PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost