Ah, but I think you missed  the main point of the new patch.
There is a counter incremented and decremented.  So, the State is set 
only when the first thread enters a guarded section, and then restored 
when the final thread leaves the section.  There is no per-thread 
storage of the previous state anymore, only static variables storing 
the "unmodified state" to restore to when the last thread leaves a 
guarded section.  As long as some thread is in a guarded section, 
assertions and stuff is turned off, otherwise, we keep the old ones.

This is similar in some ways to the special version of 
Py_BEGIN_ALLOW_THREADS (FILE_BEGIN_ALLOW_THREADS, was it?) used in 
fileio.c.