Your Eval Suite Has a Flake Rate: Gating LLM Changes on Paired, Powered Tests