The interpreter’s trap: how explainable AI launders uncertainty into justification—a socio-legal case study of COMPAS risk assessment