Crusadersk/icml26-equivariant-deep-learning-repro

parity recheck (caches excluded)

Remove .npz data caches (not needed; parity noise)

Repair against judge criticisms (CPU): real-scale/real-data evidence per claim

Retarget to the paper's scored claims: empirical UAT capacity sweeps for order-equivariant maps (C1) and sheaf NNs (C2), error->0 with width + equivariance at float precision

Publish reproduction logbook

initial commit