o MLX

PSA: Qwen3.6 ships with preserve_thinking. Make sure you have it on.

Qwen3.6 introduces preserve _thinking flag to keep reasoning context in-context, fixing KV cache invalidation.