Generating classes at build time vs. generating at run time

[rmst]

I'm trying to understand the trade off from the title. Generating classes (or rather deleting unused classes) at build time has the drawback that it requires static code analysis which is complicated and imperfect. E.g. it currently interferes with assigning classes dynamically even in relatively simple cases. So instead, why aren't the necessary classes generated at run time? E.g. assume we only care about making it work for React. We could hook into the render process, observe all the class names and generate them. Sounds complicated but it isn't really, especially compared to the rest of the React render process. The mapping from Tailwind class names to css is super trivial, e.g. the margin classes m-${size} simply map to margin: ${size * 0.25}rem and once the class has been generated the only overhead left would be the hash map lookup for the fact that it already has been generated. What am I missing?

[rmst]

Okay I guess https://www.npmjs.com/package/@tailwindcss/jit exists which sounds like its ditching build time generation for run time generation and seems to have been pitched as the future of Tailwindcss but for some reason it still depends on postcss and the main tailwindcss package and also nobody seems to be using it (6k vs 10M weekly downloads on NPM) after existing for 4 years. What's going on there?