And although it's bad form to reply to my own mail...

take older binary of screen, and create a screen session.  install
new binary, create and detach new session.  attaching to the 'old'
session works; attaching to the 'new' session does not.
 
However, I don't think it's the compiler - my -current system with
gcc 3.3.6 produces a similarly broken screen. What's worrisome is this:

on older -current:
- both sockets and fifos usable. let's take fifos.
configure: checking select return value...
- select is ok

on newer:
- both sockets and fifos usable. let's take fifos.
configure: checking select return value...
- select can't count

This definitely seems like a regression.