CI: auto-file ai-fix issues from coverage manifest diffs after data updates

[jay9297]

Goal

Close the loop from audit to fix: when a data update lands on dev and the coverage manifests change, automatically file well-scoped issues for the newly-unsupported items so the existing ai-fix pipeline can work through them. This turns "new patch merged" into "triaged work queue" with zero manual effort.

Depends on: #67 (mod-coverage manifest) and ideally #68 (skill-stat manifest). Do not start until #67 is merged.

Background (read first)

• Read CLAUDE.md and the "AI Workflow Escape Hatches" section of README.md.
• This fork already has .github/workflows/ai-fix.yml: adding the ai-fix label to an issue spawns an agent that implements it, runs the test suite, and opens a PR. The missing piece is generating the issues.
• After Tooling: mod-coverage audit — committed manifest of unsupported mod lines + CI regression guard #67/Tooling: skill stat coverage audit — detect gem stats with no mod mapping #68, audit/mod-coverage.txt and audit/skill-stat-coverage.txt are committed manifests that change exactly when parser coverage or data changes.

What to build

A workflow .github/workflows/audit-triage.yml:

Trigger: push to dev, with a paths filter on audit/*.txt.

Steps:

1. Check out with enough history to diff (fetch-depth: 2 is sufficient for the pushed range; use git diff HEAD^ HEAD -- audit/ — handle the first-commit edge case by skipping).
2. Parse the diff with a small script (Python is fine for workflow tooling; keep it in tools/triage_audit_diff.py):
   • Added unsupported/partial lines in audit/mod-coverage.txt → new unsupported mod forms.
   • Added lines in audit/skill-stat-coverage.txt → new unmapped stat IDs.
   • Removed lines are good news — ignore them (the manifest commit itself is the record).
3. Group added mod forms into batches of at most 15 per issue (one giant issue is not actionable; per-line issues are spam). Group by shared leading keyword where easy, otherwise just chunk.
4. For each batch, create an issue via gh issue create (the default GITHUB_TOKEN with issues: write permission) with:
   • Title: Coverage: support N new unsupported mod lines (from <short-sha>) or Coverage: map N new skill stat IDs (from <short-sha>)
   • Body: the exact manifest lines in a table, a pointer to CLAUDE.md rules (ModParser changes need spec/System/TestModParser_spec.lua cases; SkillStatMap changes need --tags data to pass), and a note that src/Data/ModCache.lua must be regenerated and committed.
   • Labels: enhancement and ai-fix — applying ai-fix is what queues the fix agent.
5. De-duplication guard: before creating, search existing open issues (gh issue list --search) for the same manifest lines / short-sha and skip duplicates. The workflow must be safe to re-run.
6. Rate guard: cap at 10 issues per run; if the diff would produce more, create the first 10 plus one summary issue listing the remainder (a giant upstream merge should not file 60 issues).

Acceptance criteria

• Workflow YAML passes actionlint (or at minimum loads without syntax errors on a dummy push).
• tools/triage_audit_diff.py has a unit-testable core: given two manifest texts, returns the batched issue payloads. Include a few test cases (plain python -m unittest style is fine, or a fixture-driven main; this is workflow tooling, not Lua, so the Busted suite is not required for it).
• Running the script against a synthetic diff (5 added unsupported lines) produces exactly one issue payload with the right body content.
• No issue creation happens when the manifests didn't change.
• Re-running on the same commit creates no duplicates.

Out of scope

• Changing ai-fix.yml itself.
• Fixing any mods.