A central lesson of strict-parity prerequisite-QA evaluation is corrective rather than algorithmic: once the encoder, candidate pool, cutoff, matching rule, and split policy are matched across systems and benchmark leakage is audited, some graph-aware retrieval gains do survive, but they are narrower than broad graph-RAG superiority claims in the prior literature would suggest. The contribution is therefore methodological — better controls reshape which graph-RAG claims are supportable — not a new headline retriever.