Did you just literally combine two models? Or did you combine parameters so that they show up in both models? With the former, there shouldn’t be a problem. With the latter, it will be depend on the geometry introduced in the posterior by the combination. If the data’s compatible with both models and a single set of parameters, this should also sample OK.
The model you present is too complicated for me to understand just by inspection.
I’d recommend moving the definition of alpha to be a local variable in the model so you don’t have to save it all in the output.
It looks like the parameters that are mirroring each other are intercept and correlation terms? What I’d recommend doing at this point is generating a much simpler model to see if you can simulate/understand what’s gong on. Have you standardized predictors? If you wind up with all positive (or all negative) predictors in a regression, the slope and intercept wind up being highly correlated.