This one's here for completeness. In the Bullet example, we're always releasing the mouse and could maybe apply the same tactic as Auto-Keyframe. With Auto-Keyframe enabled, a script is called after you've let go of a manipulator, and whatever value is in the channel box is written to keyframes. What if we did that to read the worldmatrix?
But that won't work. Not only does it limit the functionality of the solver to the UI, it also doesn't apply to hierarchies being translated at once. We need these things to happen interactively, irregardless of UI or events.