Add capability-review skill for in-chat advisory PR review

[enricopiovesan]

Summary

Adds .agents/skills/capability-review/SKILL.md: the AI-advisory review of capability publish PRs (spec 004's duplicate-detection and boundary-quality analysis) now runs interactively in the reviewer's chat session, since ANTHROPIC_API_KEY is intentionally not provisioned for CI (decision-log entries 19 and 25). The skill takes a PR number or contract path, loads the new contract plus compact summaries from the latest index.json release (same input design as scripts/ci/ai_advisory_review.py), applies that script's rubric verbatim (one meaningful business action, not a CRUD wrapper/utility), and reports likely_duplicates and boundary_concerns in the script's exact comment format, stating explicitly that it is advisory and human judgment overrides it.

Docs: README's publish flow and CLAUDE.md's key rules (AGENTS.md is a pointer to CLAUDE.md since #34) now describe the review path as deterministic CI + in-chat advisory skill + human approval.

Governing Spec

no-spec-needed (agent/session tooling, same category as registry-ops from #16; spec 004 remains the advisory function's normative home and its CI-side requirements are untouched)

Project Item

• Closes Add capability-review Claude skill for in-chat advisory PR review #23 (Project 3, In Progress)

Definition of Done

• Skill exists and is invocable in a chat session against a publish PR or contract path
• Output covers likely duplicates (vs the latest published index) and boundary concerns with reasons, and states human judgment overrides it
• Reuses the prompt rubric and output format from scripts/ci/ai_advisory_review.py (named as the canonical wording in the skill) rather than a divergent second prompt
• README.md + CLAUDE.md (canonical target of AGENTS.md) note the review path: deterministic CI + in-chat advisory skill + human approval
• Zero changes to .github/workflows/ci.yml or scripts/ci/ai_advisory_review.py (decision 19 state preserved)

Validation

• Fixture near-duplicate (traverse-starter/document-pipeline vs simulated published traverse-starter/process): flagged as likely duplicate with the specific overlap
• Fixture CRUD wrapper (acme/update-record, raw table-row update): flagged as boundary concern with reasoning
• First-real-use run against Publish seed capability traverse-starter.process 1.0.0 (artifact release + contract PR) #21's publish PR deferred until that PR exists (index-v13 is empty; noted in issue Add capability-review Claude skill for in-chat advisory PR review #23)
• git diff --stat confirms only SKILL.md + two doc files touched