inputs_from_state typos pass construction when a tool takes no input

Describe the bug

Tool validates inputs_from_state against the tool's actual input parameters — but only when that set is non-empty. For a tool that takes no input at all (a zero-argument function, or a ComponentTool wrapping a component without input sockets), every possible value of the mapping is by definition a typo, yet construction succeeds silently. The mistake then reappears much later, when an agent actually calls the tool, as a ToolInvocationError that no longer mentions inputs_from_state.

Error message

No error at construction time. The failure shows up only when the tool runs:

haystack.core.errors.ToolInvocationError: Failed to invoke Tool `clock` with parameters {'city': 'Berlin'}. Error: no_param_tool() got an unexpected keyword argument 'city'

Expected behavior

The mapping should be rejected while the tool is being built, exactly as it already is for a tool that does take input:

ValueError: inputs_from_state maps 'state_key' to unknown parameter 'not_a_parameter'. Valid parameters are: {'city'}.

An empty set of valid parameters is not "unknown", it is "known to be none" — any reference into it is a typo.

To Reproduce

from haystack.tools import Tool


def no_param_tool() -> dict:
    return {"output": "12:00"}


# 'not_a_parameter' is not a parameter of this tool, and this tool has no parameters.
tool = Tool(
    name="clock",
    description="What time is it?",
    parameters={"type": "object", "properties": {}},
    function=no_param_tool,
    inputs_from_state={"city": "not_a_parameter"},
)
print("constructed without error ->", tool.name)
tool.invoke(city="Berlin")  # ToolInvocationError

The same happens with ComponentTool:

from haystack import component
from haystack.tools import ComponentTool


@component
class NoInputComponent:
    @component.output_types(reply=str)
    def run(self) -> dict[str, str]:
        return {"reply": "Hello!"}


ComponentTool(component=NoInputComponent(), inputs_from_state={"state_text": "text"})
# constructed without error; tool.parameters == {"type": "object", "properties": {}}

For contrast, the identical mistake on a tool that does declare an input is caught immediately:

def weather(city: str) -> dict:
    return {"forecast": city}


Tool(
    name="weather",
    description="d",
    parameters={"type": "object", "properties": {"city": {"type": "string"}}},
    function=weather,
    inputs_from_state={"state_key": "not_a_parameter"},
)
# ValueError: inputs_from_state maps 'state_key' to unknown parameter 'not_a_parameter'.
#            Valid parameters are: {'city'}.

Additional context

Tool._get_valid_outputs() distinguishes "there are no outputs" from "outputs are unknown" with an explicit set[str] | None sentinel that the caller checks with is not None. _get_valid_inputs() returns a plain set[str], and the validation site additionally requires it to be truthy — so the two sibling validations use different notions of "nothing to check".

FAQ Check

• Have you had a look at our new FAQ page?

System:

• OS: Windows 11
• Haystack version: 3.2.0-rc0 (editable install of main @ b717d00), Python 3.13.5