The standard version, which would be pretty much what I've been running all along, but with some improvements, of course.
This is the production version of qpsmtpd; the other two are less well tested, though by no means broken. There are several ways to run the trunk code, including pperl, tcpserver, and (my personal favorite) forkserver. I am using the forkserver on two very underpowered servers and have no complaints whatsoever.
The apache module version, which I presume hooks into apache somehow, and saves time by not having to load each time (?) ...
There is at least one person on the list running the Apache version in production; the benefit is indeed that the Apache engine manages all of the process creation bits from a resident master service (but then again, so does the forkserver mode above).
The high-perf version, which I presume also saves time by not having to load each time, but by using something similar to PPerl or FastCGI (???), neither of which I could get to work on my current server.
Not exactly; the high-perf branch (which is intended to eventually replace the standard version) is a more or less complete rewrite of the core connection code. It uses async-IO to support many thousands of simultaneous connections. It is currently being used primarily for spamtrap applications (right Matt?), where the mail is not actually accepted for final delivery.
My recommendation is to get Subversion running and check out the trunk for testing, but if you aren't completely comfortable running not exactly bleading edge but not completely shaken out code, then you should just download the 0.29 tarball. Definitely run in forkserver mode, though; all the speed of pperl, none of the hassle.
HTH
John
