And example, lets say your windows 99 device pixel wide and a DRP of 2.0 and lets say you ask for 2 side by side elements (via flexbox or whatever) each width:50%. At the HTML level the browser will compute there are 49.5 CSS pixels available. It will then compute the size of each 50% as 24.25 CSS pixels but in order to fill the 99 pixels, one of those 2 elements has to be 50 pixels and the other 49. There is no part of the spec that says which side gets 50 and which get 49. Most browser compute 50 for both sides, the pass it up their compositor which will use 50 pixels of one and 49 of the other. The point is, you can't know which one will get which by doing CSSpx * dpr.
This is why ResizeObserver device-pixel-content-box was added. So that you can ask the browser what it did.
TIL. Oh gosh I don't know if I have the brain power to work through all the necessary rendering layers logic to fix up my game engine to make the pixels as they should be, but if I ever do and figure it out, great thanks to you! (and even otherwise, thanks!)
I hack around it by taking the CSS size of the container, finding the closest smaller integer device pixel size, and set the canvas context to that and the canvas CSS size to the fractional size, which can be smaller than the containers CSS size.
Basically rounding down.
I feel your pain about non-integer scaling ratios. The way I hack around it is to set the canvas context to an integer number of device pixels, and then set the CSS size of the canvas to a fractional size: device pixels / DPR.
We use this along with some other hacks on keep.google.com and some other sites to map 1:1 to display pixels to try to ensure hardware overlay promotion for the inking surface.