The paper's second contribution is a set of bounded empirical findings, each released with an explicit boundary check rather than as a general graph-RAG superiority claim. Concretely: on LectureBank-Full, diffusion and quotas account for the main gains over flat dense retrieval and survive target-concept-disjoint evaluation; on MOOC-CS, multilingual controls show that encoder–language alignment dominates the graph effect; on QASC, reranking remains stronger than either hierarchical or adaptive traversal; and on an auxiliary HotpotQA stress test, adaptive-depth gains do not transfer cleanly to a denser non-prerequisite graph. Each finding is paired with the specific boundary it does not cross, which is what makes the headline retrieval claims defensible under strict-parity controls.