Excuse me if what i say below is nonsensical, for I haven't read much about compression techniques and hence these ramblings are just out of common sense.
I think the debate about level (row, page, file) of compression arises when we strictly stick to the axioms of compression which require that all the info that would be needed for decompression must also be presented in the same compressed unit. Can't we relax this rule a bit and separate out the compression-hints into separate file, like the way we have a table data in one file and the positional references [indexes] in another file? will it not eliminate this dilemma about the boundaries of compression? perhaps a periodic auto vacuum like compressor daemon can take up the job of recompression to have the compression-hints updated as per the latest data present in the file/page at that instant. Regards, Samba
