nle() and nre() fits are interchangeable everywhere downstream of the
estimator, and they differ in exactly three things: which function produces
the n_theta x n_obs matrix of scores, which one produces its row sums
without building it, and whether standardizing x needs a
change-of-variables term. A density reported in the simulator's units needs
one; a ratio does not, since the Jacobian cancels between its numerator and
denominator (see nre()).
Details
Keeping the three together in one table is what lets surrogate_score() and
surrogate_potential() be plain functions rather than a generic each. It
also means the fact that a ratio has no Jacobian is written down once.