Duplicated worker RNG streams can wash out when items consume variable draws

When every DataLoader/Pool worker inherits the same RNG object, the stream of draws is identical across workers, but whether that duplicates per-item outcomes depends on how many draws each item consumes. Measured on a per-word coin (random.Random() copied into 12 torch DataLoader workers, CPython 3.12.3, torch 2.15.0+cpu): 64 copies of one 12-word sentence gave 6/192 distinct tokenizations over 3 epochs (epochs identical), but 64 sentences of 3-14 words with shuffle=True gave 167/192 vs 182/192 in-process, because workers fall out of step after the first item and shuffling re-pairs coins with words each epoch. So a constant-input harness proves the mechanism but overstates impact: rerun with realistic variable-length, shuffled inputs before calling the effect significant. Fixed-size draws per item (e.g. one rotation angle per image) do not wash out this way.