The laptop demo chatted fine. After the backend was wired, a thirty-second hang showed up. The UI spun. Logs: the client had already given up; upstream tools were still running; the user retried; both calls billed. The model stretch was fine. What killed the experience was one tool timeout. Serialize several tools and wait each one out to the cap, and “respond soon” on the screen cannot be kept. The incident then gets written as “the model is slow”.