Add support for #to_a in ActiveRecord relations compiler

[olivier-thatch]

Motivation

This PR adds support for #to_a to the ActiveRecord relations compiler.

Sorbet can usually infer that #to_a returns the same type as #to_ary, but because Active Record explicitly aliases #to_ary to #to_a (here), Sorbet thinks #to_a is a separate method with no signature and thus returns T.untyped.

Cf. sorbet.run bug repro.

Implementation

Updated every place where a signature is defined for #to_ary to also define the same signature for #to_a, with the exception of the signature that returns NilClass (cf. comment below).

Tests

Updated the existing test for the compiler with the new signatures.

[paracycle]

These relation classes already have to_ary methods defined on them (with the model method returning NilClass so that things like flatten work correctly). See here for reference:

tapioca/lib/tapioca/dsl/compilers/active_record_relations.rb

Line 299 in 7f4c7cd

klass.create_method("to_ary", return_type: "T::Array[#{constant_name}]")

As for to_a, that should already be working since the relations are Enumerable with the correct Elem type, no?

[paracycle]

Here is a sorbet.run demo I created from the generated relation file to show that to_a should work as expected.

[olivier-thatch]

Oh, weird. The whole reason I opened this PR is because I got a T.untyped... I'll try to reproduce my case on sorbet.run.

[olivier-thatch]

It turns out #to_ary returns the proper type, but #to_a doesn't: sorbet.run bug repro

This is because Rails explicitly aliases #to_ary to #to_a (here), so Sorbet picks up on #to_a as a distinct method with no signature (since Tapioca only generates signatures for #to_ary). Without the alias call, Sorbet correctly infers that #to_a has the same signature than #to_ary.

@paracycle How do you think this should be handled? Should the relations compiler use alias too? Or duplicate the #to_ary signatures for #to_a?

[paracycle]

It turns out #to_ary returns the proper type, but #to_a doesn't: sorbet.run bug repro

This is because Rails explicitly aliases #to_ary to #to_a (here), so Sorbet picks up on #to_a as a distinct method with no signature (since Tapioca only generates signatures for #to_ary). Without the alias call, Sorbet correctly infers that #to_a has the same signature than #to_ary.

Ah ok, thanks for the explanation.

@paracycle How do you think this should be handled? Should the relations compiler use alias too? Or duplicate the #to_ary signatures for #to_a?

I find being explicit in RBI files to be better, so I tend to not use alias or attr_xxx methods to define things in RBIs. I think the approach I would prefer is to add to_a implementations right below the to_ary ones in the current code (except for the one that returns NilClass since that's special case for to_ary on the model itself).

[paracycle]

Thanks, this looks great. You'll need to run bin/docs and commit the result, since you've made a docs change as well.