Reviewing the way things are now, I don't think it's worth rearranging this yet again for 3.7 - instead, I'll just put it back the way I intended it to be (and the way the PEP describes) for 3.8.
The issue is that Victor had good reasons for moving the handling of the "PYTHONCOERCECLOCALE=warn" case to be after the low level C stdio streams were properly initialised, and retaining that aspect of the changes while still moving the C locale coercion back to the beginning of the program means changing the signature of _Py_CoerceLegacyLocale to return a stateless static string to be printed to stderr later.