The first forms of user tracking involved 1px GIFs that existed purely so that the server could log the request. If you allow any code execution at all, then the client can send data back to the server by asking for data from the server. Reads are just bidirectional writes.

I’m almost with you but in that case even shopping online would not work. Unless you force the user to manually type in the SKU of each item they want to add to their cart and so on. That’s not gonna happen :p

I'm assuming Tor Browser doesn't actually alter the CSS stylesheet layout engine, so here's a quick way around that: Construct a CSS stylesheet that styles a div a certain way based on the media query size. Encode the success or failure of the media queries in the width of the element and then read out the width of the element to tell if the browser is lying to you. Even if the browser throttles CSS queries, you'll still know that a lie happened, which is an extra fingerprinting bit.

If you're thinking of prohibiting dynamic CSS, you'll break all sorts of harmless JS - and still not fix the problem. You could maintain two sets of computed layouts, of course, but that would break JavaScript layout scripts (e.g. masonry.js). The option that breaks no scripts is to lock browser widths to rendering at specific viewports - if Tor Browser does that, then it probably has adequately resisted this particular fingerprinting vector, at the expense of some user experience.