The Decision Tree will crash if class proportions differ between train and test

Stratified splitting always produces higher accuracy than random splitting

The test set may contain very few or zero positive examples, making performance metrics unreliable or undefined

No risk — random splitting is always valid regardless of class distribution