Allow for custom node type keys

[nzakas]

In ESLint, we are looking at being able to support different formats of ASTs, and as such, we can't guarantee that the node type will always be in the type key. This patch allows a custom nodeTypeKey option to be passed in that will alter be used instead of type when specified.

I wasn't sure what kinds of tests you'd be looking for here, so I added one as an example, but please let me know if you'd like me to add others.

[michaelficarra]

This WFM. It's a bit weird that our "class" matchers (:function, :expression, etc) are aliases for a hard-coded list of estree types. But as long as they don't use these selectors, things should work fine. That begs the question, then: should we even update the class selector implementation to use nodeTypeKey if it's specific to estree which uses type?

[nzakas]

That begs the question, then: should we even update the class selector implementation to use nodeTypeKey if it's specific to estree which uses type?

That's an excellent question. While some of them will definitely not work elsewhere, it seems like :expression, :statement, and :declaration would still most likely work in other ASTs so long as they follow a similar convention as ESTree.

I was going to follow up with another PR that externalizes the class name resolution so we can pass in a custom resolver that will take the place of the internal one. That would solve the problem completely for ESLint.

[michaelficarra]

it seems like :expression, :statement, and :declaration would still most likely work in other ASTs so long as they follow a similar convention as ESTree

I don't really like the thought of someone relying on that. We currently take some shortcuts (like "ends with Expression") to avoid doing a comparison with every possible node type because we assume we know all the possible node types. I don't want this shortcut to become de facto spec.

I was going to follow up with another PR that externalizes the class name resolution

I think this is a good approach. Let's drop the class changes from this PR then until we have such a facility.

Design wise, I'm thinking the config is an object whose keys are class names that extend/override built-in class names and whose values are predicates. Not sure how much context is appropriate for the predicate parameters though. I would probably just start with passing the node for now.

[nzakas]

Sounds good. I've disabled classes when nodeTypeKind is supplied. Once this is merged, I'll get to work on the custom class resolver.

[michaelficarra]

Thanks @nzakas.