Six embedding models rate a question and its opposite as closer than the same question reworded. What that breaks in caches and retrieval, and what fixes it.