> I think the patch is *NOT* correct.

Can you elaborate?  It works fine for that example on a 64-bit build;  not sure why 32-bit would be any different.

The idea is just that the single cast forces the addition to be done as an addition of integers of type size_t.  Since the integers being added are (a) nonnegative and (b) both fit into a Py_ssize_t, the sum fits into a size_t without overflow, and the result of dividing the sum by 2 fits into a size_t.

(This is all assuming that 2*Py_SSIZE_T_MAX fits into a size_t, but that's a fairly safe assumption in practice.)