I'm getting my ass kicked by something that I'm sure has a simple solution: more off-by-one errors. I've tried various things but as a baseline I have this code in arc.rs (not pushed):

let radius = Real::from(self.diameter) / 2.into();

Point::new(
    i32::from(angle.cos() * radius),
    -i32::from(angle.sin() * radius),
) + Size::new_equal(radius.into())

The sin/cos stuff is fine, but as far as I can see the final point rarely lies on the actual arc (or circle, same thing) because the point is transformed not from the true center of the arc but somewhere near it (+/- 1) due to rounding. @rfuest you've done some magic with circles in the same vein as this, so could you offer some guidance perhaps?

I did take a very brief look at this and I think that at least part of the problem is in the existing code. I would expect an 90 degree arc starting at 0 degrees to exactly draw one quadrant of a full circle. But in my test this only worked correctly for arcs with odd diameters. Arcs with even diameters were missing one row of pixels at the bottom.

The circle code internally uses coordinates that are scaled by a factor of 2 (e.g. in Circle::center_2x). This way it is possible to use an effective resolution of half a pixel for the calculations, which is required to represent the accurate center of a circle with even diameter.