Besides, the assumption that a middlebox sees all of the host
communications is often wrong today, and will likely become wrong even
more often in the future.
That's true, and in most cases a host implementation would be better
than a middlebox implementation. Still, we shouldn't ignore the input
from the enterprise guys who really don't want to do traffic
engineering on a host-by-host basis, to name just one downside of
every host doing multihoming autonomously.
So unless having potential middlebox implementation compatibility is
a great hardship we shouldn't break it.