Embedding the Teacher: Distilling vLLM Preferences for Scalable Image Retrieval