Review this Python feature-engineering step for a churn model.
The model hits 0.99 AUC on a proper held-out split. Why is that a red flag?