A prior's mean is data the refit reads, but a self test does not redraw it with the
observations, so where the data pulled the fit away from a prior every replicate is pulled
back toward it, a bias of the design rather than of the estimator. "assessment" keeps
the priors as the assessment gives them; "truth" moves the mean of every catchability,
natural mortality, steepness, R0 and selectivity prior to the value the replicate ran on,
keeping its sd, so the prior keeps its information and loses its pull; "off" turns every
prior off, which leaves a parameter only a prior identifies unidentified. A Dirichlet (on
movement and on the recruitment apportionment) or a mean and sd beta (on the stray rate and the
tag reporting rate) is moved the same way, its mode put on the value the replicate ran on and
its concentration kept, so it has no gradient there (dirichlet_mode_at,
beta_mode_at); one with no interior mode is left as given, as is the symmetric
beta on the tag reporting rate, which sits at one half by construction.