Qwen recommends to preserve_thinking: true for agentic/coding workloads.

Thanks!! I had disabled that previously while debugging, I can confirm this is helping accuracy from what I can tell so far. (And speed since the cache is preserved more often!)