It's unfortunate that the only way to find out the new stuff doesn't work is by putting it live.  Every local test I can do turns out fine, but there are so many thousands of hardware configurations, software configurations, and just plain VOLUME of network activity that you can't properly test in a sterile development environment.

For now, I think I might hang up this attempt at a rewrite and fall back to what I know at least one other pool has done to get better multi-core utilization:  Just run the server process multiple times on multiple ports and have a proxy split the load on the same machine over the different ports.  It feels like a cop out, but I noticed during the live testing that a significant load still hit one core:  The one used for the primary network IO processing, which is extremely difficult to split over multiple cores/threads without significant overhead in mutex locking.  I get the feeling it wouldn't be any more efficient than just running multiple instances of the server.