[ML] Zero-allocation trained-model deployments are silently dropped from cluster state during ML node shutdown, producing persistent 404 Unable to find model deployment task