ML Fairness Breaks Under Distribution Shift–Here's the Fix