Whether a response is discrete cannot be read off a density: the argument
names say nothing about it, and probing the CDF numerically is unreliable
because implementations differ in whether they floor a non-integer
argument. So it is declared, as in LaMa's pseudo_res(), and
fam() takes a support override for anything not listed.
Details
Every family here is supported on the non-negative integers, which is what
lets residuals.gamRTMB() take F(y - 1) = 0 at y = 0 rather than
evaluating the CDF below its support.