Representation geometry shapes task performance in vision-language modeling for CT enterography — ThinkLLM