> It exposes too many internals whereas most consumers only need basic functions like _PyThreadState_GET().

That's basically done: pycore_pystate.h no longer defines PyInterpreterState structure. pycore_interp.h must now be included explicitly. It's now included by the following files:

* Modules/gcmodule.c
* Modules/_io/textio.c
* Modules/main.c
* Modules/_threadmodule.c
* Objects/codeobject.c
* Objects/interpreteridobject.c
* Objects/longobject.c
* Objects/moduleobject.c
* Objects/unicodeobject.c
* Parser/listnode.c
* Python/codecs.c
* Python/dynload_shlib.c
* Python/import.c
* Python/initconfig.c
* Python/pythonrun.c
* Python/_warnings.c

I also pushed many cleanup changes. I think that it's now enough, I close the issue ;-)