Use compact default agent guidance with evaluation evidence

[Raithlin]

Summary

• Make the always-present GHA guidance compact while keeping the detailed skill installed for explicit use.
• Add reproducible agent evaluations and sanitized results to show quality, token use, and command effort.

Changes

• Replace the mandatory skill-loading instruction with JSON capability discovery, uncertainty handling, and dry-run guidance.
• Document the git/gh control: both arms answered 14/14; short-guide GHA used 30% fewer tokens on these GHA-derived tasks. PR inspection favored git/gh.
• Add a held-out local pilot: compact guidance with the skill installed answered 4/4 branch questions with roughly the short-guide token cost. It used GHA on 4/4 separate PR/release checks. A third local task had a bad expected clean-tree answer because the harness added AGENTS.md; its strict failures are retained and excluded from the pilot comparison.

Validation

• GOCACHE=/tmp/gha-go-build GOLANGCI_LINT_CACHE=/tmp/gha-lint-cache make check — lint, race tests, and 95.1% coverage passed.
• make test-eval — 30 tests passed.
• Disposable gha agent install/uninstall smoke test — compact block and detailed skill installed and removed correctly.

Checklist

• Tests cover the changed behavior.
• README, evaluation method, and roadmap updated.
• Structured output contracts remain unchanged.
• Existing write previews and confirmations remain unchanged.