/kb-audit — Audit KB health

1. Acquire runs/.lock.
2. Run lint: python3 scripts/kb-lint.py. Capture errors.
3. Walk the content folders and check (named thresholds defined in CONVENTIONS.md § "Growth rules"):
   • Stale raw drafts: status: raw or status: prospect (for advisors) where date_found or created_at is older than STALE_DRAFT_DAYS.
   • Oversized files: any *.md larger than OVERSIZED_FILE_LINES (use wc -l). Candidates for multi-chapter promotion.
   • Undecided comparisons: decision_status: considered with no commit touching the file in the last UNDECIDED_COMPARISON_DAYS days.
   • Orphan files: content files with no inbound [[<slug>]] citations from any other file. Use grep.
   • Broken wikilinks: [[<path>]] where the target file does not exist. Use grep + file checks.
   • Overdue todos: todos/*.md with status: pending and a deadline before today. Report them as outcome unknown (nobody recorded it), not as "not done".
   • STATE.md age: if STATE.md exists, compare its as-of: date with the newest synthesis/ file and the newest dated content file; flag it when either is newer.
   • Convergence promotion candidates: python3 scripts/kb-convergence-scan.py — any convergence: tag on ≥3 brainstorms means the pattern has escaped its origin and deserves its own questions/ or research-dispatches/ file. Capture the script's "Promotion candidates" section.
4. Compose report: runs/YYYY-MM-DD-audit.md with frontmatter:

   ---
   type: run-log
   kind: audit
   date_found: YYYY-MM-DD
   ---

   Body: sections for each check category; counts + file lists. If zero issues in a category, say so.
5. Stage + commit: kb: audit <date>. Push.
6. Release the lock.
7. Report summary to user: counts per category + report path.

The audit is diagnosis, not remediation. Surfacing the issues is the skill's job; /kb-archive, /kb-promote, etc. fix them.