Write the four-personality sweep: run(lr) descends f(w) = w² from w=5 for 30 steps; then for lr in (0.01, 0.1, 1.0, 1.1) print f"lr={lr:<5} {verdict}" where the verdict is "converged" if |w| < 0.1, else "DIVERGED" if |w| > 6, else "oscillating in place" if |w| >= 4.5, else "crawling". These four shapes are the vocabulary of loss-curve reading.