fix(workflow): exclude thoughts from typed node inputs

[BlueRaddish]

Link to Issue or Description of Change

Problem: A workflow node output can contain model thought parts followed by the final text or JSON answer. Typed Content conversion currently concatenates both. A downstream str function receives the thought text, while a BaseModel function fails validation even when the answer alone is valid JSON.

Reproduction: On main 7d8c9502f924ed5e7a1ab00dc1f5a603e858abc3, run a real Workflow through Runner with a producer returning Event(output=Content(parts=[Part(text="Let me think first.", thought=True), Part(text=answer)])), followed by a typed function. The string case receives Let me think first.hello; the structured case reports model_type validation failure before the function runs. No LLM is needed to reproduce the conversion defect.

Expected behavior: Typed values contain the answer text. Callers explicitly requesting types.Content still receive the full Content.

Solution: Reuse the existing thought-filtering extract_text_from_content helper in shared schema validation and FunctionNode argument conversion. Preserve existing warnings for discarded media parts and the handling of ordinary multipart text.

Testing Plan

Unit Tests:

• Added regression coverage in the existing schema and FunctionNode test files.
• Affected and adjacent test files pass locally: 176 passed on Python 3.13.14.

pytest tests/unittests/utils/test_schema_utils.py tests/unittests/utils/test_content_utils.py tests/unittests/workflow/test_function_node.py -q

The six regression cases fail before the fix and pass after it: string/structured schema conversion with both preserve_content settings, plus real workflow string and BaseModel inputs. Existing ordinary multipart-text, media warning, and schema conversion cases remain green. All native pre-commit hooks pass.

Manual End-to-End (E2E) Tests: The two new workflow cases execute real Workflow, FunctionNode, Runner, InMemorySessionService, and invariant checking through run_workflow. Before: thought text leaks into the string input and valid JSON input fails. After: the functions receive hello and _OutputModel(name='test', value=42), respectively. The setup is deterministic and performs no model or network calls.

Checklist

• Read the contribution guide and performed a self-review.
• Added tests that demonstrate the bug and exercise both affected conversion paths.
• New and affected existing tests pass locally.
• Exercised the change through the real workflow execution path.

Additional context

No public signatures, dependencies, or prompt behavior change. Full multi-version tox and the entire repository test suite were not run locally; the focused suite and native hooks above were run. Separate pending PRs #7430 (null preservation) and #6750 (JSON-mode serialization) touch schema utilities but do not address thought extraction; their actual diffs were checked. The earlier #7452 addresses debug printing and is independent of this workflow conversion fix.

Prepared with assistance from OpenAI Codex.