The Rosetta Paradox: Domain-Specific Performance Inversions in Large Language Models