Note that after first run your mount remembers 3 alignment points (based on 3 resolved captures), next you adjust the mount manually with screws, so these alignment points are not valid anymore, but probably are used in the next run of the procedure. Can you check on your mount's Align tab how many points are used before the second run? I believe it might be the root cause of this issue in some circumstances. In such a case it's enough to clear all the points before the second run.
I've been toying with PAA, CCD Simulator and Telescope Simulator from the git repository and after completing the process I always get about 5' 40' - 5' 50' of PA error. In fact, the same behaviour can be seen in Jasem's video (watch it from 5:10 onwards):
As you can see, there's a PA error of 5' 41".
In my mind, both CCD Simulator and Telescope simulator should be "perfect" and have 0 PA.
So I have been using the CEM25's built-in Polar Alignment iteration routine, which is the only way to align from the equator.
It works, and just requires me to flip back and forth between two stars for ALT and AZ adjustments.
I cannot see any stars within 20 to 30 degrees of either pole. Will the EKOS polar alignment routine work in such case?
I have not tried myself at night as we have had nasty skies all week.
I would be highly interested, if this worked - because nevertheless I am living at 48 deg north, I don't have a view to north from my balcony So without using a guiding scope it is really really problematic. I think, this is a scenario, which a lot of people might have. Or think, if you are in the mountains or have trees to the north.