GeezerVison

What was wrong. Under schema-merging-behavior: v1, an allOf member that is a $ref is embedded in a struct, whatever its type. encoding/json flattens only an embedded struct, and encodes an embedded field of any other type under a key named after the type. So for an untyped, primitive, array or map member, allOf: [$ref: Name] sent {"Name":"x"} instead of "x". Decoding the value the spec describes failed, or, for a map, left it empty. This happened in every position: models, properties, items, map values, request and response bodies, and the client and strict server. It included the common 3.0 idiom allOf: [$ref: Status, {nullable: true}] for decorating a $ref. An untyped member next to members with properties added a spurious "Untyped" key to every object.

What changed.

• An allOf of one such $ref, next to members that add no fields, is now the member's type, as under v2: type Wrap = Name, One *Name, nullable.Nullable[Name].
• An untyped member places no constraint and carries no value of its own, so it never decides the composition's type: a primitive, array or map member next to it is the composition's type, and a composition of only untyped members is the first one's type (v2 and v3 give any for both). Whenever another member adds fields or brings a type of its own, the untyped member keeps its embedded field, so code that sets it still compiles, but it is tagged json:"-"; it used to add an "Untyped" key to every object.
• A member is classified by the Go type the generator declares for the schema it refers to, and each classification is remembered in the generation context. Without that, a member that is itself an allOf of a $ref is generated twice per link, and a chain of such compositions takes time exponential in its length: 24 links took a minute, and a 200-link chain now generates in a third of a second. A member whose type may have JSON methods the struct inherits keeps today's embedding, because that already works on the wire. Examples are time.Time, openapi_types.UUID, and an x-go-type naming a type or an interface with methods.
• Two guards keep v1's struct where aliasing would break code that compiles today:
  • Alias cycles. Some members would make a cycle of aliases, such as type Root = Arr with type Arr = []Root, which Go rejects. v2 and v3 generate exactly that today; v1 keeps its struct there.
  • Recursion guard. Classification generates the member in a context whose merge frames are copies. Otherwise, generating it could mark a composition in progress as referred back to, which turns a valid spec into "recursive allOf composition at the root … not supported".
• The aggregate matrix harness's echo now fills a Body-wrapped strict envelope (fix: handle untyped strict response bodies #2514), so a shape with an interface-typed body runs at every position. Only the generated harnesses change for the existing shapes.

This changes the Go types of these v1 compositions from struct { Name } to the member's type. Code that sets .Name needs updating. Those types never serialized correctly, so the old shape isn't preserved behind an option.

Not covered: an allOf whose member is an inline array (the matrix's allof_recursive_array) is a different v1 bug, and stays marked broken.

How it's tested.

• Aggregate matrix:
  • allof_array, allof_default and allof_unique_items were marked broken under v1 for exactly this bug. The marks are removed and v1 now round-trips them.
  • New shapes cover a map member, an untyped member, two untyped members, an untyped member next to a map, and an untyped member next to properties, under v1, v2 and v3.
  • All eight v1 cells fail with main's generator.
• servers/strict/responsebody_v1 and the v1 case of TestStrictServerComposedResponseReceiverCompatibility, both from fix: handle untyped strict response bodies #2514, recorded the old v1 shape. They now expect the value itself on the wire.
• pkg/codegen/merge_schemas_v1_members_test.go:
  • covers 13 kinds of member (alias or kept), a chained composition, a 60-link chain (which hangs without the memoization), the alias-cycle guard, context isolation, the untyped-member rules, and the classifier;
  • removing the isolation or the cycle guard fails its test (checked by mutation).
• Every generate target with v1 forced: I generated every //go:generate target under internal/test and examples with a binary built from main and one from this branch. That's 382 targets, 220 of them forced to v1 where the config doesn't pin a version. 372 are identical, and no target changes between generating and failing. The 10 that differ are the eight matrix cells, responsebody_v1, and schemas/nullable forced to v1, whose allOf: [$ref: DecoratableValue, {nullable: true}] goes from nullable.Nullable[struct{ DecoratableValue }] to nullable.Nullable[DecoratableValue], as v2 generates.
• Standard checks: make generate changes no other committed file. make test passes, make lint gives 0 issues, and the changed tests pass under Go 1.25.

Fixes #2590

🤖 Generated with Claude Code