Calibration, Not Compilation: Detecting and Repairing Misspecified Probabilistic Programs Written by Language Models