[python] Preserve nested union container values

[markstuart-oai]

Why

A Python union containing an array/map of another generated union can silently lose valid data. For example, with Message.content = oneOf(string, array<Part>) and Part = oneOf(Text, Image), Message.from_dict({"content": [{"text": "x"}]}) serializes the content as [null]. The outer union lets Pydantic construct the inner wrapper without selecting its actual_instance. to_json() also exposes wrapper fields for explicitly constructed nested values.

Reproduced on 7.25.0 and current master 6f77724ce5d4a9d61ccb91fde60d80f1c38b0bed, using Python 3.12.13 and Pydantic 2.12.5/2.13.5. The reduction needs no OpenAI extensions, seed, nullable fields, or OpenAPI 3.1. Related: #22261 and OpenAI's serialization report; this is an independent reproduction, not an identification of that reporter's unknown toolchain.

Change

• Decode composed-model leaves inside union arrays/maps through their generated from_dict, then retain the existing Pydantic container validation and union matching rules.
• Leave ordinary models, primitives and enums on their existing parsing path. Routing ordinary models through from_dict would change omitted nullable fields into explicit nulls; regression coverage prevents that.
• Project nested union values recursively for to_dict/to_json, without depending on ApiClient.
• Keep constructors, class names, actual_instance, and legacy public/wire dictionary behavior. The intended observable correction is that modern to_dict returns nested data instead of wrapper objects. This is not a general rewrite of Pydantic model validation and does not fix the separate python-pydantic-v1 generator.

Four implementation files and two fixture/test files are authored changes. The remaining files are regenerated samples, including the two Pydantic-v1 samples that share the fixture.

Validation

• Native PythonClientCodegenTest: 90 passed; CLI build passed.
• Existing Python model/deserialization/serialization tests plus new regressions: 120 passed, 2 existing skips. The exact new regression suite is 14 failed / 14 passed on unmodified master, 28 passed with the fix.
• Default sample mypy: 153 files clean. Legacy dictionary sample: 11 tests passed, 17 files mypy-clean.
• Regenerated all 12 affected sample configurations with bin/generate-samples.sh.
• Independent model-only generation verifies nested unions without supporting API files, enum/primitive handling, ambiguity rejection, typed construction, and omission/null preservation on both Pydantic versions.
• Actual Chat-operation schema slice: 17 cases. Text/image/audio content arrays are repaired; all other captured request bodies are unchanged. Exact signed-64 endpoints survive and adjacent invalid values remain rejected. No network request was sent.

Native command: ./mvnw -pl modules/openapi-generator-cli -am package -DskipTests=false -Dtest=PythonClientCodegenTest -Dsurefire.failIfNoSpecifiedTests=false -Dmaven.javadoc.skip=true (optional remote build cache and build-scan upload disabled locally). No full live-Petstore integration claim.

PR checklist

• Read the contribution guidelines.
• Build the affected generator, run its tests and regenerate affected samples. No generator option/docs change.
• Request Python generator review: @wing328.

---

Summary by cubic

Fixes Python union models losing data when they contain arrays or maps of other generated unions: from_dict no longer produces [null] for nested values, and to_dict/to_json return the nested data instead of wrapper objects.

• Decodes composed-model leaves inside union containers through their generated from_dict, preserving Pydantic container validation and union matching rules.
• Leaves ordinary models, primitives, and enums on their existing parsing path so omitted nullable fields stay omitted.
• Recursively projects nested union values for to_dict/to_json without depending on ApiClient; constructors, class names, and actual_instance behavior are unchanged.
• Adds regression fixtures and regenerated samples across affected configurations; the python-pydantic-v1 generator is not affected.

^{Written for commit 2b275f9. Summary will update on new commits.}