I have found and fixed the problem in Windows.
I had to make sure, when tcp_listen was disabled, to not look at possible TCP sockets when deciding which socket number was the highest socket number to select() with. Took a couple of hours to hunt down and fix this problem.
It's a good thing I keep daily snapshots around; I tried different snapshots until I figured out where things broke, looked at the changes I did (add tcp_listen support, namely), then fixed things so we can disable tcp_listen without breaking Deadwood on Windows.
While debugging this, I set things up so it is now possible to run Deadwood as a non-service as "Deadwood.exe --debug dwood2rc.txt". This is so I can more easily debug Windows-specific problems.
Anyway, I just posted a snapshot where things work in Windows. Now, it looks like the DNS servers for example.com have changed, so I have to revise that particular test. Once I do that, I will run the full SQA battery on the CentOS Linux version of Deadwood, then make the Windows binary of Deadwood (making sure things work in Windows, of course).
I should be able to release Deadwood-2.4.02 today or tomorrow.