openai/gpt-chat-latest → A (3:2)

Side A introduces a substantial architectural improvement by moving vote persistence and ranking recomputation into a dedicated settlement worker, adding cached ranking reads (`ranked_items_cached`), warming the cache at startup, and switching HTTP paths from write locks to read locks. It also removes obsolete demo-counter code and event types, simplifying the application, whereas Side B improves the pair-selection heuristic with more nuanced bridge and rank-based prioritization plus tests, but its impact is limited to recommendation quality rather than core system architecture and performance.