Sounds like a good idea to me. Although all major platforms support IEEE
Python should guarantee a stable behavior on all platforms.

The lines should be a safe replacement of the downcast:
double fromd;
float tof;
tof = abs(fromd) >= FLT_MAX ? FLT_MAX : fromd;
tof = (float)copysign((double)tof, fromd);

By the way the release notes of Python should mention the our work on
better IEEE-754 and numeric support in bold letters. "Python 2.6 - now
with even better support for professional math in the core". It's a good
advertisement.  :)