Make bare None a partial type again

[ddfisher]

Fixes #3741. Only merge after reading discussion in #3741 and deciding this is the right thing.

[gvanrossum]

This gives lots of new errors in our internal codebase, so I think this PR is dead on arrival.

[ilevkivskyi]

I think a better idea might be to only do this for self.attr = None assignments. IIUC this is how PyCharm does this.

[gvanrossum]

I wish we could do partial types across methods (and include class variables too).

[ddfisher]

Yeah. See my comment in 3741. I'm not actually convinced this is the right idea, but I put it up in case you thought it was and wanted to merge it for 0.521.

I spot-checked some of the internal errors, and a decent number were from self.attr = None assignments, so I'm not sure that would help.

Is this a real user-confusion issue, or is it only theoretical at this point?

[ilevkivskyi]

Is this a real user-confusion issue, or is it only theoretical at this point?

FWIW, this came form @vlasovskikh on gitter.

[vlasovskikh]

@ddfisher I haven't encountered anyone besides me who complains about this problem. In PyCharm we infer Any for such fields instead of None since we don't want to force users into using type hints just to make the warnings go away. But since you already have such warnings for empty lists, it makes sense to handle the None-initialized fields in the same way. It is highly unlikely that somebody wants an attribute of type None, so inferring None makes little sense.

[gvanrossum]

@JukkaL, what do you think?

[JukkaL]

If we change this, it will potentially affect a lot of existing code, so it looks we can't really include the change in the 0.521 release. Let's reconsider this for 0.530 after 0.521 has been released. Just doing this for self.attr = None is a potential compromise, but again we can't include this in 0.521 because it could start generating errors for code that's clean in 0.520.

[ddfisher]

I think we've basically decided we're not going to do this.