> What is the definition of "archive" for this purpose?
I would say any buffer that has certain major mode, e.g. 'archive-mode' or 'tar-mode'.
> Also, when you visit a .tar.gz file, isn't most of the buffer readable
> text (because Emacs automatically decompresses the archive)?
Yes, it appears readable, but dabbrev still comes with various "useful" binary data. Perhaps another approach would be to not search for abbrevs in hidden text, but I don't know how that is implemented internally and if it would be easy enough.
Paul