fix: handle non-standard responses from OpenAI-compatible API proxies

[zhalice2011]

Problem

When using CodeWiki with an OpenAI-compatible API proxy, documentation generation fails with:

pydantic_core._pydantic_core.ValidationError: 1 validation error for ChatCompletion
choices.0.index
  Input should be a valid integer [type=int_type, input_value=None, input_type=NoneType]

This happens because some OpenAI-compatible proxies (Azure OpenAI, vLLM, LiteLLM proxy, internal corporate proxies, etc.) return choices[0].index as null instead of an integer. The pydantic-ai library validates responses strictly against the OpenAI spec and rejects these responses.

Root Cause

pydantic-ai's OpenAIModel._validate_completion() calls ChatCompletion.model_validate() which requires choices[].index to be int. The library already has a workaround for Ollama's finish_reason: null (line 777 in openai.py), but doesn't handle index: null.

Solution

Created a CompatibleOpenAIModel subclass that overrides _validate_completion() to patch non-standard fields before validation:

class CompatibleOpenAIModel(OpenAIModel):
    def _validate_completion(self, response):
        if response.choices:
            for i, choice in enumerate(response.choices):
                if choice.index is None:
                    choice.index = i
        return super()._validate_completion(response)

This approach:

• Uses the existing _validate_completion hook (designed for subclass overrides)
• Only patches None values, doesn't touch valid responses
• Minimal surface area — single method override
• Follows the same pattern pydantic-ai uses internally for Ollama compatibility

Test plan

• Works with standard OpenAI API (no behavior change for valid responses)
• Works with proxy that returns choices[].index: null
• Fallback model chain still functions correctly
• No changes to call_llm() direct client path (only affects pydantic-ai agent calls)