Numpy's implemention of uint64 is 'unpredictable'. It randomly switches datatype to float64 and doesn't allow bitwise operations. This is seemingly just because such operations are not that common so the issue hasn't been fixed.

This issue cost me a lot of time to debeg but I eventually realised that appending an uint would make the entire array change to float64 so when the value was reread it had lost the precision to represent the least significant bits. Annoying isn't it?

uint32 is much more reliable!