vLLM V0 to V1: Correctness Before Corrections in RL