Thanks Eric for the report. I had tested earlier on XP, but didn't do so
with the latest changes. It appears that a couple of modules didn't
behave well under XP. The latest patch (16) addresses these issues. In
particular, many tests use the presence of os.symlink to presume
symbolic link support, but under Windows XP, os.symlink exists but
raises a NotImplementedError when called. To address this, I added
test.support.has_symlink function which can be used to determine whether
the OS has symlink support (and not just a symlink function).

This raises the issue that it might be worthwhile to remove the symlink
and readlink methods at the os module init time if symlinks aren't
supported on the OS.

For now, this implementation should be adequate to move forward.