No Single Tokenizer Feature Reliably Predicts Downstream Language Model Performance (Findings of EMNLP 2026). Continue to the code and data.