These are important concerns.
The gait23 models without patellae reflect muscle path kinematics that are identical to those with the patellae present. If you evaluate the moment-arm using muscle length to joint angle relationship (dL/dtheta) then the moment-arms should appear to be identical. Unfortunately, the effective torque experienced by the knee joint are very different (OpenSim switched to the effective torque method as of 2.4). Why? In the models with patellae, a patella is constrained to a tibia via coordinate couple constraints. Mechanically, they represent massless gearing between the rotation of the tibia (w.r.t. femur, specifically the knee_angle) and the rotation and translation of the patella (w.r.t tibia). When muscles pull on the patella they create constraint reaction forces that cause the knee to rotate. That is, the effective torque on the knee_angle coordinate is due to the coupling mechanism of the patella to the tibia. In the model without a patella, the points of attachment where also carefully prescribed with respect to the tibia as functions of the knee_angle and act as extensions of the tibia. The problem here, is that there isn't a representative mechanism for dragging the via point around and in the the case the via point motion is under tension it is doing additional (unaccounted) mechanical work on the system. In the patella model, part of the definition of a coupler constraint is that it does no work, and the work to move the point (fixed on the patella) has to come from the applied tension.
That makes a strong case for not using moving path points altogether and we are considering that. However, for the purpose of performing muscle-driven forward simulations, moving path points do enable modelers to eliminate light/massless bodies whose only purpose for existing is to manipulate the muscle path to produce desired moment-arms. The use of fake bodies are equally non-physical in my book but also have a very detrimental effect on computational performance. If one thinks of the moving path point as only specifying the instantaneous location of a via point (which neglects the work required to move the point around under tension) then you can see how an applied tension leads to the effective torque. This might be a trade-off a modeler may be willing to make for significant gains in performance. In fact, modeling the patella as fully constrained w.r.t the tibia is also incorrect, in my opinion. The patella articulates w.r.t. the femur (not the tibia) and constraint (or contact) forces should reflect the interaction between these bodies in order to correctly capture the transmission of muscle tension to bones. The next full-body model that we release will reflect these concerns and will not use moving path points or massless bodies.
For physically plausible mechanisms the dL/dtheta and effective torque (r = Tau/tension) must be (and are) the same. When OpenSim users' scalar multiply muscle tension with the the muscle moment-arm, they are clearly trying to compute the effective torque produced by that muscle. Here, the moment-arm is the transmission of tension to torque (more accurately generalized force), so we have adopted the effective-torque definition for moment-arm to provide consistency at the level of system dynamics. It is also fast and robust to compute across a range of constrained and unconstrained multibody mechanisms (see
http://simtk-confluence.stanford.edu:80 ... 9103515834 for more details).