openai/gpt-chat-latest → A (9:1)

Side A fixes a real correctness bug by moving the zero-ratio guard before `ensure_item` and `voted_pairs.insert`, preventing ghost items and incorrectly recorded voted pairs. It also updates the test to verify that no items, edges, or voted pairs are registered, whereas Side B is a large refactor/feature addition with dependency changes and code movement but no similarly clear, focused correctness improvement.