New jev_classify tool: batched single-label classification. The class catalog is sent once in shared state; every item becomes an independent Choice question, so a 4-item batch costs ~670 input tokens instead of ~940 per item. Auto-acceptance requires both top probability (default 0.85) and winner-to-runner-up margin (default 0.5); caller IDs are preserved verbatim; malformed responses surface as invalid_response; batches cap at an 8,000 item-class budget. Install: npx -y @jkudish/jev-mcp