Uninhabited should have all attributes

[ilevkivskyi]

Although this can hide some mypy bugs, TBH this always bothered me as something conceptually wrong. The work on checkmember stuff inspired me to actually try it.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

Tanjun (https://github.com/FasterSpeeding/Tanjun)
- tanjun/clients.py:2294: error: "Never" has no attribute "__iter__" (not iterable)  [attr-defined]
+ tanjun/clients.py:2294: error: "Never" not callable  [misc]

ibis (https://github.com/ibis-project/ibis)
- ibis/expr/types/relations.py:1041: error: "Never" has no attribute "__next__"  [attr-defined]
+ ibis/expr/types/relations.py:1041: error: "Never" not callable  [misc]

prefect (https://github.com/PrefectHQ/prefect)
- src/prefect/server/models/deployments.py:63: error: "Never" has no attribute "name"  [attr-defined]
- src/prefect/server/models/deployments.py:767: error: "Never" has no attribute "model_dump"  [attr-defined]
+ src/prefect/server/models/deployments.py:767: error: "Never" not callable  [misc]

manticore (https://github.com/trailofbits/manticore)
+ manticore/wasm/types.py:56: error: Unused "type: ignore" comment  [unused-ignore]

xarray (https://github.com/pydata/xarray)
+ xarray/core/variable.py:922: error: Unused "type: ignore" comment  [unused-ignore]
+ xarray/core/variable.py:924: error: Unused "type: ignore" comment  [unused-ignore]
+ xarray/core/variable.py:2808: error: Unused "type: ignore" comment  [unused-ignore]
+ xarray/core/variable.py:2810: error: Unused "type: ignore" comment  [unused-ignore]

[sterliakov]

For me "Never" has no attribute "foo" is an early indicator of failed/impossible inference. Allowing arbitrary lookups on Never may prevent that from happening, potentially breaking some funny typing patterns like Never-returning overloads. This will be even more painful for users who don't run mypy with --warn-unreachable enabled.

While I understand your motivation, I think it's completely impractical here. Furthermore, if I'm not mistaken, the "Never" not callable error is no better than "Never" has no attribute "foo" - it's as safe to call Never as it is to access attribute foo on it. Allowing one but not another is an inconsistency and source of confusion. Allowing both can hide too many issues.

[A5rocks]

I think this change is good (and there's like 2 other similar places), especially if combined something that marks any block with a variable bound to Never as unreachable.

Edit: Nevermind I see you address that w/r/t warn-unreachable not being set. I don't think this is actually an actionable error message though? Like, "ok so my variable is Never" doesn't convey much especially to those not familiar with mypy internals.

[sterliakov]

marks any block with a variable bound to Never as unreachable.

Well, then we won't see "Never has no attribute ..." at all, so it doesn't really matter whether such attribute access is allowed?

I see some merit in saying "we do indicate some important issues via [unreachable] errors, please keep them enabled", but it's easier said than done - notably some parts of mypy itself still don't pass --warn-unreachable AFAIC.