nx_provclass_gate.nx

nx_provclass_gate.nx -- proves the provenance-class ruler tightens without breaking anything. THE RISK THIS HOLDS DOWN. nx_provclass_lib adds a SECOND admission axis beside nx_refcorpus's payload wall. A second wall in front of a live journal is exactly the change that silently refuses work that used to pass -- and the live journal already carries FIVE licence vocabularies nobody declared (ORIGINAL, sovereign, royalty-free, CC-BY-mixed, REFERENCE_MEASURED). So the load-bearing tooth here is not "does the classifier work"; it is T2: EVERY DISTINCT LICENCE TAG IN THE LIVE JOURNAL STILL CLASSIFIES. That is a POPULATION check over the real file, not a hand-listed sample, because a name list of tags would go stale the same way every name list in this estate has. THE DISCRIMINATING PAIR is T5/T6: the SAME tag `unverified` must REFUSE at n=8 and ADMIT at n=32. n=8 clears nx_refcorpus's global RC_KMIN, so if the raised per-class floor were not actually binding, T5 would pass anyway and the whole class distinction would be decoration. One tag, two sample sizes, opposite verdicts -- that is the only shape that proves the floor binds. T9 is the STRICTLY-ADDITIVE proof: no declared class may carry a floor BELOW RC_KMIN=8. This lib is allowed to tighten admission and never to loosen it, and a conf edit is exactly how someone would loosen it by accident. BITE NOTE 2026-09-03: on its FIRST run (laptop, no journal present) this gate went 15/16 RED on `journal-population-was-actually-read` while `every-live-journal-licence-tag-classifies` scored a VACUOUS actual=0 expected=0 PASS. The anti-vacuity tooth caught its own author on the first run, which is exactly the case it exists for. license_tier: ORIGINAL expect_exit: 0