I should have said in my other reply how neat it was to see the guidance system almost instantly kill the roll after booster sep, no sign of over-correction. Also so good that you have video cameras. Pretty for entertainment, but also so critical for technical review of guidance effectiveness.
I have a couple of ideas. One, pretty simple, but may not work. The other...fuzzy.
The simple one would be to orient the guidance section so its control fins would be in-line with the booster fins, or possibly right in-between. IIRC, you may have to remove and reposition a stage coupler alignment guide to be able to do that. Reason why it might make a difference would be how the airflow may be interacting downstream of the control fins. As it is right now, deflected airflow from an angled control fin seems to be impinging on the booster fin to cause slightly higher pressure on it to cause a roll.
I take note that you do not have the guidance control fin alignment exactly between the booster fins. Rather than 60 degrees apart, it looks more like 30 (fin seen on left) and 90 (fin to the right). So this aerodynamic effect probably is way more on the fin on the left and possibly not at all on the fin on the right. I take note than on flight #6's video, at first the rocket started to roll right (camera view) just a bit, then the guidance system stopped that right roll, stopped for a moment, then began a left roll that got faster and faster. So that may be a clue that the aerodynamic interaction is limited b how far away (in roll orientation) the lower fins are. So if the fins had been exactly in-between, 60 and 60, then this effect likely would be less effective on the fin at left, but then would affect the fin at the right equally.
What I am not sure of is whether the effect would be solved by having the control fins directly inline with the booster fins, or not. It may still have an effect, possibly make it even worse. If not, then at least the effect could be reduced by having the control fins exactly mid-way, at 60-60, not the current orientation that looks to be about 30-90.
The other idea is "fuzzy logic", in case you do not think this is a repeatable effect. If it is repeatable, then the simplest programming fix would be to have the control surfaces move the opposite way for first stage boost, fighting reversal with reversal to get the desired result.
But if it was not repeatable, but a random effect, the programming would need to be able to try to figure out if it is in a reversal situation or not, and correct accordingly. So if the rocket begins to roll one way, the guidance acts to correct the roll, but the roll increases almost directly in response to the control output trying to STOP the roll, the programming can consider this may be a control reversal situation. It could do a VERY brief test for that, by going to neutral for say 1/4 to 1/2 second (possibly even half-throw the opposite way), and detect whether the roll gets worse, or if the roll rate begins to reduce. If it begins to reduce, then the programming can assume a roll reversal situation and then give reverse roll control commands to the servos. Which should solve that for the rest of the first stage portion of the flight, then go to normal after booster sep. Though that probably should not be a one-shot programming tree branch, but some sort of routine that keeps monitoring whether actual rocket response seems to be as a result of control surface deflection, or not. Could even have a "When in doubt, go to neutral" routine in it if it had more than one roll reversal anomaly during boost and after the first changeover to roll control reversal to solve it, it started to roll too fast the OTHER way.
Also, at those speeds, and since aerodynamic forces increase with the square of the velocity, maybe you have a lot more control surface deflection in roll than you really need. For example the aerodynamic forces at 400 mph are 16 times greater than the forces at 100 mph. And the aerodynamic forces at 600 mph are 36 times greater than the forces at 100 mph. I mean, the ideal
guidance system would take into account velocity if possible, to have a lot more angular control surface deflection at lower speeds, and as the rocket gets faster, to have less control surface deflection for the same desired level of corrective guidance response (When you fly on an airliner at 600 mph, the ailerons are NOT as responsive to pilot commands as they are for landing!). But that's not at the core of this particular problem, not the reversal itself. But I think it may be a contributing factor. You could simply just try it with 1/3 the control surface deflection when the booster is attached (if not the entire guided flight), and see if it is sufficient to maintain roll control. And if nothing else was done to address the roll reversal problem, at least that should greatly reduce how fast it would roll
The one good thing is that the rolling tends to keep the rocket mostly straight, it's not like it is veering off course but is ballistic. Then after booster sep for the coasting vertical guidance phase it kills the roll fast and that allows for good control in pitch and yaw. More a matter of higher stresses on the rocket and some increase drag reducing the potential altitude.