feat(skills): land audit-rule-author skill (rule + agentic-audit authoring)

[anmolg1997]

What

Lands Akhilesh's audit-rule-author Claude Code skill (authored on the evidence-synthesis branch) into .claude/skills/. This is the rule-authoring skill — and the agentic-audit authoring guidance — we had been waiting on.

The skill helps author/maintain compliance audit rule config (backend/app/compliance/rules/):

• SKILL.md — 3 paths (add/update rules, new framework, analyze results), evaluation-strategy decision tree, VC-* visual-check tags, merge-behavior reference, preflight checklist.
• references/rule_authoring_playbook.md — authoring conventions & field semantics.
• references/agentic_audit_guide.md — the agentic_audit strategy guide: target/context section design, context_sources, pass_criteria patterns (signature images, summary-vs-subrow tables, NOT_APPLICABLE thresholds, exclusion counter-examples), postpass validation.
• scripts/ — peek_pages.py (OCR + section inspection), analyze_results.py (filter compliance_result.json). evals/ — 3 triggering/behavior evals.

Self-contained against main

All infra the skill references is already merged: validate_cli, registry, profiles, document_profiles.yaml, run_checklist_agentic_postpass.py, and the app/compliance/agentic/ stack. No dependency on the rest of evidence-synthesis (that branch's ~1,600-line rules/segmentation/report changes are a separate review and are intentionally not included here).

Review + main-compat fixes

An independent skill review (plugin-dev:skill-reviewer) found it high-quality and self-contained, with one real drift against main, which I fixed (the only changes to the original):

• peek_pages.py read a per-page classification.yaml that doesn't exist on main (the pipeline writes segmentation.json, page-range DocumentSection). Rewrote the loader to read segmentation.json (legacy classification.yaml kept as fallback). Verified --list-sections renders real sections on a live doc.
• Both scripts now honor AT_STORAGE__BASE_PATH (the pipeline's own storage var) and peek_pages.py gained --data-root, instead of only the hardcoded backend/data/documents cwd path.
• SKILL.md data-location section updated to match.

Follow-ups for the author (non-blocking)

• Description triggers are broad ("analyze results", "look at findings") — could over-trigger; worth scoping.
• SKILL.md lists text_primary/llm_arbitrated under validate_cli, but validate_cli evaluates those text-only (doesn't run the escalate/arbitrate merge) — worth a caveat.

Testing

• Both scripts py_compile clean; peek_pages.py --list-sections smoke-tested against a real doc → correct page-range sections. Skill auto-discovered by the harness (valid structure).