arXiv:2611.06492v1 Announce Type: cross Abstract: Largest-angle bisection (LAB) repeatedly divides a triangle by bisecting its largest angle. Although this procedure is known to preserve nondegenerate triangle shapes, less is known about the distribution of shapes across the resulting refinement tree. We show that its long-term behavior is governed by a distinction between pure and mixed branches. A branch that eventually selects only one of the two children may approach an exceptional family of shapes. By contrast, every branch that changes children infinitely often converges in shape to a right-isosceles triangle. This follows from a projective Lyapunov function that is non-increasing at every refinement and decreases strictly whenever the selected child changes. Consequently, a random branch converges to the right-isosceles shape with probability one, and the proportion of triangles near that shape tends to one across complete refinement levels. Largest-angle bisection therefore produces statistical regularization even though exceptional branches remain present.