Ah, turns out where we place the Compound point (center of mass?) has an effect.
Changes made:
Placed Compound point in the center of Capsule (between both spheres and in the center of cylinder)
setAngularVelocity of Compound to zero in every simulation step (seems to not have an effect after the first change was made)
Not perfect yet, but getting very close 
: https://www.babylonjs-playground.com/#3EDS3A#96
The spheres can still slightly clip into the red slide. I think this is because while the Compound point (center of mass) is now in the center of the Capsule, it is not in the center of each sphere?