AI Assistant fails with a generic error for Ollama models without thinking support

Using the AI Assistant with a local Ollama model that does not support thinking (for example LFM2), every request fails with:

The AI provider could not complete the response: The provider stream failed before completion.

The Base URL validation still succeeds, because it only lists models. The AI Assistant enables thinking by default, and the Ollama integration requests thinking mode on every request, which the Ollama server rejects with an error for models without that capability. Turning off "Enable thinking" in Settings makes chat work immediately.

Expected: chat works out of the box with any installed Ollama model, and when a request is rejected, the reported message names the actual cause.