fix: harden tool invocation and function-call validation

[Ridwannurudeen]

Summary

This PR fixes two runtime correctness issues in the recursive tool-calling loop and adds regression tests.

Problems Fixed

1. Tool execution used positional dispatch via *function_args.values(), which is order-dependent and can misroute arguments for multi-parameter tools.
2. Function-call schema validation skipped type checks for falsey argument values (0, False, "", None), allowing invalid calls through.
3. Integer/number checks relied on Python isinstance, which incorrectly treats booleans as integers.

Changes

• Added tool_runtime.py:
  • invoke_tool(...) for deterministic invocation:
    • uses .invoke(dict_args) when available (LangChain tool objects)
    • otherwise uses callable(**dict_args)
  • format_tool_result(...) for JSON-safe <tool_response> payloads.
• Updated functioncall.py to route execution/serialization through the new runtime helper.
• Hardened validator.py:
  • always validates provided args, including falsey values
  • enforces JSON-schema-consistent numeric typing (bool is rejected for integer/number)
  • added explicit error for unsupported schema types.

Tests

Added focused regression coverage:

• tests/test_tool_runtime.py
• tests/test_validator.py

Local run:

python -m pytest -q tests
# 9 passed

Why this matters

These are correctness fixes in the core tool-use path, reducing silent argument corruption and invalid tool-call acceptance during multi-turn function-calling loops.

[Ridwannurudeen]

Maintainer note: this fixes argument misrouting in tool invocation plus falsey-value validation bugs, with 9 focused regression tests to lock behavior.

[Ridwannurudeen]

@interstellarninja @teknium1 when you have a moment, could you take a look at this fix? It targets core tool-use correctness and includes regression tests.