fix(workflow): preserve explicit null values in typed node inputs

[hulkbig]

Describe the bug

Workflow input normalization drops valid nullable fields by serializing Pydantic models with exclude_none=True. A required nullable value becomes missing before the function validates its input again, and explicit null on a field with a non-null default can silently become that default, including in nested models.

Preserve None during input normalization with a keyword-only exclude_none option on validate_node_data. Its existing True default retains output validation and other callers' behavior. Regressions cover required/defaulted/nested nulls, omitted fields, message and predecessor inputs, and model/list/dictionary serialization.

Reproduction

Define a model with value: str | None and a workflow function accepting node_input of that model type. Run Workflow(edges=[(START, consume)]) through InMemoryRunner with a user message containing {"value": null}, or pass {'value': None} from a preceding node. Before the fix, this raises Field required; a non-null field default instead replaces explicit null. After the fix, explicit null reaches the function as None, an omitted field uses its default, and an omitted required nullable field remains invalid.

The seven public-workflow regressions give four failures and three control passes with the pristine base methods; all seven pass with this change. The original built-wheel replay through InMemoryRunner also passes all seven scenarios. No model API is involved.

Current revision and verification

Current head b80d408ba012cdcc7a786ca6e249d832c7a0a3bc normally merges upstream main 7099ce45fa7b55d708c96e1696aa03f15067756d. Relative to main, the diff is the original four-file nullable-input fix plus a five-line GKE test-fixture correction (146 insertions, 3 deletions across five files).

The mocked Kubernetes API now returns a real V1Job with string owner fields. This restores five GKE test scenarios that also fail on pristine main with Kubernetes 37.0.0. The production GKE implementation and dependency declarations are unchanged; no dependency downgrade or test filter is used.

Full supported-version tox matrix

The original pytest tests/unittests command completed on dad5ca17894ad4a0c34962326db118d24e7f0181 in the repository's isolated tox environments:

Python	Passed	Skipped	Xfailed	Non-strict xpassed	Failed/errors
3.11.17	18,136	90	24	2	0
3.12.15	18,127	91	24	2	0
3.13.11	18,127	91	24	2	0
3.14.8	18,127	91	24	2	0

All use Kubernetes 37.0.0, Pydantic 2.13.5, and the optional Agent Platform / Vertex SDK pair 2.4.0 on macOS arm64. The full command is unfiltered; the two xpasses are non-strict upstream cases and all four tox exit codes are zero.

Final published-head verification

Subsequent normal upstream merges include LiteLLM parsing and the latest A2A, CLI, evaluation-rubric, Runner and cache-analyzer changes through 7099ce45. The authored PR patch, dependency declarations, lock file, tox and hook configurations remain byte-identical to the full-matrix commit.

• On this final head, the complete workflow, utils, A2A, CLI, evaluation and flow suites, runner tests, Model Armor tests, GKE executor tests and affected LiteLLM module pass on each of Python 3.11–3.14: 6,814 passed, 57 skipped, 8 xfailed, with zero failures/errors. Unaffected full-suite evidence above is reused with its exact source scope; this is scoped validation, not a second full-suite run on the final SHA.
• The final isolated offline flit wheel builds, passes CRC verification, and contains byte-identical source for all 12 checked runtime modules, including the five latest upstream modules.
• Mypy 2.4.0 against exact pristine main in the same Python 3.12 dependency environment: both baseline and PR have 808 existing diagnostics, zero new PR errors. This is the local test-extra comparison, not the hosted all-extras/platform matrix.
• The public seven-case InMemoryRunner reproduction on current main gives four expected failures and three control passes; the fix passes all seven. The original four-file pre-commit proof remains applicable by byte-identical patch/config checks. The new GKE fixture passes pinned Pyink 25.12 / isort 8.0.1, compliance, header and diff checks; current guide links also pass.

Required hosted CI (including other platform/API-version matrices), repository-wide hosted pre-commit and maintainer review remain pending. AI assistance was used for implementation and independent review/validation.