Build a production-shaped canary corpus for LLM pipelines with stage-level contracts, noise floors, and release gates that catch semantic regressions.