improve CAD segmentation

[DamienGilliard]

This PR contains changes that improve the CAD segmentation in "difficult" CAD-pc cases

It includes;

• a homogenization of normals that appear to be easily flipped on horizontal surfaces: now all normals are facing up
• a fallback solution for beams that have only 2 parallel faces in the scan: we use bounding boxes to find secondary axis

[DamienGilliard]

Hi @eleniv3d , This PR seems ready for review !
Here is a file to test it;

2026_01_28_test_df2.zip

[eleniv3d]

Hi @eleniv3d , This PR seems ready for review ! Here is a file to test it;

2026_01_28_test_df2.zip

unfortunatelly the pcd input is internalized in a cockroach cloud parameter which i have tried before to install but doesnt work :/ would it be possible to put i in a gh pcd parameter? :)

[DamienGilliard]

damn.
Here it is:
2026_01_28_test_df2_with_gh_pointcloud.zip

[eleniv3d]

i remember last time we also discussed the option to introduce another filter in the loop that associates each beam that looks more or less like this:

#build a BRep-sampled point cloud for each beam
#run ICP with source = brep_cloud, target = associated beam cloud
#use the resulting transform to compute a registered bounding box of the brep
#“unassociate” any face-assigned points outside that bbox

i just leave it here as a thought..