I think your test is a bit more complicated than it needs to be again because you're trying to create a square. You could have, for example, used 4 different colored shadows off a single character and compared the result in vertical-rl to the same result in horizontal-tb. But the test is not incorrect.

Two options for you:

• I can approve the test as-is. (Though I would recommend using green instead of purple and giving the div a red background.)
• You can rewrite it using something like text-shadow: -1em 0 orange, 1em 0 blue, -1em purple, 1em yellow; margin: 1.5em against a single-character test div and, now that it's simpler and easier to understand, include a div for every writing mode combination.

4 different colored shadows off a single character and compared the result in vertical-rl to the same result in horizontal-tb

This flower-and-petals-like formation is indeed easier to understand and its design is more simple. I have created

http://www.gtalbot.org/BrowserBugsSection/CSS3WritingModes/s76-text-shadow-orientation-upright-002.html

http://www.gtalbot.org/BrowserBugsSection/CSS3WritingModes/reference/text-shadow-orientation-upright-002-ref.html

and I will make the necessary github adjustments to this PR later today.

I'm a little bit concerned that the text will have anti-aliasing but the image reference won't, so probably a better reference would be to have a horizontal-mode copy of the text-shadow test code. But I'm not noticing any problems at the moment, so I will approve this and maybe the reference can be tweaked later if needed.

https://wpt.fyi/results/css/css-writing-modes/text-shadow-orientation-upright-001.html?label=experimental&label=master&aligned

There is indeed a failure in Edge 96 due to anti-aliasing but only in Edge 96. I will revisit the reference and implement your proposal later this fall: it is on my to-do-list now.

the text will have anti-aliasing but the image reference won't, so probably a better reference would be to have a horizontal-mode copy of the text-shadow test code.

I almost forgot about this. A proposed correction and improvement is now in PR35337