> Given that IP only has names for interfaces, rather than machines,

No, IP has no names, only addresses.  BIND has names for all kinds of
things, including addresses (not interfaces), but a name may have more
than one address associated with it.  Also note that interfaces may
have more than one address, and that you can (usually) use any of the
addresses associated with any of the interfaces of a given host to talk
to it (hence I vastly prefer to refer to my host as 192.160.125.65, the
ethernet address which comes from my class C network, rather than as
199.232.59.148, which is the Ultranet-chosen address for my end of the
PPP link).  Unfortunately, because BIND implementations do caching,
names cannot easily have dynamic addresses associated with them.

Granted, it's all a confused mess.  Even though it's not technically
necessary, I still invent a unique name for each interface on my system.
But I still treat my ethernet's address as the "machine" address, and it
works fine.

> I'm having problems talking to some SMTP servers when my hostname is
> "localhost", for example.

Um, yes, I would think so.

> Is anyone interested in developing some kind of standard handling for
> such situations (i.e. periodic change of location, link status, etc.)

Yes, the IPng group is tremendously interested in fixing this whole mess
so that addresses, routing, identification, and naming can all be disentangled.
It's not clear that there are any tremendously satisfactory solutions available
under IPv4.