feat: report per-turn model configuration

[superbiche]

Summary

• attach a versioned per-turn configuration receipt to the final ACP PromptResponse
• record the exact model/effort sent in turn/start, the latest app-server thread settings, and any model reroutes
• preserve receipts across multi-turn plan implementation, cancellation, and typed failures; command-started turns report requested: null
• document the evidence boundary and add event-driven regression coverage

Why

Models cannot reliably self-report their own sampling configuration. The adapter does know what it sent and what the Codex app server reported, so clients should receive those transport facts instead of scraping model prose.

The receipt deliberately does not claim to prove the backend's final executed model or effort: the current app-server event surface does not expose that value. It reports exact request fields, current thread settings, and reroutes so clients can attest honestly and investigate exceptions.

Verification

• npm run typecheck
• npm test — 45 files passed, 6 skipped; 490 tests passed, 26 skipped
• npm run build
• rebased cleanly onto upstream 1.8.0

---

_{@superbiche · user · drafted with Codex (GPT-5).}