openai/gpt-chat-latest → B (3:2)

Side B fixes a concrete correctness issue in vote comparison by making highlight gradients operate per ranking group instead of globally, aligns vote-history visualization with actual left/right vote polarity through new mapping helpers, and adds focused tests for orientation and ranking invariants. Side A adds useful UI enhancements (vote counts beside compare links and an unpin action in the HUD) plus tests, but these are primarily feature and usability improvements rather than a core correctness fix.