Seppev If you have sub-frame samples in your point cache, they will be considered when solving substeps. If the solver steps are temporally mismatched with the point cache samples, then a linear interpolation is performed with the point cache samples surrounding the current substep.
Regarding the muscle jiggle, that is affected by the number of substeps and the time integrator you choose.
The default integrator (Backwards Euleur), is very stable, but suffers from numerical damping.
So if you add additional substeps, you'll notice that the tissue will likely appear to be more energetic.
The higher-order integrators require less substeps to fit to the fully converged solution, but could tax our current collision model, as it just tests when the simulation advances on each substep.
Regarding making a cleaner result, yes, keeping your bones connected to your joints will technically make a better result, provided that the rotations are doing the correct thing between frames.
When working that way, we'll usually use the "Bake Simulation" option in Maya to bake the joint rotations to keyframes on the joints, and then apply an Euleur filter to the curves before running a simulation.
-jj