Ollama reasoning is dropped: responses render empty with gemma4

Using Ollama as an openai-compatible provider, models that emit reasoning (gemma4:26b here) show a thought block in the TUI and then no response text at all.

Ollama puts reasoning on a field called reasoning. The OpenAI Chat protocol in packages/llm/src/protocols/openai-chat.ts only reads reasoning_content, so the reasoning is dropped and the assistant message comes back empty.

A streamed chunk from Ollama looks like this:

{"choices":[{"delta":{"content":"","reasoning":"The user"}}]}

Steps to reproduce:

1. Run gemma4:26b under Ollama 0.33.2 (the OpenAI-compatible /v1 endpoint is on by default)
2. Point an openai-compatible provider at http://<host>:11434/v1
3. Ask anything that makes the model think, the thought appears, the answer doesn't

OpenCode version: dev (also seen on 1.18.20)
Operating System: macOS 26.3