fix(langgraph): keep JSON Schema semantics for object values and validate node I/O

[fede-kamel]

Fixes #220, #231, #233, #239, #251, #252. Fixes the LangGraph part of #241 (the Wayflow part lives in wayflowcore).

Root cause

Values flowing through the LangGraph adapter lost the JSON Schema semantics of their Agent Spec properties, and nothing checked them against the declared schemas:

• _build_type_from_schema (adapters/_utils.py) turned a bare {"type": "object"} into an empty pydantic model, which silently stripped every key (Bare object schemas in tool parameters silently strip all keys (empty create_model + pydantic extra='ignore') #220); dropped additional properties even when the schema allowed them (LangGraph does not preserve additional object fields during Flow state propagation #239); and replaced omitted nested properties with None instead of their declared default (Inconsistent handling of default values for optional fields inside ObjectProperty inputs #241).
• LangChain validates tool calls with the generated args_schema and passes getattr(result, field) to the callable, so tools received pydantic model instances for object inputs, and those instances leaked into flow outputs and client-tool interrupt payloads (LangGraph adapter: tools receive pydantic model instances for object inputs, and the instances leak into flow outputs #252).
• _cast_values_and_add_defaults only cast top-level primitives, so StartNode inputs and EndNode outputs violating a nested schema (missing required property, wrong type, unexpected key) were accepted (LangGraph runtime does not enforce nested schema validation on Flow/StartNode input #231, LangGraph runtime does not enforce EndNode output schema constraints #233), while Wayflow rejects them. The integer-cast branch also compared against a Python error message that never occurs, so a non-numeric string for an integer input crashed with a raw int() error.
• ToolNodeExecutor._format_tool_result used if instead of elif for the "no declared outputs" case, so any non-dict/tuple tool return value raised Unsupported multi-output mapping (LangGraph adapter: ToolNode without declared outputs fails with "Unsupported multi-output mapping" #251).

Changes

• adapters/_utils.py: generated object models derive from a new AgentSpecObjectModel. Bare object schemas and typed dictionaries map to Dict[str, ...]; extra keys are kept unless additionalProperties is false (JSON Schema default); declared defaults are applied; to_json_value() converts models back to plain JSON, omitting optional fields that were neither provided nor defaulted. New helpers apply_json_schema_defaults and get_json_schema_validation_errors (uses the jsonschema package that is already a core dependency).
• langgraph/_langgraphconverter.py: tool callables (server, remote, client, and BaseTool func/coroutine) are wrapped with _with_json_arguments so they receive JSON values; user-provided StructuredTools with their own args_schema are untouched (to_json_value only converts models generated from Agent Spec schemas).
• langgraph/_node_execution.py: node inputs and outputs are converted to JSON values, nested defaults are applied, and each value is validated against its property's JSON schema with an error naming the node, the property and every violation. Structured agent outputs go through to_json_value. The zero-output ToolNode branch short-circuits.
• Tests: tests/adapters/test_schema_models.py (unit, no runtime needed), tests/adapters/langgraph/flows/test_schema_fidelity.py (the scenarios of the issues end to end, plus the zero-output ToolNode), tests/adapters/langgraph/test_tool_json_arguments.py (sync, async and client-tool interrupt payloads). One existing test declared an object output but expected the scalar 1 to come out of it; it now declares an untyped output.
• Changelog entry under Bug fixes.

Behaviour changes to be aware of

• Flows that used to run with values not conforming to their declared schemas now fail at the node boundary with an explicit error. This matches Wayflow and the spec's intent; lenient casts (numeric strings to numbers, anything to string) are kept.
• Objects whose schema does not mention additionalProperties now keep extra keys instead of silently dropping them (JSON Schema default).
• The CrewAI adapter shares create_pydantic_model_from_properties, so its nested object models gain the same defaults/extras behaviour. Its tests were not run (CrewAI is not installable alongside the other adapters).

Verification

• New tests: 35 passed. Full suite with the LangGraph, AutoGen, Agent Framework, OpenAI Agents and evaluation extras: 1214 passed, 586 skipped (SKIP_LLM_TESTS=1), up from 1179 on main.
• Public CI steps (black, isort, flake8 + copyright, bandit, mypy, tests/run_tests.sh with core dependencies only) reproduced locally on Python 3.10 through 3.14.
• The same flows were run on wayflowcore 26.3.0 for comparison: outputs and rejections now match across the two runtimes.

Notes for reviewers

• Touches extract_outputs_from_invoke_result, which the fix for LangGraph AgentNode returns tool-output field names instead of values #224 also edits; whichever lands second needs a trivial rebase.
• PR feat(adapters/langgraph): make agent structured output not hang or silently vanish #210 adds a helper right after _build_type_from_schema; no textual conflict.