lingauss_mean() is the only place x meets de$B, so it is also the
place the width of x is checked. It passes de$dim_x to
as_theta_matrix() for the reason every neural estimator passes its own: a
wrong-width x otherwise gets as far as the matrix product and is reported
as "non-conformable arguments", which names neither the argument nor the
width expected of it. An estimator fitted before dim_x was recorded has
NULL here and keeps the old unchecked behaviour.