CARP: cloud-adaptive robust prompting of vision-language models for ship classification under cloud occlusion | Synapse