Why wider neural networks usually break—and how a simple mathematical constraint lets deep models grow smarter without collapsing.