docs: trim internal training-provenance from the model card

Add multilingual email-gate (MiniLM + folded LogReg head, INT8 ONNX) + model card

initial commit