Failing within a stream

Because agent responses are streamed (Chapter 8), failures often occur after you've started responding. Handle it:

• Wrap the orchestration so that on error you can still emit an error event and a terminating sentinel, so the client shows a clean failure rather than hanging.
• Don't lose persisted progress: the pre-saved user message and any incrementally-saved results survive.
• Clean up server-side resources (abort the in-flight model call) when the client disconnects, so abandoned turns stop burning tokens.