Small polygons under rotation result in SDF flipping its sign

[sorki]

These two should be equal, but internally when sampled at (V2 (-1) 0)
the sign of the SDF differs yet they both get rendered correctly.

let funPoly = polygon [V2 0 0, V2 0 (-0.1), V2 (-2) 0, V2 0 (-1)]
    rotFunPoly = rotate (2*pi) funPoly

> getImplicit funPoly (V2 (-1) 0)
-4.993761694389224e-2
> getBox funPoly
(V2 (-2.0) (-1.0),V2 0.0 0.0)

vs

> getImplicit rotFunPoly (V2 (-1) 0)) 
4.9937616943891996e-2
> getBox rotFunPoly
(V2 (-2.0000000000000004) (-1.0),V2 0.0 4.898587196589413e-16)

This has something to do with isIn

ImplicitCAD/Graphics/Implicit/ObjectUtil/GetImplicit2.hs

Line 60 in 874cb76

minimum dists * if isIn then -1 else 1

and

ImplicitCAD/Graphics/Implicit/ObjectUtil/GetImplicit2.hs

Lines 49 to 56 in 874cb76

	relativePairs = fmap (\(a,b) -> (a - p, b - p) ) pairs
	crossing_points =
	[x2 - y2*(x2-x1)/(y2-y1) | (V2 x1 y1, V2 x2 y2) <- relativePairs,
	( (y2 <= 0) && (y1 >= 0) ) || ( (y2 >= 0) && (y1 <= 0) ) ]
	-- FIXME: use partition instead?
	seemsInRight = odd . length . filter (>0) $ nub crossing_points
	seemsInLeft = odd . length . filter (<0) $ nub crossing_points
	isIn = seemsInRight && seemsInLeft

and it only seems to affect small polygons + rotation. If the isIn mechanism is dropped, they don't get rendered at all.

[sorki]

be1ee7a adds golden test (since we don't have any 2D ones). The two renders differ in one pixel and 9 bytes..