6K 93% [Random Samples] Instance-Adaptive Inference-Time Scaling with Calibrated Process Reward Models