How Much Regularization Survives Averaging? Update Masking in Federated Learning