Cluster deformation is calculated from the transformation of the cluster handle relative to handle's parent. So when you parent the handle and transform the parent node, there is no actual change in the local transformation of the handle since everything is inherited from the parent.
However, the absolute (world space) position of the cluster handle get changed and maya just applies that transformation to a handle and every cv that is member of the cluster deformer set.
Reasonable solution for the rig you want might be:
Create the cluster deformer for the object and make sure that all of the objects' cvs are included in the deformer set ( this is important for hierarcy to work correct ). Paint the weights. Cv's that you don't want to deform must have weight of 0 but they still must be in the deformer set.
Now parent constrain the handle to a locator. Create another locator and parent the first locator and the cluster handle to it. The later locator will now control the transformation of the whole system and can be parented to other hierarchy.
Of course, you can add more clusters following the same logic. As long as all of the cvs are exactly in one and only one of the cluster sets, it should be ok.
