out is the coverage table before sorting, with noise only
when a noise pattern was given. Testedness is judged per term from
term_ok, not from out$tested: a factor's rolled-up coverage
is the union of its levels' trees, so it can pass trim while every
level, and so every level's z, was trimmed away. Returns
invisible(NULL) when the screen passes.
.varpro_coverage_guard(out, max_untested, trim, term_var, term_ok)invisible(NULL); called for its error.