Untangling Component Imbalance in Hybrid Linear Attention Conversion Methods