Just trying out your suggestions now (minus rescaling; the only real data input is the response times and those are modeled by the lba distribution which is not necessarily normal).
One more relevant factor: previously I observed that chain iterations started off fast and got slower. That is the general trend, but I do sometimes see a chain speed up again. I guess that lends credence to the idea that slow chains are exploring difficult regions of the parameter space. I’m not sure if it’s a good sign that the other chains have avoided those areas entirely!?