These NAT discussions are starting to become annoying..... especially
the misinformation that is going around.
There is NO single defination for what Owen calls 'a NAT box'.
Translating IP addresses is a concept and it can occur in routers,
in hosts, in just about any system.
Okay, granted, RFC 1631 considered NAT as an end-user technology,
but has little to do with the more broader question of translation
of IP addresses and checksum calculations in intermediate systems.
RFC 1631 was only _one_ approach to address translation. I refer
the reader to the unfinished draft:
That discusses this in more detail.....