Why LLM-as-judge fails for code evaluation. Here's what works.