I've spent the day ironing out little weirdnesses in the Oracle version, in preparation for the beta release ... it will happen!
Anyway, I'd make my decision as to OpenACS 3 vs. 4 based on:
1. The packages I need. Some OpenACS 3 packages have, as of yet, no OpenACS 4 analog.
2. Long-term goals for your site. OpenACS 4 is the future and our community seems intent on focusing efforts there. I really mean it that way, OpenACS 3 is there for bug fixing and enhancement for those who interested in that path rather than the OpenACS 4 path. But general community interest does seem to mostly lie in pushing the OpenACS 4 toolkit forward. But if you need some of the packages in OpenACS 3.2 that have no OpenACS 4 analogs, and don't expect to be forever seeking additional functionality as web technology shifts, OpenACS 3.2 may be right for you.
3. OpenACS 4 supports Oracle as well as Postgres, so if you strike it rich you can buy a Really Expensive Oracle Enterprise Edition license and impress women! Yes! Really! NOT! Damnation, where is justice in this world! :)
I wouldn't worry about stability at this point. As Gilbert points out the core datamodel's stable for our OpenACS 4.5 release and overall the toolkit's in pretty good shape if you accept it for what it is (a port of aD's work to a framework supporting multiple RDBMSs along with a bunch of bug fixes, performance enhancements, and minor improvements but lacking any large-scale rewriting to get rid of package deficiencies).
