openai/gpt-5.2-chat → B (3:1)

Side B introduces a substantial architectural improvement: a new asynchronous settlement worker with batching (`settlement.rs`), cached ranking via `ranked_items_cached`, and refactors `AppState::record_vote` to decouple disk I/O and score recomputation from request handling. It also removes the demo counter feature and switches to read locks for rendering, improving performance and concurrency. Side A meaningfully fixes OAuth test mocks (e.g., correct request body reading, safer token parsing, try/catch in handlers), but its impact is limited to test stability rather than core runtime design.