Parse allowed* = null properly in structuredAttrs

[Ericson2314]

Motivation

This is the explicit way to indicate that there is no allow list.

Context

Should help with failures in #14688 CC @Radvendii

---

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[xokdvium]

This seems like a breaking change? `null' was never allowed.

[xokdvium]

This will fail if it's not an object. That's not what the code above used to do.

[Ericson2314]

OK sure we can put that back, this is the only breaking change of that nature?

[xokdvium]

This isn't a regression. I don't see why this needs to change.

[Radvendii]

Looking at this discussion, I'm realising this should really be fixed in nixpkgs not nix. I was reluctant to come to that conclusion because

1. 25.11 is already out, so fixing this in nixpkgs doesn't solve the problem for us.
2. I still don't understand why what's in Nixpkgs works when __structuredAttrs = false, so I thought maybe I was missing something / holding it wrong.

But I think the bottom line is that nixpkgs should strip out allowedReferences in inputDerivation (and probably any time it's set to null)

With regard to (1), I will see if this can be backported, but I don't think it's a recent regression, so we might have to think of something else.

[Radvendii]

To be clear, this might be a worthwhile change to make to Nix as well, but it is a breaking change, so we should discuss it more thoroughly.

Thanks for looking into it @Ericson2314 💜

[Ericson2314]

@Radvendii to be clear, I think there is still a good chance that I am fixing an accidental breaking change from a much older version of Nix.

[xokdvium]

much older version of Nix.

When did this first break? I think I've tested 2.4 and it also disallowed nulls.

[xokdvium]

I've checked the code and the issue has existed since 2.3 at the very least, so we are better off testing current behavior and being done with this. I don't see how changing the semantics of the store layer isn't a rat's nest. We simply must preserve it at all costs when it comes to subtle behavioural differences.

[Ericson2314]

Yeah sounds good. We can devise some sort opt-in at a later point if we want to.