I could see that approach being viable when using a component-based framework, such as React. However, if implementing with a more traditional framework or application, we’re back at our starting point: presentational, declarative UI elements that are segregated into different chunks of front-end code.
The end result of all our hard work will always be HTML, CSS, and JS, each performing their specific function. Part of the intent is to logically bundle code together that’s inherently bound together anyway. HTML and CSS are more tightly-bound, and thus it makes more sense to bundle them together.
Also, while the concept of bundling the Atomic classes together into JS variables that describe their appearance sounds nice, we’re beginning to approach the level of indirection that concerned me in the first place: begin at HTML, find some sort of tagname or attribute that indicates what JS component we’re using, find JS file containing component, find any variables containing lists of CSS classes, etc. It’s a lot of overhead.