From the dense seeds, the bidirectional-diffusion retriever runs deterministic breadth-first traversals in the prerequisite DAG, one upward toward ancestors and one downward toward descendants, each truncated at the per-query depths $d_\uparrow$ and d_downarrow chosen by the adaptive controller. Because the traversals are deterministic BFS rather than a stochastic random walk, the truncated ancestor and descendant path sets $\mathcal{P}_\uparrow(v)$ and mathcal{P}_downarrow(v) that feed the path-score formulas are reproducible given the graph, the seeds, and the chosen depths.