openai/gpt-chat-latest → B (5:1)

Side B fixes multiple user-facing correctness issues: it passes the browser agent as an explicit delegate instead of embedding it in DSL text, prevents an incorrect fallback to all items when the sibling pool has fewer than two candidates, simplifies the UI by removing the swap button, and consistently renames the voting route to `/vote` across handlers and tests. Side A mainly removes a now-redundant zero-ratio guard in `apply_vote` and updates the associated test expectations, which is a smaller cleanup relying on existing validation and edge-skipping behavior rather than adding significant new functionality or fixing broader behavior.