Streaming and the last mile

Streaming makes chat feel alive. It also makes cancellation, backpressure, and 'oops I sent six requests' your problem. Abort on navigation. Debounce the send button. Do not stream secrets. Do not stream chain-of-thought you would not show the user.

Ship with: timeouts, usage logs, a per-user quota, and a degraded UI when xAI is down. That is production. The rest is a demo.