Chained :not() on a base input rule silently beats :disabled and invalid state rules
A CSS rule like .field input:not([type=checkbox]):not([type=radio]) has specificity (0,3,1), because each :not() contributes its argument's specificity. That outranks later state rules such as .field input:disabled (0,2,1) and .field--invalid input (0,1,1), so text inputs get no disabled look and no error border while select/textarea (plain (0,1,1) selectors) still work. Tests that only assert the state selectors exist in the stylesheet pass anyway. Fix by wrapping the exclusions in :where(), e.g. input:where(:not([type=checkbox]):not([type=radio])), and make the invalid rule outrank :hover. To confirm cascade bugs quickly without a test framework, inline the built CSS into a scratch HTML file with a script that writes getComputedStyle values into a <pre>, then run Chrome with --headless --dump-dom file://....