CSGPolygon3D union has inconsistent behavior with coplanar faces

[OffsetMOSFET]

Tested versions

4.4.1
4.5 dev 4

System information

Ubuntu 22.04.5 LTS

Issue description

CSGPolygons do not consistently join when sides are coplanar.

Steps to reproduce

In the MRP, I have constructed four CSGs

• Bottom of large CSGPolygon connecting to tops of smaller CSGPolygons (Bottom-Top)
• Side of large CSGPolygon connecting to sides of smaller CSGPolygons (Side-Side)
• Buttom of large CSGPolygon connecting to sides of smaller CSGPolygons (Bottom-Side)
• Side of large CSGPolygon connecting to tops of smaller CSGPolygons (Side-Top)

The CSGs have flipped faces for easy viewing, but the behavior is the same without flipped faces. You can see this by using wireframe camera mode and unflipping the faces, that the resulting mesh isn't correct.

In the cases of Bottom-Top and Side-Top, the example is able to join the CSGs correctly.

In the case of Bottom-Side, the large CSG does not join with the smaller CSGs, which do join with each other.

In the case of Side-Side, the small CSGs that are level or lower (y-axis) with the large CSG do not join with the large CSG. If you increase the depth of one of the unjoined CSGs, you can see even glitchier behavior where only sections of the small CSGs will join.

Minimal reproduction project (MRP)

example_csg.zip

[OffsetMOSFET]

Note that the example uses whole numbers to avoid floating-point errors, though that does not guarantee the internal math avoids it too.

[kleonc]

This indeed looks like a floating point errors. What seems to be problematic is the Basis of the CSGs, as 90 degrees rotations introduce errors:

The above produces Basis with such basis vectors:

X: (1.000000000000, 0.000000000000, 0.000000000000)
Y: (0.000000000000, -0.000000043711, 1.000000000000)
Z: (0.000000000000, -1.000000000000, -0.000000043711)

Getting rid of those near zero values on Combine Side-Side and Combine Bottom-Side makes the issue go away for them:

@tool
extends CSGCombiner3D

@export_tool_button("Fix children basis") var fix_children_basis_action = fix_children_basis

func fix_children_basis() -> void:
	for child in get_children():
		print()
		print(format_basis(child.basis))
		for axis in 3:
			child.basis[axis] = child.basis[axis].round()
		print(format_basis(child.basis))

func format_basis(basis: Basis) -> String:
	var coord_format := "%14.12v"
	var axes_format := "X: %s, Y: %s, Z: %s" % [coord_format, coord_format, coord_format]
	return axes_format % [basis.x, basis.y, basis.z]

Bottom-Side	Side-Side

Note that:

• For the Bottom-Side case the large CSG is not rotated, so it somehow makes sense that the small rotated CSGs with the erroneous Basis ended considered as not coplanar, and hence faces were not joined.
• For the Side-Side case the large CSG is also rotated, it had the same erroneous Basis as the small rotated CSGs, hence it's rather unexpected that some of the small CSGs were not joined correctly. And this indicates it might an issue about CSG merger, not just problematic Basis.

Also:

• For the Bottom-Top case no CSGs are rotated, no erroneous Basis involved, works fine.
• For the Side-Top case only the large CSG is rotated, it has mentioned erroneous Basis, yet merging seems to work fine.

(cc @fire)

[fire]

we recently merged a new manifold version upgrade. so need to check if if your report is before or after

[kleonc]

we recently merged a new manifold version upgrade. so need to check if if your report is before or after

Oh, I've tested previously in v4.4.1.stable.official [49a5bc7] only. 🤦‍♂
But just checked now and got the same results on the current master (v4.5.dev.custom_build [be3ecae]).

[fire]

What happens if you avoid the euler mode and use quaternion or basis mode?