Replace SFT reward curve with baseline-anchored learning curve (tool-aware baseline → checkpoint-40 … final)