Update public comparison roster; preserve model weights and benchmark scores

Publish complete decision benchmark, task diagnostics, and current latency