test: add deterministic Moonshine integration and enforce session resume identity

[Charlie-Wang-03]

Problem

The archive runner had deterministic offline coverage, but it still lacked a regression layer against the real Moonshine runtime.

In particular, resume state only passed session_id back into Moonshine and then checked project_slug afterward. A same-project session created with the wrong Moonshine mode or agent could therefore be accepted by the archive runner.

This matters because the persisted session_id is part of the runner's recovery contract: resuming the wrong runtime identity can silently continue work under incompatible agent semantics.

Contract

A resumable archive session must be bound to the complete runtime identity expected by the workflow:

• mode="chat"
• the target object's project_slug
• agent_slug="moonshine-core"

The runner must fail closed before continuing when a persisted session violates any of those constraints.

The integration harness should exercise that contract through the real Moonshine runtime while keeping model behavior deterministic and offline.

Implementation

• Add a dedicated Moonshine integration workflow on Python 3.11.
• Pin Moonshine to commit 72a64f8984e0e7e1b0a50f6e53e4d00e4fffbb7d.
• Keep Moonshine integration coverage outside the offline test_*.py discovery layer so the existing Python 3.9 / 3.11 suite remains dependency-free.
• Exercise real MoonshineApp, session persistence, skill/tool infrastructure, material staging, verification-event storage, recovery, and archive publication.
• Fake only main-provider and verification-provider responses.
• Resume sessions by passing the complete expected identity into MoonshineApp.start_shell_state(...).
• Translate Moonshine identity mismatches into runner-level RunnerError.

Deterministic integration coverage

The integration suite covers:

1. A verified run through the real Moonshine runtime path.
2. Recovery from persisted verification on a fresh MoonshineApp without another model turn.
3. Wrong-project session rejection.
4. Same-project wrong-mode session rejection.
5. Same-project wrong-agent session rejection.

Regression evidence

Candidate A was originally developed TDD-style: workflow run 33901898148 reproduced the wrong-mode / wrong-agent acceptance bug. After the fix, run 33902182551 passed all five integration tests.

After upstream PRs #1 and #2 merged, the work was rebuilt directly on current main (5f7fbb72) and revalidated:

• Moonshine integration run 33951560203: 5/5 passed on Python 3.11.
• Offline regression run 33951560213: 17/17 passed on both Python 3.9 and Python 3.11.

No real LLM/API credentials or Tavily/MCP network calls are required.

Scope / limitations

This PR verifies the Creative-Intelligence ↔ Moonshine runtime boundary. It does not claim end-to-end model quality, live-provider compatibility, or web/MCP behavior.

The Moonshine commit is intentionally pinned so future runtime changes cannot silently alter this regression baseline.