[claude]

Output change: none for any existing schema (no snapshot entry moves; three are added for the new cases). Invalid schemas that master compiled around are now rejected: see "Also rejected" below. Draft until its check passes.

A $ref standing in for a whole allOf, anyOf, oneOf or prefixItems list -- allOf: {$ref: "mixins.json"}, or {$ref: "#/definitions/base/allOf"} -- crashed before anything was dereferenced with TypeError: arr.forEach is not a function (the walk in src/utils.ts assumed an array). Now it compiles when the $ref leads to a list of schemas, and is rejected with the keyword, its position and the $ref when it leads to anything else.

Before / after, for record: {type: "object", allOf: {$ref: "SchemaLists/ObjectSchemas.json"}} (a file holding two object schemas) and value: {anyOf: {$ref: "SchemaLists/lists.json#/scalars"}}:

// before: TypeError: arr.forEach is not a function
// after:
export interface RefInPlaceOfSchemaList {
  record: { id: string } & { createdAt?: number };
  value: string | number;
}

When the $ref leads to something that is no list of schemas:

anyOf at properties/status must be a list of schemas: $ref "#/definitions/status/enum" leads to a list whose item 0 is the string "on".
allOf at properties/person must be a list of schemas: $ref "Person.json" leads to an object (one schema, not a list of them).
allOf at properties/record must be a list of schemas: $ref "#/definitions/base/allOf" leads to one, but the keywords beside it (description) make an object of it (a $ref standing in for a whole list stands alone).

A $ref in a list position leading to a string, null or an empty file was already reported (#877); that report is unchanged and still comes first. A title beside the stand-in is dropped (as beside any $ref); any other sibling keyword is rejected as above.

Also rejected, where master compiled around an invalid schema: a null, false, 0 or "" written in place of a list (allOf: null compiled to the object type; prefixItems: null to unknown[]) now gets the same "must be a list of schemas: found …" message; and a bad stand-in under OpenAPI components/schemas, which master never walked and printed as a literal type (export type Pet = "x" & "y"), gets the message too. Incidentally fixed: a property named allOf under a keyword this tool does not know (contains, propertyNames) crashed master with the same TypeError and now compiles.

How

• traverse visits a non-array value under a list keyword as a schema, so the {$ref} reaches the ref-parser like any other.
• The resolver judges what it led to once dereferencing is done (rejectNonSchemaLists): a walk over the schema positions from the root, throwing for a list keyword whose value is not an array, or is an array a $ref led to whose members are not all schemas. A list written in place keeps today's member-by-member judgement, so this does not overlap Reject a $ref in place of a schema that leads to an array (an enum, a required list, an array document) instead of emitting a tuple #886.
• That walk costs about what the validator's does (2 % of compile() on the Azure template), so it runs only for documents the prenormalizer marked as holding a $ref or a non-schema at a list keyword (ListsToJudge), and for a file that is itself a list of schemas. Real schemas have properties named allOf (Azure's Microsoft.Security does), so a schema object without a $ref is not a mark. Master vs this head, 7 alternating runs, min/median ms: Azure 5209/5263 vs 5154/5256; FHIR 537/551 vs 525/552.

Tests

Seven e2e cases, test/e2e/refInPlaceOfSchemaList.*.ts: external file (a whole-file array and pointers into a file), in-document pointers, a stand-in inside a file that is itself a list, and four rejections (an enum's values, one schema, a sibling keyword, null in place). Six fail on master with the TypeError; inArrayDocument already compiled there and guards the gate.

• CI=true bun run test: 690 pass, 0 fail (master 5caacfc: 686).
• bun run corpus and bun run conformance: same as master (26 × 3 compile and type-check; all 2404 groups match the baseline).
• The 154-row $ref-target probe differs from master in exactly the whole-list rows pos.wholeAllOf.str (TypeError → the Report a $ref whose target is not a schema (empty file, null/string/number document, plain-text YAML) instead of crashing #877 report) and arr.wholeAllOf.schemasDoc (TypeError → string & number); the array probe in its four allOf.whole.*/anyOf.whole.* rows only.
• CLI: json2ts on a schema with the stand-in prints the type; on a bad one prints the message above.

Not changed: a list written in place whose members are no schemas (allOf: [1, 2], a member $ref onto an enum; #886 covers the latter), and a schema object written in place of a list (allOf: {type: "number"}), which still reaches the parser's members.map TypeError as on master.

git merge-tree against every open draft, #879 to #887: all clean.

Changelog

• Fixed: a $ref standing in for a whole allOf/anyOf/oneOf/prefixItems list compiles, or is rejected by position, instead of throwing TypeError: arr.forEach is not a function.