The in-domain validation on subsume's trained WordNet boxes surfaced a real
gap: strict subsumers (parent box fully contains child box) is 0% because
trained Gumbel boxes nest softly -- the child's center is inside the parent but
the child's box pokes out -- while membership recovers the is-a ancestor 98%.

subsumers_soft(query, min_prob) filters candidates by entailment_prob(query) >=
min_prob (vol(S ∩ query)/vol(query)), returned ranked by probability; min_prob=1
reduces to strict containment. On the trained WordNet boxes it recovers the
ancestor at 98%, matching membership but returning ranked region-region results.
Validated by examples/wordnet_boxes and a partial-overlap unit test.