fix: fallback to json_object when json_schema response_format is unsupported

[mjnovice]

Summary

• Adds a try/except fallback in LLMMocker eval LLM calls so that when a model rejects the json_schema response format, it retries with json_object format instead
• Prevents eval failures caused by strict schema validation differences across models

Context

The json_schema response format requires schemas to meet strict validation rules (all properties required, additionalProperties: false everywhere, no unsupported keywords, etc.). While _cleanup_schema handles some of these, certain edge cases can still cause rejections.

In practice, gpt-4o-2024-11-20 enforces strict schema validation and rejects schemas that don't fully comply, while gpt-4.1-2025-04-14 is more lenient and accepts them. This fallback ensures evals work across both models by catching the json_schema rejection and retrying with the simpler json_object format.

Behavior	gpt-4o-2024-11-20	gpt-4.1-2025-04-14
json_schema strict validation	Strict — rejects non-compliant schemas	Lenient — accepts them
json_object fallback needed	Yes	No

Test plan

• Run evals with gpt-4.1-2025-04-14 — verify behavior is unchanged (json_schema succeeds)
• Run evals with gpt-4o-2024-11-20 — verify it falls back to json_object and completes successfully

🤖 Generated with Claude Code