To this end, I've considered setting up spam@ and notspam@ accounts on the gateway itself, and having local users send appropriate samples to these accounts, then running sa-learn against these. Does this approach make a great deal of sense?
Only if you can get your local users to send them in a way that you can reconstruct the original headers and body. (ie: regular forwarding won't work here, but forward as attachment might).
Check the wiki, there's a bit of information on this kind of stuff for various kinds of mailclients up there.
