Antoine, some questions about the patch:

First, I think it omits expat_capi from the state. Is that intentional?

Second, I'm not sure if this approach is fully aligned with PEP 3121. A global, shared state is still used. Instead of actually having a different module state per subinterpreter, this patch will have shared state. Another problem seems to be using PyModule_FindModule without using PyModule_AddModule first. 

These problems could be shared to all of Robin's original patches. Of course, there's also the possibility that I don't fully understand PEP 3121 yet :)