In hanger-3dseeds, a 3D cloth simulation designs a shirt hanger, and the best designs had blunt ends that sat on the 6 cm end-radius limit in three of eight seeds. We raised the limit to 9 and 12 cm and ran eight seeds each.
The numbers. At 12 cm no seed sits on the limit. Whole-sheet error goes 1.12 cm (6 cm limit), 1.02 (9), 1.01 (12); neither difference is distinguishable from seed noise (p = 0.41 and 0.45). The chosen radius ranges from 4.8 to 10.4 cm across seeds at nearly equal error, so the score barely cares about end radius.
What it means. A wider box is not the fix. A 10 cm end is not a real hanger tip, so the next step is a contact-width penalty. Stylised cloth constants, one garment and one arm depth; nothing here is measured on a real garment.