Discrete Ricci curvature estimation (Ollivier/Forman)

Builds a k-NN/r-neighborhood graph from a layer's activations and computes a graph-theoretic discretization of Ricci curvature at each edge — Ollivier-Ricci (via Wasserstein-1 optimal transport between neighbor distributions) or augmented Forman-Ricci (a combinatorial degree/triangle-count formula) — tracking how this intrinsic, graph-based curvature evolves across depth or training.