feat(benchmark): record hierarchical handoff evidence

[JaimeCernuda]

Summary

• records direct planner-selected tool calls as owning expert handoff events
• keeps inferred expert attribution when provider errors happen after successful tool observations
• tightens the demo benchmark runner so partial recovery is reported as partial, not clean pass
• adds the natural NDP -> SAC -> visualization demo case and refreshes the ALCF benchmark report

Evidence

• Live ALCF/Metis run on gpt-oss-120b: 13/15 clean passes, 1 expected surfaced error, 1 partial recovery, 0 hard failures
• Strongest demo: ndp_seismic_waveform_to_plot with ndp_catalog -> analysis -> sac_format -> visualization handoffs and SAC PNG artifact

Verification

• uv run ruff check src/ tests/ scripts/run_demo_benchmark.py scripts/create_benchmark_data.py scripts/create_demo_data.py
• uv run pytest tests/test_core/test_agent_planner.py::TestRunAgentLoop::test_direct_tool_action_records_owner_handoff tests/test_core/test_agent_planner.py::TestRunAgentLoop::test_provider_error_after_tool_keeps_inferred_expert tests/test_core/test_agent_planner.py::TestRunAgentLoop::test_tool_observation_then_planner_failure_synthesizes_partial_answer tests/test_core/test_agent_dispatch.py::TestForwardDispatch::test_parquet_file_followup_reuses_last_session_path -q
• uv run pytest tests/ -> 1156 passed, 37 skipped