A powered test: do small AI models quote a student's misspelling back less exactly? (No clear difference)