I see the engineering desire to make this work better, but feel that optimizing the failure case (beyond not killing sessions) may be too much complexity for too little gain. --Jari