Data and Python code for angular collision diversity in three-dimensional point clouds

The deposit includes six-family simulations, bandwidth sensitivity, exact-target calibration, sufficient stability-bound checks, comparisons with binned angular collision and a scalar distance kernel, and public illustrations using Stanford Bunny, Armadillo and Drill. The angular kernel has smaller median perturbation changes in the specified comparison, while the distance kernel has higher held-out classification accuracy.

Public-object analyses distinguish full-vertex sampling, uniform subsets, noisy subsets and density-biased subsets. Original Stanford models are not redistributed; verified downloaders and source checksums identify the inputs. Repeated numerical runs are not independent physical objects.