Strip namespace when using policy_class

[Burgestrand]

Closes #689
Closes #694
Closes #666

Using policy_class and namespacing are two different ways of doing the same thing (specifying which policy to use). When both are given, we ignore the namespacing and rely on policy class since it's more specific — when we do so, we need to strip the namespace when passing the record to the policy class.

This is a refactor of #694.

TODO:

• Update the changelog.

[sedubois]

NB: I used to override def policy(record) (to add a namespace to the records and thus load namespaced policies when in an admin context) but that method is not called by def authorize any more, so it makes the overriding more difficult. Now I have to override both def pundit and def authorize.

[dgmstuart]

@sedubois have you tried the suggestion laid out in https://github.com/varvet/pundit#policy-namespacing?

If it doesn't work, please raise an issue.

[adherr]

This appears to be a breaking change, and I'm curious whether or not that was intentional. Our code uses the authorize(record) method below which used to find the policy_class with policy but this now uses policy! in that situation.

Can you please verify that you meant to force a more explicit definition of policy_class? Thanks!

[Burgestrand]

Knee-jerk reaction is that's indeed unintended! We had some other things that were affected by this move that might also be unintended, so I'd wager it's very likely we'll make some adjustments to this code in the near future.

[adherr]

Thanks for the reply and opening a new issue. I'll stick to 2.1.1 until this is worked out. Let me know if there's anything I can do to help.

[Burgestrand]

Hi @adherr! This was a long time ago, but I'm here again for other reasons.

While the previous implementation used to use policy(record), it still relied on the method returning an actual policy because the next call would be policy.public_send(query) which would fail with a NoMethodError if nil was returned for the policy.

The main change is that after this we instead raise Pundit::NotDefinedError as opposed to the NoMethodError.

What part of this behaviour were you relying on that prevented the upgrade?