Review this Python feature-normalization function.
Downstream models trained on this perform worse than on raw features. What's the bug?