When Do Local Score Models Extrapolate Across Size? A Diagnostic Theory and Benchmark