fix(ollama): support gpt-oss thinking parameter

[goutamsidharth287]

Summary

Fix Ollama native structured-output requests for gpt-oss models.

Previously, think was hardcoded to False, which causes structured
fact extraction to fail with gpt-oss models.

This change uses think="low" for gpt-oss models while preserving
the existing think=False behavior for other Ollama reasoning models.

Testing

• Added regression coverage for the Ollama native API path.
• Verifies the think parameter for both gpt-oss and non-gpt-oss models.
• Verifies structured-output format is still included.

Fixes #3246

[nicoloboschi]

Thanks for catching #3246! I opened #3344 as an alternative approach.

Instead of the model-name heuristic ("gpt-oss" in model → think="low"), it makes think configurable through the existing extra_body surface — set HINDSIGHT_API_LLM_EXTRA_BODY='{"think": "low"}'. This avoids guessing the thinking level, gives an opt-out, and keeps model-specific behavior out of the transport layer (it also lets any future native field / options param be set without a code change).

Proposing #3344 supersede this one — but the fix is entirely thanks to your diagnosis here.

[handnewb]

LGTM! Ollama recently added native thinking/reasoning support — gpt-oss parameter enables this. 1-line change in openai_compatible_llm.py + comprehensive test. Good catch!