Racing Thoughts: Explaining Large Language Model Contextualization Errors