Learning Calibratable Policies with Programmatic Style-Consistency