After releasing a Py_DEBUG build to some users who were experiencing 
problems, I noticed a pattern in some of the crash reports I got back:

msvcr90d!_wassert+0xb64
python25_d!FILE_TIME_to_time_t_nsec+0xac
python25_d!attribute_data_to_stat+0x67
python25_d!win32_wstat+0x6f
python25_d!posix_do_stat+0x51
python25_d!posix_stat+0x24
python25_d!PyCFunction_Call+0x65
python25_d!call_function+0x34f
python25_d!PyEval_EvalFrameEx+0x4741

The only way I can see _wassert being hit in FILE_TIME_to_time_t_nsec is 
in the Py_SAFE_DOWNCAST used to downcast an __int64 to int. 
Py_SAFE_DOWNCAST checks that there is equality between the casted and 
non-casted values with Py_DEBUG enabled--maybe in this function we 
should remove Py_SAFE_DOWNCAST?

I can't find a way to see the actual value for "in" before the assert is 
hit--I'm unfamiliar with picking through minidumps with WinDbg, which 
for some reason will show me the stack for these dumps when Visual 
Studio won't. But if I need to investigate more about them I can.

You almost gave the answer in your question - the FILE_TIME is about to be 
converted to a time_t greater than 2**31.

in posixmodule.c::FILE_TIME_to_time_t_nsec(), a comment says:
/* XXX Win32 supports time stamps past 2038; we currently don't */
just before the assert()...

And indeed to reproduce the same crash it is enough to call os.stat() on a 
file with a creation date in 2050 for example.

The attached patch corrects the problem: since VS2008 time_t is a 64bit 
integer; now os.stat() can return times after 2038 on Windows.

This prevents the crash, but the functionality is far from complete: 
os.utime() at least should accept 64bit times.

IIUC, the patch relies on VS 2008. So it is not a candidate for 2.5.3.

Oh, this issue was already fixed by r87666 to fix the duplicate issue #8278.

> Oh, this issue was already fixed by r87666

... in py3k, and then merged into release31-maint (r87668) and release27-maint (r87669).