Computing the log density of a bernoulli is easier numerically if you keep everything on the log scale. If you transform back and forth from probabilities you can have loss of precision issues.
The bernoulli function doesn’t like the parameter value 1 (which will occur for log scale parameters negative enough). This has come up before: Initialization between (-2, 2) failed after 100 attempts . I think this is a bug. Made an issue (https://github.com/stan-dev/math/issues/879). Tnx for being patient and posting this.