Thanks for the reply!
Regarding re-parameterization: as far as I can tell from reading the literature on this problem (time series with both structural breaks and additive outliers), there’s no way around the correlated variables, and the solution is specialized sampling algorithms (e.g. this paper).