ScopeProvider sometimes returns nodes that don't appear in the CST

[zsol]

Consider this example code:

def f() -> "None":
  pass

Scope analysis will return an Access for a node like Name(value="None"), but there's no such CST node in the code. This is a problem because users of LibCST would expect to query metadata (for example the position) of nodes returned by scope analysis.

Here's some code to demonstrate the problem:

import libcst as cst
m = cst.metadata.MetadataWrapper(cst.parse_module("def f() -> \"None\": pass"))
pos = m.resolve(cst.metadata.PositionProvider)
global_scope = list(filter(None, m.resolve(cst.metadata.ScopeProvider).values()))[0]
alien_node = list(global_scope.accesses['None'])[0].node
pos[alien_node]  # BOOM

This produces an exception like the following:

KeyError: Name(
    value='None',
    lpar=[],
    rpar=[],
)

[zsol]

I suspect the issue exists since #373, where we construct a new CST for each string annotation, and then nodes from this constructed CST are exposed by ScopeProvider. I think we should instead expose the original string annotation CST nodes as the referrer nodes.

[jimmylai]

I suspect the issue exists since #373, where we construct a new CST for each string annotation, and then nodes from this constructed CST are exposed by ScopeProvider. I think we should instead expose the original string annotation CST nodes as the referrer nodes.

Nice catch! The change does expose new CSTNode which not existed in the original node.

The suggestion makes sense for simple type annotation with a type name.

For the case that the String is an expression with multiple parts, is exposing the entire string annotation still useful?
E.g. a: "Optional[A, B]" = None

[zsol]

I think it's still better than nothing. Not to mention it would be great to keep the Access.node attribute as not Optional