feat!: rewrite the skill around measured evidence

[AllanOps]

Phase A measured twelve practices against seven neutral baselines. Six were applied reliably with zero prompting — including complete idempotency on money-moving endpoints in 6 of 6 runs — and are now cut. The skill ships only what the model demonstrably misses, in two parts: habits absent from every baseline (API versioning, feature flags, circuit breakers) and habits present only when the task cues them (deadlines on every I/O including the database, work off the request path, business-metric counters, tombstone-aware reads).

• SKILL.md: 254 -> 137 lines. Description rewritten to pure triggers; the old one enumerated all twelve practices, creating a shortcut agents took instead of reading the body. Each habit gains a "Skip it when" conditional.
• README: repositioned around the measurement, scorecard published.
• evals/: protocol, the three neutral scenario prompts, results, and the two places our own methodology was wrong.
• CONTRIBUTING: a habit now ships only with evidence the model omits it.
• Manifests and issue templates realigned.

BREAKING CHANGE: field-limited responses, validation at the boundary, idempotency keys, graceful degradation, and intention-revealing naming are removed from the skill. Pin v1 if you depend on their presence.

What this changes

Type of change

• New trick
• Fix / sharpening of an existing trick
• Docs / CI / tooling

Checklist

• Commit messages follow Conventional Commits (feat:, fix:, docs:…)
• Ran node scripts/validate-repo.mjs and node scripts/scan-content.mjs locally — both passed
• Content follows the trap-vs-fix format in CONTRIBUTING.md

Security checklist (required for any change under skills/)

• No new external URLs — or they're added to scripts/url-allowlist.json with a justification
• No invisible / bidirectional Unicode or encoded blobs
• Nothing that would direct an agent to fetch or run remote content, or to read or transmit environment variables or secrets