openai/gpt-chat-latest → A (4:1)

Side A fixes a concrete concurrency bug by ensuring Tokio RwLock read guards are dropped before nested read/write awaits, preventing deadlocks in RPC handlers for room creation and grants, and adds an integration test covering room creation. The remaining changes improve test reliability and align expectations with stored usernames. Side B is primarily a schema/UI cleanup that removes legacy GitDiscovery projection paths and requires evidence-only records, updating tests accordingly; while valuable for maintainability, it mostly deletes compatibility code rather than fixing a core runtime issue.