Offline evals catch regressions before release. Online evals score real production traffic. How to run both for LLM apps and connect them in one feedback loop.