Spotlights and Blindspots: Evaluating Machine-Generated Text Detection