Seeds Fixed, Validation Accuracy Still Differs Between Runs

A CNN training script calls torch.manual_seed(42) and nothing else. The config has cudnn.benchmark = True, a DataLoader with num_workers=8, a NumPy-based augmentation class that creates np.random.default_rng(0) inside __init__, and a data-parallel-free single A100. Five identical runs give final validation accuracies 91.2, 90.8, 91.5, 90.9, 91.1 (%). A colleague claims a new augmentation "improves accuracy by 0.2 points" because the mean of five runs went from 91.1 to 91.3.

1. List every plausible source of run-to-run difference in this setup, grouped by what you can fix and what you cannot.
2. Show exactly what you would change to make the run (near-) deterministic, and the cost.
3. Is the 0.2-point claim supported? Compute the spread of the five runs and say what you would do.
4. What does determinism buy you, and when should you not enable it?