test: Strengthen cassette parity checks and multi-turn replay

AGENTS.md

@@ -82,6 +82,7 @@ The user's job as a senior software engineer is dependent on you implementing cl
 - All tests must have docstrings that state their intention, goal, purpose, or the regression they guard against. Use phrasing such as "Ensures ..." or "Asserts that ..." to describe the behavior being protected.
 - Add or update regression tests when fixing a reproducible defect. For any public provider behavior with an OpenAI SDK-equivalent contract, include cassette-backed parity coverage that compares Molto against parsed OpenAI SDK responses or events; IR and provider-setting unit tests may supplement this coverage, but must not replace it.
 - Read `tests/cassettes/README.md` before recording, editing, or reviewing cassette fixtures.
+- When a cassette fails validation against the pinned SDK because recorded fields are missing or incompatible, identify the cassette, SDK version, and validation error, and ask the developer for a fresh recording from an approved source. Do not add defaults, coercions, or weaker validation to runtime code or test helpers solely to accommodate the recording, or fabricate fields in the fixture unless asked by the developer. Continue unaffected work and report dependent verification as awaiting the recording. See [SDK schema mismatches](tests/cassettes/README.md#sdk-schema-mismatches).
 - Never run tests in cassette-recording mode. Test runs must replay them only.
 - If any of the committed cassettes are demonstrably wrong, you may update them, but you must explicitly mention the changes to the user in your final review or reply.
 - Never update uncommitted cassettes, always commit them first, and use their diff to guide your changes.