[almog]

That was my thought too, pointing out why NTP smearing might be fragile is a crucial point in any argument against leap seconds, and the reasoning in this post are lacking (regardless of the conclusion's correctness).

My only guess is that because smearing takes place at Stratum 2, if the network partitions part of the NTP servers downstream (Stratum 3+), they'll have an offset as large as T/(17 x 3600) (T being the partition duration in seconds). Yet I guess it must be something else for I cannot see why that won't be tolerable.

More generally AFAIK the NTP RFC does not include smearing period, which is why the best practices are to only use smearing in a well controlled environment rather than on public facing NTP networks, but why is this not something that can be fixed? I'm not sure.