openai/gpt-chat-latest → B (4:1)

Side B fixes a substantive concurrency bug in the RPC handler by ensuring Tokio RwLock read guards are dropped before nested read/write operations, preventing deadlocks during RoomCreate and RoomGrant. It also adds a regression test for private room creation and improves test reliability with HTTP timeouts and optional server log redirection, whereas Side A is primarily a refactoring that centralizes shared test helpers and reduces duplication without materially changing project behavior.