Here is what I have found out so far.
Python/bltinmodule.c:builtin_compile takes in a PyObject and gets the
char * representation of that object and passes it to
Python/pythonrun.c:Py_CompileStringFlags. Unfortunately no other
information is passed along in the call, including what the encoding
happens to be. This is unfortunate as builtin_compile makes sure that
the char* data is encoded using the default encoding before calling
Py_CompileStringFlags.

I just tried setting a PyCF flag to denote that the char* data is
encoded using the default encoding, but Parser/tokenizer.c is not
compiled against unicodeobject.c and thus one cannot use
PyUnicode_GetDefaultEncoding() to know what the data is stored as.

I'm going to try to explicitly convert to UTF-8 and see if that works.