Hi, I think this discussion and Stan code for the GEV may be useful to you.
We discuss the implementation in a little more detail in this Rpubs and this brms issue report.
Broadly speaking, the trick, as @avehtari made elegant use of in the case study with the generalized pareto, is to reparameterize the distribution such that support constraints can be forced onto the shape parameter. This allows the bounds to be expressed as a continuously differentiable function of the other parameters (and the data). The quantile-based reparameterization used by Castro-Camilo et al., 2022 is particularly useful for the GEV.
Hope this is helpful