The paragraph positions this paper's contribution as complementary in scope to the recent graph-RAG evaluation-first lineage (Zeng et al. 2025; Han et al. 2025; Ju et al. 2025). On top of the matched-interface comparison among flat dense retrieval, graph baselines, reranking, diffusion, and best-first traversal, the paper adds four specific layers: (i) leakage-controlled prerequisite QA, (ii) question- and target-concept-disjoint split audits, (iii) separate token-cap diagnostics, and (iv) a claim-to-artifact traceability layer over each headline result. The framing is explicitly additive: these four layers extend the eval-first contract rather than replacing it.