It turns out this option is very problematic. For example, we can't take derivatives with respect the scale_radius of the moffat profile if trunc is non-zero. The issue is that convolutions track whether or not profiles have hard edges and this causes it to attempt to branch on a traced value.