Narrow phase incorrect incremental state update upon collider reparenting

When reparenting a collider to another rigidbody, the narrow phase code does not handle the reparenting correctly:

Two colliders' interactions are ignored if they belong to the same parent rigidbody, which makes sense. However, this is only checked when adding the collider pair to the narrow phase (to either the interaction or contact graph), in add_pair.
As a result, if two colliders used to be in the same parent but one is reparented, narrow phase ignores their collisions until some other mechanism notifies it of the pair again. The symptom of this is that if at the moment of reparenting the two colliders are already overlapping, narrow phase completely ignores the penetration until they are moved apart and have a chance to re-collide.
On the other hand, if two colliders used to be in different parents but are reparented to the same rigidbody, narrow phase does not correctly ignore their collisions. The symptom of this is that if two objects are already overlapping (via e.g. squeezing each other) when the reparenting occurs, and if the positions of the two colliders in the rigidbody are also overlapping, then narrow phase continues to treat the overlapping as normal collisions and tries to resolve them, resulting in spontaneous accelerations of the parent rigidbody.

To reproduce (sorry don't have time to write a test case right now but here's exactly what to do):

First issue: make two overlapping colliders in one rigidbody. Simulate. Then reparent one collider to a new rigidbody while keeping global position. Simulate. The two colliders will remain overlapping.
Second issue: make two colliders in two separate rigidbodies that overlap. Simulate but without resolving the overlapping completely. Then reparent one collider to the other's rigidbody, keeping the global position. The new rigidbody will spontaneously accelerate.

To fix, we should handle collider parent change correctly in handle_user_changes_on_colliders in narrow_phase.rs, but I'm not sure if this can tackle the first issue, because when the pair was rejected earlier, narrow phase does not know of the pair. The second case can be correctly tackled this way by removing the pair that is no longer supposed to be in the graphs. But there may be other implications of parent change that I'm not aware of.

As a workaround, instead of reparenting I am removing and re-adding the same collider. This solves both issues.

dimforge / rapier

Narrow phase incorrect incremental state update upon collider reparenting #734