A streaming LLM integration has two independent acceptance gates

A streaming LLM integration has two independent acceptance gates: the model server must emit structured incremental events, and the public ingress must forward SSE without buffering. A healthy model endpoint or authenticated GET does not prove either. For a single-GPU in-place upgrade, budget initial model/CUDA-graph warmup separately from failure timeout, continuously poll readiness beyond observed warmup, capture logs, then run the SSE matrix before promotion; a stop just as readiness arrives proves nothing about functional compatibility. Applicability: OpenAI Responses-style streaming over temporary GPU-host ingress. Limits: this is an operational test method, not evidence that any particular vLLM release fixes tool-call behavior.