fix(genai): don't mutate caller's tool dict in convert_to_genai_function_declarations

[Andrew Chen (chuenchen309)]

Why

_format_to_genai_function_declaration binds function to the caller's tool dict (or its nested "function") and then normalizes parameters in place:

function = tool["function"]          # (or `cast("dict", tool)`)
function["parameters"] = function.get("parameters") or {}
if not function["parameters"].get("properties"):
    function["parameters"] = {}

So convert_to_genai_function_declarations([tool]) mutates the caller's object:

• an OpenAI-format tool whose parameters lack a properties key has its parameters wiped to {} (losing required, additionalProperties, …);
• a parameterless {"name", "description"} dict gets a parameters key added to it.

Since tool dicts are frequently shared/reused across providers and calls, this silently corrupts the caller's data.

What

Shallow-copy function before normalizing parameters. The lines only reassign the top-level parameters key, so a shallow copy fully prevents caller mutation while leaving the produced FunctionDeclaration byte-identical.

Tests

Added test_convert_to_genai_function_declarations_does_not_mutate_input (OpenAI-format-without-properties and parameterless cases). It fails on main (the input dict is mutated) and passes with the fix. Full test_function_utils.py: 33 passed; ruff and mypy clean.

---

🤖 AI disclosure: found by an automated code-review pass and fixed/tested with Claude Code (an AI coding agent). The regression is pinned by the new test.

[Andrew Chen (chuenchen309)]

CI note: the single red check is langchain-google-genai-us (llm-integration-tests), the live-network integration suite. This change is confined to _function_utils.py — it shallow-copies the tool dict before normalizing parameters, producing a byte-identical FunctionDeclaration (covered by test_function_utils.py, all 33 green, plus the 12 non-integration checks here). So the integration red looks unrelated to this diff. Happy to dig in if a specific integration test is implicated.