Resample or Reroute? Budget-Aware Test-Time Model Selection for Large Language Models