Represented but misread: Post-training changes how language models read out what they represent — ThinkLLM