Thanks, @seberg. I'll take a look and see if we can restore the old behaviour, at least for Python 3.9 and 3.10. I'm not convinced that the new behaviour is objectively wrong, but I agree that the *change* in behaviour is problematic.

I do find it a little odd that we're special-casing integer exponents for complex numbers when no such special-casing exists for floats. It's not purely for optimization (despite what the code might say). And all the messing around with trying to figure out exactly _which_ doubles can safely be cast to long (which was the original cause of this issue) seems like a symptom of a deeper problem, which is that it doesn't make a whole lot of sense to be distinguishing in the first place. But if we're going to change the behaviour, we should only be doing that in 3.11.

OTOH, there *is* real value in z**2 being fast and accurate; we definitely want to keep that. (Perhaps the math and cmath modules should grow a `square` function.)