Assert matching response shapes in fields_will_merge

[rmosolgo]

Oops, it seems like one part of FieldsWillMerge was missing. Adding it now might break some queries that used to work, so I may have to ensure there's a way to opt out of this change.

From the spec:

SameResponseShape(fieldA, fieldB)

• Let typeA be the return type of fieldA.
• Let typeB be the return type of fieldB.
• If typeA or typeB is Non-Null:
  • If typeA or typeB is nullable, return false.
  • Let typeA be the nullable type of typeA.
  • Let typeB be the nullable type of typeB.
• If typeA or typeB is List:
  • If typeA or typeB is not List, return false.
  • Let typeA be the item type of typeA.
  • Let typeB be the item type of typeB.
  • Repeat from step 3.
• If typeA or typeB is Scalar or Enum:
  • If typeA and typeB are the same type return true, otherwise return false.

https://spec.graphql.org/draft/#SameResponseShape()

In GraphQL-JS:

https://github.com/graphql/graphql-js/blob/f201681bf806a2c46c4ee8b2533287421327a302/src/validation/rules/OverlappingFieldsCanBeMergedRule.ts#L691-L718

https://github.com/graphql/graphql-js/blob/a24a9f35b876bdd0d5050eca34d3020bd0db9a29/src/validation/__tests__/OverlappingFieldsCanBeMergedRule-test.ts#L738-L767

TODO:

• Does this make other tests fail?
  • One now-removed test was added here: Interpreter type at path fix #2004 I think it was a test for code that doesn't exist anymore but ... will it cause a bug to make that fail?
• Is this too big of a change to make at this point?
• What about people who depend on this check not being here?
  • Added a compat layer for deprecation and future removal
• Performance implications?
  • I ran bench:validate and didn't see any regression. The recursion here would only occur for list/non-null types, but in the average case, the overhead is a few simple attr_reader-type method calls.

[gmac]

Adding it now might break some queries that used to work, so I may have to ensure there's a way to opt out of this change.

Unfortunately, yes... this is a breaking change that will start rejecting queries that used to be accepted. Probably the smoothest path to reaching spec would be to make this an opt-in setting on GraphQL::Schema that is off by default for now and activates in the next major release.

[rmosolgo]

opt-in setting

Done in 75c164c with a handler method for tracking occurrences of this problem.