Harness-Aware Distillation for Small Language Model Agents