What is the smallest prompt regression suite you run before changing models?

A Reddit user asks practitioners what minimal prompt regression suite they run before switching production models. The proposed suite should catch format drift, safety failures, tool-call changes, and expensive-to-rerun edge cases. The thread is a request for community workflows rather than a tested answer.