[MrJohz]

Sorry, I didn't see this earlier. The problem is that it's very difficult to determine what properties are observable for fingerprinting purposes. I used the background image as an example because it's very simple, but you can also trigger requests in more obscure ways. For example, you could have a lazy-loaded image in the rendered HTML - the image will only be loaded if the user's viewport contains the image. Then you create a rule where if the user is using dark mode, the element immediately before the image becomes really long, forcing the image off the screen. Now, if the user loads the website and doesn't immediately also load the image, you know that they were using dark mode.

Alternatively, everywhere where you have a link, you could have one link for each combination of bits that you want to send to the backend. Then using CSS, you can hide or display this links so that only one version of each link is displayed at a time, and then monitor what gets clicked. If the user clicks the link that says `/?dark-mode=true&orientation=vertical`, you now know two extra bits of information.

This is obviously all excluding Javascript, which can just read this information straight out and use it.

The problem ends up being that there's so many different (and often valid) ways to customise a website that it's very difficult to limit these customisations to only the "safe" ones. Even if the only properties I was allowed to use were colour/background-color, I'm sure I could come up with some sort of way to use them to convey information. So the only safe option here is to turn off the customisations altogether. Yes, it's still possible to track if a user is using light mode or not, but now they're all using light mode, so that bit of information becomes useless.