No, that's not the point. The point is that when velocity is constant, there's no indicator that your pedal manipulation is good as opposed to bad. For that to happen, you would need a period of time where you messed up with the pedal, and saw that you messed up. Then you could make an argument like: Hey, when we let hillAngle-pedalAngle [or whatever] drift away from zero, things go bad [velocity will fall].

Now it is true that you can measure collinearity between the hillAngle and the pedalAngle and show that they are correlated... but we sorta knew this already, didn't we? After all, we're setting the pedalAngle based on the hillAngle. We created that functional relationship. There's no need to empirically discover it.

Put it another way. Suppose v = w1 * hillAngle + w2 * pedalAngle + v0, and you want to find w1 and w2.

If your dataset shows pedalAngle = -hillAngle, then you have

v = w1 * (-pedalAngle) + w2 * pedalAngle + v0 = (w2 - w1) * 0 + v0

So you're dataset tells you nothing about w1 and w2.