https://developers.google.com/time/smear
https://aws.amazon.com/blogs/aws/look-before-you-leap-the-co...
The smear technique results in seconds that are 11.6μs longer than nominal, and Google makes the case this is within the "manufacturing and thermal errors of most machines' quartz oscillators". So at face value, I'd say software already copes with discrepancies of that order.
However I would infer the maximum delta actually seen by an NTP client is larger and depends on the client's polling interval. e.g. If a machine runs a normal clock locally and synchronizes every 20 minutes (which is probably on the high side), it would see about a 14ms jump on each sync during the 24 hour smear window.
That's assuming the NTP client doesn't already use a smear algorithm locally to correct for drift (Windows doesn't, but I think Red Hat Linux might? Per https://access.redhat.com/solutions/39194 "NTP polling does not directly synchronize the local system clock to the server clock; rather, a complex algorithm calculates an adjustment value for each tick of the local system clock")
Maybe someone more well-versed can elaborate further.
https://www.freebsd.org/doc/en_US.ISO8859-1/articles/leap-se...
That's fine as far as it goes, but make sure never to use a BSD machine as an NTP server or you will cause a lot of confusion for any servers downstream.
(btw, the word "bigotry" has a real meaning and this isn't it.)
But a more important distinction than whether you smear over +/-12 hours or +/-1000 seconds is whether you broadcast a bogus time signal to all your hapless servers, which is what Google does, I think, or whether you fix your kernel and C library so that the system knows the real time but reports UTC-SLS to those programs which use the old APIs and are therefore presumed to be incapable of understanding the real time. For example, clock_gettime(CLOCK_REALTIME, &res) might provide UTC-SLS (or whatever) for naive programs while the same syscall with CLOCK_TAI or CLOCK_UTC might give you the actual real time correctly.