A practical break-even test for plain RAG, GraphRAG, and hybrid retrieval based on question shape, indexing cost, freshness, latency, permissions, and quality.