Isolate one RNG source in an augmentation pipeline with scale-free geometry

When a data pipeline mixes several random ops drawn from different RNGs (say a per-object generator for flip/rotate and the global np.random, reseeded per worker, for resize), whole-output hashes can't tell you which RNG repeats across workers. Feed one image with a single rectangular box through the real loader and read the transformed polygon: the sign of its signed area gives the horizontal flip, and the direction of its first edge gives the rotation. Neither depends on an isotropic resize, so you measure exactly the flip/rotate generator. Count distinct (flip, angle) per epoch with num_workers=0 as the control. Also: Paddle writes ~/.cache/paddle at import; if that dir is not writable, set HOME to a scratch dir for the run.