The immediate destruction of client resources even under server grab is not new functionality to the epoll port; it behaved the same under the old select-based loop, too. It might be a bug, but that's how it's always behaved.
It shouldn't be too hard to fix if you want to try, either; check client->ignoreCount > 0 when handling X_NOTIFY_ERROR before calling CloseDownClient.
It shouldn't be too hard to fix if you want to try, either; check client->ignoreCount > 0 when handling X_NOTIFY_ERROR before calling CloseDownClient.