Ok, let's start over.
I am using a 170N gas strut, which I think could be too much tension.
If anything it's on the low end. I use a gas strut with 2x that force. Sometimes drive wheel spin can throw the belt off that would have stayed on (albeit off center). For example, when using a very high friction attachment, a little bit of water on the belt of my grinder causes the drive wheel to spin and throw the belt even though it was tracking properly.
When running the belt in normal operation the tracking wheel is canted over at quite an angle to get the belt to run central on the platen.
In reverse it comes straight off, no matter how I adjust the tracking.
This says to me that there is run out between the drive wheel and the platen wheels. When the belt is traveling from the drive, over the tracking, then to the platen, it can be corrected by an extreme adjustment of the tracking wheel. When running from the drive, over the platen, and then the tracking wheel, it comes off before the tracking wheel can correct it.
I built all of my grinders in very similar fashion as yours with slot and tab assembly of coil plate. The problem with this is that plate is not terribly flat, and the one side that is acting as a reference plane may be throwing the drive wheel axis out of relation, or the stacking of the tool arm, and plate platen arm may be doing the same to the platen wheels. This plate is unpredictable and 75% of the time seems flat enough everything works, but 25% of the time there's some glaring issue that results in symptoms like yours. Without a CMM it's almost impossible to measure which axis has the most deviation from the theoretical reference plane.
I have experimented with moving the tracking wheel position on the shaft (in and out) and it doesn't have any significant effect. What seems to make a bit of difference is the height of the tracking arm, the higher it is the more angle is required on the tracking wheel. That said I only have 40mm of adjustment on the arm, and I imagine the force exerted by the strut changes over that distance too.
I would shim the tracking wheel out to the point that when the adjustment axis is "level" and the tracking adjustment is nominal (face of the tracking wheel is parallel to the face of your reference plane) the center is in the same plane as the center of the drive wheel. Then leave it for the time being.
I didn't notice if you mentioned, is your drive wheel crowned or flat? I'm a huge proponent of only 1 crowned wheel per system. Every tracking issue I've personally had was impossible to resolve until I eliminated any crowned wheels other than the one doing the tracking.
That said, start with shimming the drive motor in one axis like P.Brewster described. Shim one side out. Test a belt by hand. What does it do? If it wants to slide off the machine, note which direction. Remove the shims and shim the other side of the motor. Test by hand again. Does the belt move off in the same direction as before? I like doing this kind of trouble shooting with drastic changes, because the results are more observable. Shim it .030" one way, then the other. If the belt comes off in opposite directions, you know that the resolution lies between +.030 and -.030. If the belt comes off with +.030 but at -.030 it just goes over center without coming off, then you likely know that the solution is between .0 and -.030.
If the results are indeterminable with shimming the drive wheel, then the problem may lie in your platen wheels.
I can't work out what element of the geometry is skewing it. Or whether it is the pulleys which are the problem. I bought the pulleys from an ebay seller in poland who manufactures them. They were quite cheap. From eyeballing it it could be that the tracking wheel is not quite evenly crowned.
I've bought really cheap pulleys, chinese, american made, fiber, aluminum, even made a couple myself. Outside of limiting a system to a single crowned pulley, they all work fine.
