Local Agent Bench: Test 11 small LLMs on tool-calling judgment, on CPU, no GPU