Yes I would assume that the NTP daemon would handle a leap second arrival with guarantees of (1) gradual application over a longer period and (2) never moving the clock backwards, only slowing it a little.

I wonder if all NTP implementations don't follow those guarantees?

Oh and another app that hates clock jumps used to be sshd; it would just bail out and drop all connections. We found that out while chasing ANOTHER bug in SunOS on a T4: they didn't have it mutexed right so it possible to read its RTC register while it was in the middle of getting updated so the client would read a garbage time. We chased NTP for a week before realizing it was the kernel.