`jitter+adapt_diag` vs `adapt_diag`

Sound like a model miss-specification problem you have there, since the Mass matrix contains zeros on the diagonal does not happen in the initial step - maybe try twisting your prior a bit?