vLLM Omni #6182 rebuilds a new SamplingParams with the pipeline sampling_constraints applied. That fixes the wholesale replacement in #6177. A caller SamplingParams(seed=1234) used to replace the stage defaults, so stop_token_ids and detokenize=False disappeared. The original report said Gepard then ran to max_tokens: 1000 and produced about 46 seconds of noise instead of about 2 seconds of speech.

The merged test keeps the caller seed and max_tokens, forces detokenize=False and stop_token_ids=[42], and checks the original object is unchanged. It also feeds detokenize=True and stop_token_ids=[7] and expects that silent overwrite.

The docstring says pipeline constraints override caller values. I still prefer a typed error when the caller names a constrained field, so the request cannot look like it succeeded. The overlay keeps the stop token. It does not tell the caller that detokenize=True was ignored.