Second independent review: production-path mismatch

The cache correction is sound for string content, but independent inspection found one remaining mismatch. Correct it without expanding scope:

1. append_persistent_model_status writes a list-valued api_content for multimodal content, but the real replay branch in agent/conversation_loop.py only honors non-empty string sidecars. The next request therefore drops the earlier multimodal status, while the test helper _wire_copy incorrectly replays any type. Either support typed sidecars safely throughout the real persistence/replay path or fail closed by not enabling this feature on unsupported content. Do not leave a test model that is more permissive than production.
2. Refactor the sidecar replay decision into production code that the test can call, or add coverage through the actual production request-building path. The contract must exercise the same type check used by conversation_loop.py, including string, list/multimodal, empty, and unsupported values.
3. Update the report with this second review round and exact verification.

Run the focused tests, the same replay/cache regression set, compilation, and git diff --check. Do not change the book, commit, push, or claim downstream task improvement.