A team migrates from Word2Vec to BERT embeddings for a legal document similarity system. Their Word2Vec system correctly identified that "bank" in "river bank erosion" and "bank account fraud" were semantically different — they had trained separate downstream models for each context. After switching to BERT embeddings and using cosine similarity directly, the system stops distinguishing these contexts. The team lead says: "BERT embeddings are contextual — they should handle this automatically." What is the precise failure mode?