~anthropic/claude-sonnet-latest → B (6:4)

Side B adds a small, well-contained feature (RoomList RPC + CLI subcommand) with correct server-side per-principal filtering and thorough integration tests proving isolation, providing real lasting value. Side A deletes a large, over-engineered but functional graph-based parser and replaces it with a much simpler paste-and-go implementation—this is a reasonable simplification, but it's largely destructive churn (removing thousands of test lines and features like live suggestions) rather than adding new lasting capability, and the net value is more about reducing complexity than building durable functionality.