I can reproduce the problem (or at least get the same symptom) by doing
this (in 2.4.6, 2.5.4 and 2.6.2):

import subprocess, signal
signal.signal(signal.SIGCLD, signal.SIG_IGN)
subprocess.Popen(['echo','foo']).wait()

The echo command completes, but the subprocess command throws the "no
child" exception. It seems like the subprocess command should either:

- detect that SIGCLD is not set correctly and throw a more informative
exception before running the command.
- override SIGCLD during the running of the sub-command. Not sure what
the UNIX correctness implications of this are.
- or allow the child to zombie without throwing an exception. The fact
that the programmer has overridden SIGCLD sort of implies that reaping
of zombie children has been switched off.

I don't have good enough understanding of the underlying implementation
to know if this is a reproducer as requested or if this should be a new
bug. Please advise and I will file a new bug if requested.

This is a follow-up to trying to resolve this problem in the PEP 3143
python-daemon module. See this thread:
http://groups.google.com/group/comp.lang.python/browse_thread/thread/9a853d0308c8e55a