The abstract closes with a corrective and narrow field lesson: graph-specific gains in retrieval-augmented generation are easy to overstate when evaluations omit any of four ingredients — (i) matched-interface evaluation (same encoder, candidate pool, cutoff, matching rule, and split across systems), (ii) leakage audits (question-disjoint and target-concept-disjoint controls on canonical prerequisite splits), (iii) separate token-cap diagnostics (token caps reported alongside, not folded into, headline comparisons), and (iv) auditable artifacts (claim-to-artifact traceability so each empirical number can be reproduced). The lesson is framed as corrective rather than algorithmic: it does not propose a new retriever, but specifies the minimum reporting discipline under which graph-RAG superiority claims can be trusted.