From Feelings to Metrics: Understanding and Formalizing How Users Vibe-Test LLMs