I was hoping to see the reasoning_content mess get robustly fixed, but all we got was this doc change: https://github.com/vllm-project/vllm/pull/50624