fix: route Claude images to vision-capable models