Eval-driven development for LLM features

Treat prompts like code: write the test first, diff the scores, and stop shipping on vibes.