Remove message.updated -> complete classification in the mobile foreground
SSE monitor. The processor cleanup stamps time.completed on every assistant
message after each LLM step, not just at session end, causing premature
complete events. The sole reliable signal is session.status with type idle.
Remove the redundant status.set(idle) from the processor halt() path. The
Runner onIdle callback already transitions to idle when the loop exits,
so the explicit set in halt was firing a duplicate complete notification
alongside the error notification.
057208e022