Picked up via python/cpython#31. This is more of a design mistake than a bug, as it's not clear who made the right decision

# a c long on this platform is indeed 8 bytes
>>> np.dtype('l').itemsize
8  
>>> struct.calcsize('l')
8

# but when an endian mark is specified, l always means 4 bytes to the struct module
>>> np.dtype('>l').itemsize
8
>>> struct.calcsize('>l')
4  

Note that this only occurs on platforms where sizeof(long) == 64, ie not 64-bit windows.

Do we care about consistency with struct? If not, then we need to document that we are not consistent with it.

For searchability: PEP3118