Consistent Labeling Across Group Assignments: Variance Reduction in Conditional Average Treatment Effect Estimation