The problem with the first is speed and query building ... and the second hides that data away where it can't be easily queried.

Yup. I'm using metadata tables for cases in which most of the information is stored this way, and packed hashes for cases in which most of the data is in normal columns and only a few extras need to be packed away.

While I do think this technique is useful, I haven't yet found a way to abstract this into a general-purpose solution.

Both kinds of implementation are reasonably easy to code as a layer over one of the standard RDBMS-OO mapper frameworks, like Class::DBI or my own DBIx::DBO2.