Seol mar théacs é seo: On guaranteed optimal robust explanations for NLP models