fix(compilers/openapi): place or keep every union's discriminator

[fuad-daoud]

Summary

Closes #593. Closes #595.

A union's discriminator either went somewhere it was wrong or went nowhere at all.

• openapi: a oneOf beside type: object with a discriminator fails validation #593: a valid document exited 1. A oneOf written beside type: object with a discriminator is degraded: the body lowers to a Model and the union is kept verbatim under Unmodeled. The discriminator, though, stayed on that Model with a mapping to Cat, which was neither a variant (there is no union any more) nor a subtype. pass/validate rejected it with pass/discriminator-missing-variant.
• openapi: a oneOf of one schema and null drops its discriminator when it collapses to a nullable reference #595: discriminators vanished. When oneOf: [X, {type: 'null'}] collapses to a nullable reference, the discriminator beside it went into no field and no Unmodeled entry, with nothing reported.

The rule for a degraded union with a model body

The Model keeps its discriminator only if every target it can route to is a real subtype of that model. A target counts as a subtype when it reaches the model through the allOf base the compiler elects, which is what lowers to Model.Base. The targets are the mapping values, defaultMapping, and, when no mapping entry resolves, the union branches. Otherwise the whole discriminator moves verbatim to Unmodeled["openapi:discriminator"] (degraded_lowering) with an info diagnostic.

• Why it reads raw schemas. pass/validate judges legality from lowered Base links. But component schemas lower in source order, so a subtype declared after its base has no Base yet when the base is lowered. The compiler therefore decides from the raw schemas, mirroring what Base will be. Both declaration orders now produce identical documents.
• One place. The rule lives in the beside closure that every degraded outcome goes through: discriminated, inline branch, oneOf beside anyOf, validation-only branches, and unresolved branches. openapi: a oneOf beside type: object with a discriminator fails validation #593's shape is one of five. The other four failed the same way, and each was probed before the fix.
• True subtypes are unaffected. If Cat and Dog declare allOf: [{$ref: Pet}], Pet keeps its discriminator and validates clean, in either order.
• Out of scope. openapi: keep a discriminated oneOf a union when structural keywords sit beside it #594, keeping such a union a Union, is separate.

Every place a discriminator could be dropped

The discriminator is now kept verbatim, with an info diagnostic, wherever a union lowering cannot place it:

• the nullable collapse, at a component, property, array items, additionalProperties, parameter or response body;
• a null-only union;
• a union beside an enum, a const or a scalar type;
• a $ref-site union, as an alias or on its carrier.

Unions that do place it (oneOf beside anyOf with no body, model-bodied shapes whose targets are subtypes) are unchanged. Docs: ir-design §4.8 and the §14 OpenAPI row.

Test plan

• make gate passes after merging main (build: move to Go 1.27.2, golangci-lint v2.14.0, openapi v1.25.5 #804), with 100% statement coverage, lint (golangci-lint v2.14.0), fuzz and bench smoke.

• Probes with morphic validate: openapi: a oneOf beside type: object with a discriminator fails validation #593's document and the inline-branch, oneOf-beside-anyOf and validation-only shapes all exit 0. On main, each exits 1 with pass/discriminator-missing-variant. openapi: a oneOf of one schema and null drops its discriminator when it collapses to a nullable reference #595's document keeps the discriminator under Unmodeled with an info diagnostic.

• Every degraded shape is tested with pass.Validate in both declaration orders, the subtypes declared before the base. Each has a true-subtype counterpart, with a cmp.Diff between the two orders.

• Each guard was checked by planting the defect and watching the test go red, then reverting:

  Planted defect	What went red
  subtype check always true	the non-subtype tables
  subtype check always false	the subtype-keep tables
  gate bypassed on the other degraded paths	all three shapes
  each collapse or keep site disabled	TestCollapsedUnion_KeepsItsDiscriminatorAtEveryPosition and TestUnplacedDiscriminator_IsKeptAtEverySite
  the moved-discriminator diagnostic removed	the validation-only case
  the moved-discriminator diagnostic duplicated	the two other shapes

• Conformance: two new specs, discriminated-union-beside-model and nullable-union-discriminator, witness tagged-unions, plus inheritance and optionality-vs-nullability respectively. Deleting the discriminator from the second spec reddens its golden. No existing golden changed.

🤖 Generated with Claude Code

https://claude.ai/code/session_016MQFAkYBVaqgSxGwe9ZJSn