Prefer responses.parse for structured parsing; on clear non-support
(404/NotFound/AttributeError or error mentioning “responses”), fall back
to chat.completions.create with response_format: {type: "json_schema"}
Build JSON Schema from Pydantic v2 (model_json_schema) with v1
(schema) fallback
Preserve reasoning-model temperature behavior (gpt-5/o1/o3) in both
primary and fallback paths
Normalize provider output to a JSON string and wrap with a minimal
response exposing .output_text
Update imports and minor lint fixes
Motivation: Improve compatibility with OpenAI-compatible providers that lack
/v1/responses while keeping the native OpenAI path unchanged.
Notes: No breaking changes; existing tests pass.
998e60baf8