dplyr_reconstruct() is not enough on its own. dplyr calls that
generic only when the result's class differs from the input's, and
dplyr::select() on a tibble subclass subsets through [, which
keeps the class; the generic is then never reached and a
column-dropping verb returns something still labelled a recovery
object with its contract broken. Measured on dplyr 1.2.1: without this
method dplyr::select(x, "term", "estimate") stays a
bmmtools_recovery and the next print() fails inside a metric with
a length mismatch. Demoting here is what makes the documented
behaviour — a verb that drops a contract column drops the class —
actually true.