As @jd_c points out, the intercept in brms is not what you generally think of. This is what I meant earlier:
Hopefully the additional details provided @jd_c gives this additional context as to why the prior specification is misleading in this case.
Can you clarify what you mean by “performs much better”? Is this from a comparison of the LOOICs, effective parameters (p_loo), comparison of effective sample sizes, or visual PPCs? A model can perform better than another one on a lot of different things, so it’s just helpful to know what is being used as the criterion. I see later on that you reference the loo comparisons, but there can be additional information from the loo package than just brms::compare_loo() function returns. See this excellent thread on how loo can be used to make decisions about a model: A quick note what I infer from p_loo and Pareto k values