That is a beautiful picture, Soren!
Regarding the FCUSB and the AccuFocuser/analog DC motor focuser: I basically use the same system, although the motors I just got from Amazon for $15, so cheaper than the (now discontinued) AccuFocuser system.
It works beautifully in my hands, but there are a few things to consider.
I think the main issues you are facing results from the differential load on the motor due to the weight of the camera/filter wheel. That means the motor will move faster in the outward direction than the inward direction.
But that can easily be fixed. This is the way how I did it:
1) Note the position on the focus tube where your telescope is in focus.
2) Then attach a rubber band or a string with a spring to the telescope and the other end to the focuser tube.
3) Next, you move the telescope into a 60 degree vertical position, so that you have average gravity force pulling on the tube.
4) Loosen the clutch screw, so that the tube now moves freely
5) Add or remove rubber bands until the tube is in equilibrium at the focus point, i.e. it neither is being pulled inside by too much force from the rubber bands and neither does it slide out.
That pretty much equalizes the force the motor has to work against when moving the tube inwards and outwards and now the timed movements will be approximately equal. That will dramatically improve the ability of any of the algorithms to find focus. Basically, the principle is no different from balancing the mount. You want to minimize the strain on the motors there as well.
A couple of other things that I find important:
1) Do not make the steps the motor moves too small. If they are smaller than the variability of the HFR, the algorithm will have trouble calculating the V-curve.
2) Hy suggests using SEP, which is definitely better than the Gradient algorithm you are using. However, for some reason I find that the Centroid Algorithm works best for me, closely followed by SEP.
3) I just use Full Field, without Autostar.
Unless there was a change in the code recently that may have caused these problems, this methods works beautifully for me. Here another screenshot showing the quality of the V curve using Hy's Linear algorithm: