force a fresh CI run" as a reason to stall, when the normal fix-and-push cycle never depends on that command at all.

This PR makes lanes react to what's already visible instead of waiting on a clean baseline or a full CI matrix, and documents why an admin-gated rerun is a non-issue for the normal flow.

What changed, by commit

fix(pr-swarm): react to first failing CI check instead of full matrix — references/lane-prompt-template.md, references/verification-gates.md

• Lane ORIENTATION now runs gh pr checks {PR_NUMBER} alongside gh pr view, and instructs the lane to treat a PR that already has failing checks or unresolved feedback at launch as normal starting state — fix it as the first action, don't wait for a clean baseline.
• WATCH LOOP rewritten from waiting the full ~10-minute quiet window before checking anything, to polling per-job status (gh pr checks) and reacting to the first individually-failing check or new feedback, without waiting for the rest of that run's still-pending checks. The passive quiet-window wait now only applies once nothing is currently actionable.
• New verification-gates.md section "React to the first failing check, not the full matrix": explains gh pr checks is per-job, not run-level, and that a lane rarely needs to force a fresh CI run at all since pushing a fix triggers one automatically — gh run rerun (admin-gated) is reserved for the narrow flaky-retest case, not the default response to red CI.

docs(pr-swarm): note first-failure reaction in skill and bump version — SKILL.md

• Phase 6 (Monitor & Verify) now states that a lane inheriting red CI or open feedback at launch is normal input, not a swarm-level stop condition — Phase 2's independence check is what actually gates the swarm, not the target PR's current health.
• Adds a matching row to the Error Handling table and a bullet to Red Flags.
• Updates the verification-gates.md reference-table description to mention the new first-failure-reaction content.
• Bumps metadata.version 1.0.0 → 1.1.0 for the behavioral change.

test(pr-swarm): cover pre-existing CI failure at swarm launch — evals/cases.yaml

• New case pr_with_existing_ci_failures: a swarm request where one PR already has failing checks and open feedback at launch. Asserts the skill still activates and the response treats the existing state as immediate work, not a blocker.

chore(pr-swarm): regenerate manifest for version bump — manifest.yaml

• scripts/generate_manifest.py output reflecting the pr-swarm version bump. No other manifest content changed.

Verification

• uv run python scripts/generate_manifest.py — regenerated cleanly, diff limited to the version bump line.
• uv run python scripts/validate_evals.py — all packages pass, including the new eval case (73 skills validated).
• uv run python scripts/evaluate_package.py --path skills/pr-swarm — 96/100 (PASS), unchanged from before this change.
• python -c "import yaml; yaml.safe_load(open('skills/pr-swarm/evals/cases.yaml'))" — new eval case parses correctly (6 cases total).

Skill Evaluator Results

Package: pr-swarm (skill)
  D1 Frontmatter Quality:            20/20
  D2 Trigger Coverage:               18/18
  D3 Structural Completeness:        20/20
  D4 Content Depth:                  18/22
  D5 Consistency:                    12/12
  D6 Compliance:                       8/8
  Overall:                          96/100 (96%)
  Status: PASS

Checklist

• SKILL.md has valid YAML frontmatter with name and description
• Skill name is kebab-case, under 64 characters
• Description is 200-1024 characters with trigger phrases and "Use when" clause
• No angle brackets or pushy language in description
• No secrets, credentials, or internal URLs in any file
• Tested locally (scripts/validate_evals.py, scripts/evaluate_package.py)
• All file references in SKILL.md resolve to existing files
• Skill evaluator score is 70% or above (96%)
• No CRITICAL or HIGH findings from skill evaluator