Hello Dirk,
Could it be that it is just an artifact of having a relatively small sample? If I use your code and run the model with n=200 and n=2000 I for instance get the following estimates for \lambda_4:
The above one is for n=200. In that case the distribution is also skewed in the uncertainty towards lower values. This seems to make sense as values above 1 are not allowed. The mean estimate will therefore also be lower. You could consider looking at the median estimates. In this case the mean is 0.83 and the median 0.87.
In the case of n=2000, you have enough information such that you get a relatively nice symmetric distribution and the mean and median are both 0.89.
I hope this helps,
Best,
Duco
Edit:
I used set.seed(12345) for both data generations in R and in both cases stan(model_code = code, data = data, seed = 123, = list(adapt_delta = .95), iter = 2000) to estimate the model.