I routinely generate more than a single gigabit NIC's worth of NFS
traffic.  I very much share SMB's concerns about wrapping ip_id space
in less than 2^16 datagrams.

ip_id collision combined with and packet drop can ucase two fragmented
UDP datagrams to be spliced together into a single datagram. The
Internet checksum is known (self-cite) to not be particularly good
at detecting such splices.

It's also pretty trivial to overload low-to-medium end switches,
causing small episodes of packet-drop.

Under those circumstances, replacing the linear ip_id algorithm with
something more "secure" but which may repeat very quickly, sounds
like a bad tradeoff.

(Yes, there's NFS-over-TCP, but the BSD NFS is not at all well-tuned
to run over TCP at gigabit speeds. Not at all.)