When RAG gives a wrong answer, everyone blames the LLM. Usually the model was fine — retrieval handed it garbage. Here's how to catch it before users do.