Admittedly, this is an edge case - however the current implementation seems a bit hacky anyway.
The issue:
We have a complex assembly of objects that are tied into a hierarchy, for example MegaParent → Parent → Box
If Box has a pivot set for it, using any attached gizmos will cause the object to rapidly fly away:
(try using gizmo attached to be Green box)
It looks like that that current implementation using Gizmo._handlePivot would handle only some specific cases, but not a hierarchy like in the example.
The core of the problem appears that Gizmo._matrixChanged uses Matrix.decompose to extract position from the local matrix. However, when pivot is set, TransformNode.computeWorldMatrix does not use Matrix.compose - instead it factors in the pivot matrix. Thus decomposing that local matrix will return a different local position and that error will rapidly accumulate.
A potential fix is to remove the position tweak in _handlePivot and handle the “un-compute-world-matrix” better in the _matrixChanged:
I’m not totally sure that fixes it in all cases thus no PR, but a monkey patch like one around line 100+ here seems to work in our case: