The paper's Conclusion organizes its takeaway as a four-part lesson: (i) what survives — on curated prerequisite DAGs, deterministic diffusion with role-aware quotas drives the main recall gains over flat dense retrieval, and the LectureBank-Full gain survives a target-concept-disjoint control; (ii) what does not survive — on MOOC-CS the dominant variable is encoder/query language alignment, not graph traversal, on QASC reranking remains stronger than hierarchical or adaptive traversal, and adaptive gating is statistically tied to a fixed-depth hierarchical baseline at $R@10$ on all three headline datasets despite pointwise tight-$k$ advantages on LectureBank-Full; (iii) what the boundary probes show — an auxiliary HotpotQA slice reproduces already-strong flat dense retrieval with no distinctive adaptive-depth advantage; (iv) what the method teaches — graph-specific deltas are easy to overstate without strict parity, stronger split audits, separate token-cap diagnostics, external-validity boundary checks, and traceable artifacts.