Uncontrolled Lexical Exposure Leads to Overestimation of Compositional Generalization in Pretrained Models