After a data-pipeline refactor, your CIFAR-10 classifier reaches 100% training accuracy, but validation accuracy sits near 10% from the first epoch to the last.
Training loss took noticeably longer to start falling than before the refactor, and the model and hyperparameters are unchanged.