322M zero-shot typed-decision model for Indian languages + Hinglish, and its open training data (real Sangraha text + synthetic, Qwen3-labelled).