Hey folks, I was able to Iprove the scale direction to get a way more consistent result, the steps are:
rebuild a base model with as uniform and low poly as possible
use subdiv to refine
Apply the geonodes for the scales (with martin’s trick to change the size dynamically)
Add a second align rotation to euler that is reading the nearest surface of a cylinder and the result is way more consisten, only requires a small manual correction

