Merging as it's a revert.

Is there any way I can help?

@bernaferrari No, I think it requires access to the failing tests to fix. The problem appears to be something along the lines of some rounding or an off-by-one error. Some thin borders disappear entirely, and others are just off by a bit (less than a pixel). I'll see if I can make a repro case that we can use as a regression test.

If I remember well, one thing we removed was if (width == 0) drawRRect (now it draws DRRect, no special case. In the perfect world, it wouldn't even be drawn). It is the only thing I can think of. Feel free to try adding it again (or skipping!) and see how the test behaves.

That's the only thing I can think.

Try on RoundedRectangleBorder.paint:

if (width == 0.0) {
  canvas.drawRRect(borderRadius.resolve(textDirection).toRRect(rect), side.toPaint());
} else {
// ...
}

or

if (width == 0.0) return;

I went ahead and made this: #109975
If you could test on Google internal and see if the issue still exists, it would be great. I don't know what happened to the bot that did this before.

I have two theories: drawDRRect has a bug and the person that made width == 0 knew this. Or, drawRRect when width == 0 does something which is better than not drawing anything (or drawing the 'wrong' way).

I don't know what happened to the bot that did this before.

It doesn't kick in until the PR has been approved.

I have two theories: drawDRRect has a bug and the person that made width == 0 knew this. Or, drawRRect when width == 0 does something which is better than not drawing anything (or drawing the 'wrong' way).

It doesn't have a bug, but I think that perhaps the anti-aliasing behavior is different between the two. I'll take a look.

It doesn't kick in until the PR has been approved.

That's how it worked, but now it doesn't. You merged your and I merged mine without the Google testing happening. It only happened after.