Not stan_num(): that one writes %.17g because a trained weight has to
come back bit for bit, and 0.40000000000000002 in the middle of a
lognormal(...) call is noise in a block a user is meant to read. Fifteen
significant digits is more precision than any hand-written prior parameter
carries.
