Validate pagination cursors on array-backed relay connections

[jwils]

Problem

widgetNames(first: 2, after: "not-a-cursor") silently returns the first page with hasPreviousPage: false. The same invalid cursor on a datastore-backed field returns a validation error, so the two paths disagree.

Before ac2c10e, Cursor.coerce_input decoded each cursor and returned nil for an invalid one, and GraphQL-Ruby rejected it at argument coercion time. That commit moved decoding into Paginator so decoding still runs when a project overrides the Cursor type to String (GraphQL only coerces at the scalar level, so a String-typed cursor argument never reaches our Cursor scalar).

Array-backed relay connections never reach Paginator. GetRecordFieldValue and ApolloEntityRefResolver pass their args straight to ArrayAdapter.build. GraphQL::Pagination::ArrayConnection then decodes with Base64.decode64, which is lenient — Base64.decode64("not-a-cursor").to_i is 0, i.e. the start of the collection. before: "garbage" yields an empty page with hasNextPage: true.

Fix

ArrayAdapter.build now decodes after and before before handing them to ArrayConnection, raising the same GraphQL::ExecutionError that Paginator raises. Both paths share a new DecodedCursor.decode_or_raise_execution_error, so they produce an identical message.

This keeps ac2c10e's design — validation lives in the consumer rather than the scalar — so it still works under the Cursor: "String" override.

Also included:

• Removed the &.encode calls in ArrayAdapter.build. They became no-ops when cursor args became strings, and the comment above them was no longer true.
• Stopped array_adapter_spec.rb from decoding cursor args itself. That emulated the pre-ac2c10ea runtime and hid the bug.
• Added an acceptance expectation for a broken cursor on widget_names, covering both the Cursor scalar and the String cursor override.

Verification

I confirmed the acceptance test is a real regression test: with array_adapter.rb reverted it fails in both the snake_case and camelCase contexts, and passes with the fix.

Valid cursors are unaffected. ArrayConnection cursors round-trip through DecodedCursor unchanged (index 1 → "MQ" → "MQ" → 1).

script/quick_build passes: 5259 examples, 0 failures, 100% line and branch coverage, no standardrb offenses, no steep type errors.

Note

DecodedCursor.try_decode is now unused within lib — it went dead in ac2c10e. It is public, documented API, so I left it in place rather than widening this change.