fix(workflow): preserve Annotated text and content schemas - #7567

Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

Problem: Node content conversion does not recognize Annotated[str, ...] or Annotated[Content, ...]. A node with a string output schema can parse JSON-looking text such as 123, true, or "quoted" before validation, either rejecting valid text or removing its quotes. Annotated Content schemas similarly parse the message text instead of preserving the message.

Solution: Unwrap Annotated in the two existing annotation predicates, including when they recurse into a union. Pydantic still receives the original annotation and enforces its metadata. The shared correction applies to node input/output validation and the function-node/parallel-worker content converters without changing their interfaces.

Testing Plan

Unit Tests:

• Added three regressions through a real two-node Workflow and Runner: the producer returns Event(output=Content(...)) with an annotated string output schema, and the consumer receives the exact original JSON-looking text.
• Added two annotated-content/optional-content regressions and a constraint-rejection control.
• The same five regressions pass with the patch and fail on the unchanged production baseline. Import paths and baseline source hashes are checked; constructor, collection and import errors are rejected as regression evidence.
• The three owning schema/function-node/parallel-worker test files pass in every supported interpreter.
• The unchanged full tox matrix passes on standard public Ubuntu runners against exact normally-hooked native commit 446f4cacda4785a95d47584de1dfd60176e08d66. Source/import origins, configuration, patch identity and the generated untracked lock are checked before/after. Standalone uv 0.12.23 uses supported TOX_UV_PATH; no project dependencies or test exclusions change.

Normal scoped pre-commit checks and the actual Git pre-commit hook pass in each job. The linked raw logs retain warnings and test-defined skips/xfails/xpasses.

Manual End-to-End (E2E) Tests: The new workflow regression executes the real offline Runner path rather than calling a duplicate converter. Reproduce it with pytest tests/unittests/workflow/test_function_node.py -k annotated_string_output_preserves_json_looking_text. It requires no model, credentials, network endpoint, or ADK Web server. No live-model or manual browser validation is claimed.

Checklist

• I have read CONTRIBUTING.md.
• I have performed a self-review of my own code (AI-assisted; not a claim of human/maintainer review).
• Tests prove the fix through its real consuming path and preserve metadata validation.
• New and existing tests pass in all four full tox environments on hosted Ubuntu.
• I have manually tested a live model or browser.

Additional context

Four production lines in the shared internal predicates; no dependency, public API, serialized-state, prompt, or unrelated formatting change. AI-assisted with OpenAI Codex.