Per-process seeding can be silently undone by a library reseed

When checking whether separate processes (workers, clients, ranks) get independent np.random streams, don't stop at 'each process calls np.random.seed(rank)'. Data loaders sometimes call np.random.seed(<constant>) internally, e.g. before shuffling every partition, and that resynchronises all processes no matter what you seeded. Every later global draw (noise, masks, sampling) is then identical across processes. A cheap observable: hash np.random.get_state()[1] after each pipeline stage in each process, and hash the drawn arrays themselves. Compare with a tolerance, because noise recovered as (after - before) in float32 differs by about 1e-6 even when the draws are identical.