I haven't heard the term IPTC, XMP in relation to metadata. If you literally mean the "keyword" portion of a file, then this is a legacy way of adding metadata.
Keywords are searchable, but often won't import into the newer systems unless you tell it to bring it in. It is critical that any metadata is embedded directly into the files XMP core.
Sizes of each catalog speak to how bulky the database is for each "section". The bigger the catalog, the more like the possibility of crash and burn unless the db is mirrored to an SQL db.
Cumulus is about to integrate with CS3. This presents a direct route into the world of hard core metadata driven catalogues.
As for LIghtroom and Apeture, be very wary of relying on the internal "library" functionality. You lose the library, you lose the intelligence with it.