Summary

Fixes #4775

When an LLM hallucinates a tool name that doesn't exist in tools_dict, _get_tool() raises ValueError. Previously, on_tool_error_callback fired immediately — before before_tool_callback and outside the OpenTelemetry tracer span. This caused plugins that push/pop spans (e.g. BigQueryAgentAnalyticsPlugin's TraceManager) to pop the parent agent span, corrupting the trace stack for all subsequent tool calls in the session.

Root Cause

The callback lifecycle contract is:

before_tool_callback → (tool execution OR on_tool_error_callback) → after_tool_callback

For hallucinated tools, the old code path was:

on_tool_error_callback → (return or raise)  # before_tool_callback never called!

This violated the lifecycle invariant — plugins that push a span in before_tool_callback never get to push, but on_tool_error_callback still pops, corrupting the stack.

Fix

Move the ValueError handling inside _run_with_trace() so that:

1. before_tool_callback always fires first (balanced push)
2. The error is surfaced within the OTel span context
3. on_tool_error_callback fires after before_tool_callback

Applied to both handle_function_calls_async and handle_function_calls_live code paths.

Testing

• 2 new tests in test_plugin_tool_callbacks.py:
  • test_hallucinated_tool_fires_before_and_error_callbacks: Verifies callback order (before → error)
  • test_hallucinated_tool_raises_when_no_error_callback: Verifies ValueError propagates correctly
• All 12 callback tests pass
• Full unit test suite: 4727 passed, 0 regressions

⚠️ This reopens #4808 which was accidentally closed due to fork deletion.