On a 64-bit OS X build of Python, built with:

./configure --with-universal-archs=64-bit --enable-universalsdk=/ MACOSX_DEPLOYMENT_TARGET=10.5 && 
make

I get the following result:

Python 2.6b2+ (trunk:65805M, Aug 18 2008, 10:59:08) 
[GCC 4.0.1 (Apple Inc. build 5484)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import pwd
>>> pwd.getpwnam('nobody')
pwd.struct_passwd(pw_name='nobody', pw_passwd='*', pw_uid=4294967294, pw_gid=4294967294, 
pw_gecos='Unprivileged User', pw_dir='/var/empty', pw_shell='/usr/bin/false')

Here the pw_uid and pw_gid should presumably be -2, not 4294967294.  I haven't had time to 
investigate properly, but the problem is almost certainly something to do with the fact that 
sizeof(uid_t) is 4 and sizeof(long) is 8 for this build.  (Though interestingly, the configure 
script detects sizeof(long) as 4, which may not be helping matters.)

This problem is causing test_httpservers to fail on 64-bit OS X.

System info: it's a MacBook Pro;  uname -a gives:
Darwin Macintosh-3.local 9.4.0 Darwin Kernel Version 9.4.0: Mon Jun  9 19:30:53 PDT 2008; root:xnu-
1228.5.20~1/RELEASE_I386 i386

It turns out that uid_t (and gid_t) actually *is* an unsigned 32-bit 
integer type on OS X 10.5, so perhaps the pw_uid and pw_gid values are 
correct.  So to rephrase: one or both of the following facts might be 
considered bugs:

(1) On a single machine, the value of pwd.getpwnam('nobody') gives 
different results for 32-bit and 64-bit builds of Python (pw_uid is -2 on 
32-bit, 2**32-2 on 64-bit).

(2) On a 64-bit OS X build, pwd.getpwnam can produce uids and gids >= 
2**31, but os.setuid and friends don't accept values larger than 2**31-1.
There's an inconsistency between pwdmodule.c and posixmodule.c:  pwdmodule 
casts uids to C longs before returning them, while the posixmodule.c 
functions parse an input uid/gid using the "i" format in PyArg_Parse*.

It's the latter problem that's causing test_httpservers to fail on 64-bit 
OS X.

Issue 1747858 looks closely related.  I propose extending the simple fix 
used there, parsing all uids and gids as longs rather than ints.

Fixed via issue5705 which I filed not having seen this one.

I'd completely forgotten about this issue!

Indeed, test_httpservers now seems to be passing on a 64-bit build
on my machine.  Thank you!