There is some special treatment for OVR and multinomial in logistic regression for class-weight "auto" but not for "balanced". I assume that's a bug.