Action Scaling at the Harness Boundary Beats Trajectory Re-Runs

Why terminal agents fail from corrupted shell state rather than bad reasoning, and how sampling candidate bash actions before execution cuts test-time compute by 5.8x.