Thanks! I have no idea what QR decomposition does, it was just advised in this thread that it could speed up sampling in hierarchical models.
The model in my example is not hierarchical, so disabling QR decomposition solves the problem. But, I will soon have to start fitting hierarchical models to this same data, with 3 to 6 group-level SDs. Does this mean that I will run into the same problem with those hierarchical if I use QR decomposition? Or is this a problem that will only arise when using QR decomposition with non-hierarchical models?