>>> The alphanumeric versions crop up   in areas with no people or too few to
merit a ZIP code. HH is (or   should be) specific to bodies of water. If this db
contains actual mailing addresses, the ZCTAs can safely be omitted.

I think we're probably safe omitting the alphanumeric codes, but these are
not mailing addresses, they are real estate listings.

>>>Such errors seem to be pretty common: "VA 37642" gets lots of Google hits
with addresses in towns like Gate City and Weber City that are right on the
border with TN, but it seems that 37642 is not actually the ZIP code for those
cities.

It seems that the post office uses the location of the post office itself to
id the state, but for real estate listings we need the actual state the
property is located in. So far it looks like our best option is to create a
limited special case subset for those zip code that overlap state borders
(including one in the southwest that serves residents of three states) based
on communities and hope that rural addresses are always associated with a
community in their state.