Add conformance test for @final combined with @property

[samwgoldman]

The language of the typing spec indicates that the @final decorator may be used with properties:

The method decorator version may be used with all of instance methods, class methods, static methods, and properties.

I noticed this case while working on Pyre. I observed that Mypy and Pyright also differ in behavior, so it seemed like a good conformance test to add.

See discussion here:
https://discuss.python.org/t/proposal-allow-chaining-final-decorator-when-previous-decorators-return-a-non-function/78918

I also added an example to the spec itself. I think it's a practical example, but also happy to remove it as well, since the existing content makes it pretty clear that the example should be allowed.

[cpython-cla-bot]

All commit authors signed the Contributor License Agreement.

[hauntsaninja]

Looks good, although I think conformance in pyright/qualifiers_final_decorator.toml should be updated to match the new conformance_automated value

[erictraut]

This should say "Partial", not "Fail".

[erictraut]

Thanks for reporting this and adding the conformance test. FYI, I've filed the following bug in the pyright issue tracker: microsoft/pyright#9795.

[samwgoldman]

Thanks for the guidance!

[hauntsaninja]

I guess always a risk changing a pyright conformant value, because odds are 50/50 that Eric will have fixed it before the PR is merged :-)

[JelleZijlstra]

Somebody on Discuss brought up that perhaps you should put @property above @final. Do we need to specify that explicitly?

For classmethod + final, currently both mypy and pyright allow putting the decorators in any order.

My preference would be to make the spec say explicitly that either order is allowed and add conformance tests for it.

[erictraut]

What is the correct behavior if a property getter is marked final but a setter or deleter is not (or vice versa)? I ask because the different ordering would seem to imply different behaviors. Does the final apply to the property object or the method that is wrapped by the property?

[samwgoldman]

I think it's a good point that the spec text here it too specific ("previous" implies an order, should read "other").

On the other hand, I think it gets a bit too precise to enumerate orders. In Pyre (and I suspect Pyright too) the non-conformance arises from attaching the decorator metadata directly to a function type, so a chained decorator that returns a non-function runs into problems.

In terms of typing, the @property and @final signatures compose in either order, so I think it's clear that the definition is legal in either order.

[samwgoldman]

What is the correct behavior if a property getter is marked final but a setter or deleter is not (or vice versa)? I ask because the different ordering would seem to imply different behaviors. Does the final apply to the property object or the method that is wrapped by the property?

This is a great question. I don't feel strongly, but one option:

1. The final-ness applies to the slot on the class, not the type itself (property descriptor or otherwise)
2. If the final property is a descriptor, then the setter should not be called on that type (either through p: Final = Descriptor() or @final-decorated descriptor)
3. [Edited to add] Specifically for the @property descriptor, defining a setter/deleter is an error when final.