For Gaussian denoising of nonnegative leaking flows, consider least squares over the full flow cone, with a fixed bound on the true root flow. We construct an N-vertex broom whose worst-case normalized risk is O(sqrt(log N)/N), while the N-vertex rooted star has risk Omega(log N/N) already at the zero signal. The ratio of the two worst-case risks tends to zero. Thus the star is not the easiest tree for this estimator when logarithmic factors are retained. The proof combines a deterministic cone-projection identity with an elementary bound for the maximum of Gaussian leaf noise. For context, we also reconstruct the known parallel-arm obstruction to the path being hardest, using a nonnegative witness. The comparisons concern the full-cone estimator, not a budget-constrained projection or minimax risk over all estimators. This is an AI-assisted, self-audited, unrefereed preprint; no absolute priority claim is made.
