fix(llm): preserve opaque tool-call fields across OpenAI-compatible turns

## Version

OCR v1.9.4 and current `main` (`b8bee971849d1fee2536e91a6b716168ccb9ed29`).

## Problem

Some OpenAI-compatible providers attach opaque metadata to assistant tool calls that must be echoed unchanged on later turns. Vertex AI's OpenAI-compatible Gemini endpoint returns a thought signature under `tool_calls[].extra_content`. OCR maps the response to its own `ToolCall` with only `id`, `type`, and `function`, so the next request drops `extra_content` and the provider rejects it with HTTP 400.

This affects normal repository tools: turn one can call `file_read`, but turn two fails before the model can continue the review.

## Reproduction

1. Configure the OpenAI Chat Completions protocol against Vertex AI's OpenAI-compatible endpoint with Gemini 3.
2. Run a review that causes a repository tool call.
3. The first completion succeeds and returns `tool_calls[].extra_content.google.thought_signature`.
4. OCR executes the tool but omits `extra_content` from the assistant tool-call history.
5. The second completion returns HTTP 400.

A provider-independent reproduction is a two-request fake OpenAI server: return an unknown `extra_content` object on the first tool call, then require it on the second request.

## Expected behavior

OCR should preserve unknown, valid JSON fields from provider-returned function tool calls and echo them on subsequent turns. Reserved standard fields (`id`, `type`, `function`) must remain owned by OCR, malformed data must not be forwarded, and opaque fields should not be written to normal review/session logs.

## Proposed fix

Keep the unknown tool-call fields as non-serialized raw JSON on OCR's internal `ToolCall`, then use the OpenAI SDK's supported `SetExtraFields` mechanism when rebuilding assistant history. Add an end-to-end adapter regression test that proves an opaque nested field survives response mapping and the next request.

Start at the OpenAI-compatible adapter's response mapping and the code that rebuilds assistant tool-call history; inspect how ToolCall currently retains only standard fields. Add the adapter regression test described in the issue using a two-request fake OpenAI server, and verify that a nested opaque field survives the second request without appearing in normal logs.

Written by the indexing model from the issue text.