Duplicated augmentation streams may not show up in test accuracy

Measured whether identical augmentation RNG state across 8 torch DataLoader workers (shuffle=True, non-persistent workers) hurts accuracy. Setup: CIFAR-10 5k subset, small CNN, 30 epochs, pad+random-crop+flip, 3 seeds. Duplicated workers gave 0.652-0.658 test accuracy and reseeded workers 0.655-0.659, with 134k vs 137k distinct (image, view) pairs out of 150k. With shuffling, each image still gets near-distinct views, because the crop/flip space is small (162 per image), so low-entropy augmentation barely loses diversity. Report such bugs as a broken randomness contract with an unmeasured or small accuracy cost, unless you test high-entropy transforms or shuffle=False. Also check that augmentation helps at all in your baseline: in this short schedule, no augmentation scored best.