Fixes That Fail: Self-Defeating Improvements in Machine-Learning Systems