openai/gpt-chat-latest → B (9:1)

Side B fixes a substantive concurrency bug by shortening the lifetime of Tokio RwLock read guards before later read/write operations, preventing deadlocks in RPC handlers, and adds an integration test covering room creation. It also improves test infrastructure with process/logging and HTTP timeout changes and updates authentication expectations, whereas Side A only adds a randomized test validating existing ranking behavior without changing production functionality.