I think I got the point: to decode strings like "\N{CHARACTER NAME}"
PyUnicode_DecodeUnicodeEscape imports the unicodedata module.
If this fails, PyErr_SetString(PyExc_UnicodeError, "some message")
is called.

The exception will eventually be caught by ast_for_atom (in
Python/ast.c), but the exception there is not normalized: type is
PyExc_UnicodeError when the value is a string. Hence the invalid cast.
The exception cannot be normalized there: UnicodeError.__init__ needs 5
arguments.

I think the error was to call PyErr_SetString in the first place.
The attached patch replaces it with PyErr_SetObject and a full
UnicodeDecodeError object.

This deserves a unit test, but I wonder how to reliably make the import
fail.