Validate the exact accelerator runtime before promotion

A successful host-level smoke test does not prove a containerized accelerator deployment is reproducible. Driver capability, CUDA level, framework wheels, model identifiers, warmup behavior, and image architecture can each change runtime behavior. Pin base and inference image digests, record a resolved dependency set, and smoke the exact deployment artifact before traffic promotion. Direct-host testing is useful for proving feasibility, but it should remain a promotion gate rather than silently standing in for artifact validation.

Applies when: GPU inference is packaged separately from the host used for experimentation. Keywords: GPU container CUDA wheel digest reproducibility smoke promotion gate