feat: refresh pi model catalog

[bcho]

What changed

• Refresh pi's generated model catalog from pi-mono in a detached background process.
• Cache the validated catalog under the ai state directory and merge it with existing pi custom models.
• Add atomic cache writes, a single-flight refresh lock, and tests for commented pi model configuration.

Why

The bundled pi catalog can lag newly published models. Refreshing the generated source makes newly available models, including openai-codex/gpt-5.6-terra, available without upgrading pi itself.

User impact

Model-aware commands continue to use the last valid cache immediately. A refresh runs at most every six hours, so newly published models are normally available on the next invocation without blocking the current command.

Validation

• npm test
• npm run check
• Live background-refresh smoke test confirming openai-codex/gpt-5.6-terra is listed