I agree completely with what you say about needing to push
the multi-address complexity to the host. As you kindly
pointed out (and I self-servingly expand on here), this is
an architecture I put forth about a decade ago in a sigcomm
paper (in Zurich, I don't remember the year).
The paper "Efficient and robust policy routing using multiple hierarchical
addresses " was published at the 1991 SIGCOMM, in Zurich. ACM papers used to
be available on line, but it seems that the ACM now wants to enforce pay per
view.
(http://www.acm.org/pubs/citations/proceedings/comm/115992/p53-tsuchiya/)
There is related work on an "Extended Transmission Control Protocol"
available at http://www.chem.ucla.edu/~beichuan/etcp/
But that architecture (hosts having multiple addresses
representing a site's multiple aggregation prefixes and
selecting among them) requires some method of identifying
hosts when they switch from one address to another
mid-connection. I would assume that what people have in
mind for this are the mobility mechanisms? (The alternative
is 8+8 or some variant, which I understand to be contentious
enough that it is a defacto non-starter.)
The rubbing point is that identifying is not quite enough -- you need
"secure identifying" in order to avoid connection hijacking, probably
through some variation of IPSEC. Which brings us back to NATs not being
terribly helpful...