Thanks for the leads!
It turns out, as you suggested, that both my troubles here, and in the previous thread were such that any change to the model or the seed moved me away from the very particular trajectories I (un)luckily happened to stumble upon at first try.
The solution was to give initial values for the ndt parameter, which is non decision time for the Wiener likelihood. The way the mixture model is set up, if rt < ndt, only the uniform likelihood component contributes to the likelihood. If the initial value of ndt is too high, too much of the data is assigned only the uniform component, and so the shape of the data does nothing to inform the fitting. Hence, the sampler can’t leave that area in parameter space, I guess.
Setting very small initial values for ndt (e.g. 0.01) solves this.