The current numpunct.mt test fails on at least one platform because the initialization code and the test code use two different methods for retrieving the numpunct data. The run_test function uses std::localeconv(), and thread_func uses the C++ locale numpunct facet. In most situations this isn't a problem.
Unfortunately, due to incompatibilities between the C and C++ locale details, the thousands seperator is different with the "C" locale. For C++ it is ',' and for C it is the empty string. Here is a quick test to show the problem that the test sees.
if (!std::setlocale (LC_ALL, "C"))
const std::lconv* const pconv = std::localeconv ();
// test against the specialization directly
const std::numpunct<char> np(0);
assert (*pconv->decimal_point == np.decimal_point ());
assert (*pconv->thousands_sep == np.thousands_sep ());
The test should be enhanced to use the C++ locale for both run_test and thread_func.