Many thanks for your advice - in the end I did what you suggested and slowly built the model toward a structure that makes sense and that I can justify. I think the lesson here is that changing estimates are simply indicative of more complex statistical interactions that unfold or become visible when I make the model more complex. I guess this doesn’t mean that the simple model becomes invalid, but that I have to discuss the different surface smooths I get: one for a more simple and one for complex representation of my system.
In that regard it would help to be able to discuss the “goodness of fit” or “variance explained” for these models, to be able to say which of the models performs better statistically. I guess to that end I could use the sigma of my dependent variables?
Or, is there something analogous to the “Deviance explained” in the mgcv package? I was already wondering about this before: Estimate variance explained with GAM brmsfit
Finally, I am wondering whether I absolutely have to scale my variables: I have always used brms with scaled variables (mean=0, sd=0.5). However, when using unscaled variables that are simply transformed to be in the same range (log/sqrt) I get a smaller sigma - is this a wrong thing to do?