Highlights:
- Migrate Gemini from deprecated google-generativeai to google-genai SDK
- Update default models: OpenAI gpt-4.1, OpenRouter openai/gpt-4.1,
  Gemini gemini-2.5-flash, Ollama llama3.2
- Handle null parts in Gemini responses (thinking model edge case)
- Surface empty response reasons to users across all providers
- Fix Ollama tool args serialization bug
- Add comprehensive LLM provider test suite (158 tests)