> What expectation?  Is this in POSIX?  If so, I might be missing it.

Yes, OK, ignore me, I think it is a POSIX recommendation
(requirement?), although perhaps specifically for the case when the
destination of a rename() call already exists.  I need to do a bit
more reading (although I still maintain my original point about the OS
we all know and love, the NetBSD man page makes no mention of
atomicity).

Cheers,
Rob