Keep HTTP Handlers Out of Inference

When inference can take seconds or minutes, an HTTP handler should accept the work and return control quickly. The model belongs behind a queue or worker boundary, not inside the request lifecycle.

That separation also clarifies responsibility: the task owns durable state, the worker owns execution, and the event stream reports progress. An SSE connection can deliver updates without becoming the system of record.

A practical implementation check is simple: restart the browser connection while a task is running. If the work disappears or starts twice, the transport and execution boundaries are still coupled.