False breaking change when direct array type changes to allOf-wrapped array #920

Follow-up to #887 / #916.

#916 fixed the ClassCastException for allOf-wrapped array → direct type: array. The opposite direction does not throw, but reports a breaking change between two equivalent specs.

Reproduction

Using the fixtures added in #916, with the arguments swapped:

OpenApiCompare.fromLocations("issue-887-2.yaml", "issue-887-1.yaml");

Output:

* Changed property `valuations` (array -> array)

#### Result
API changes broke backward compatibility

Left side:

valuations:
  type: array
  items:
    $ref: '#/components/schemas/Valuation'

Right side:

valuations:
  allOf:
    - $ref: '#/components/schemas/Valuations'   # Valuations is `type: array`

Both describe the same array, so the expected result is no change / compatible — which is what the forward direction already returns after #916.

Cause

SchemaDiff.computeDiffForReal selects the diff-result implementation from the right-hand schema only:

SchemaDiffResult result = SchemaDiff.getSchemaDiffResult(right.getClass(), openApiDiff);

resolveComposedSchema merges allOf members into the parent but leaves the instance as a ComposedSchema, so with the allOf spec on the right, ComposedSchemaDiffResult is selected. Its else branch fires because left is an ArraySchema rather than a ComposedSchema:

} else {
  return openApiDiff.getSchemaDiff().getTypeChangedSchema(left, right, context);
}

That reports a type change even though left.getType() and right.getType() are both array — computeDiffForReal already verified they match before dispatching.

Suggested fix

In ComposedSchemaDiffResult.diff, when left is not a ComposedSchema, fall through to super.diff(...) instead of declaring a type change — the type/format equality check upstream has already passed at that point. The oneOf/discriminator handling stays behind the existing instanceof guard.

More generally, the dispatch being driven solely by right.getClass() makes composed-vs-plain comparisons asymmetric; worth considering whether the result class should be chosen from both sides.

Note

Issue887Test.testDirectArrayToAllOfArrayDoesNotThrow covers this direction but asserts only that no exception is thrown, so it passes against the current wrong output. It should become assertOpenApiAreEquals(DIRECT_ARRAY, ALLOF_ARRAY) once this is fixed.