nx_mmbench_gate.nx

nx_mmbench_gate.nx -- proves the media+companion ruler is a RULER, not a rubber stamp. It imports nx_mmbench_lib -- THE SAME lib nx_mmbench ships -- so these teeth bite the code that actually produces the published number, not a re-implementation of it. The ruler's entire honesty claim is aud(). If aud() does not really downgrade an unbacked claim, then coverage_permil is decoration and every gap-queue ranking derived from it is fiction. The first live run reported liar_killed=0, which is exactly the reading a BROKEN audit would also produce -- indistinguishable from a healthy one without these tests. Hence: T1 CLEAN : HAVE + an artifact that exists -> stays HAVE. (the ruler can say YES) T2 NEGATIVE CTRL : HAVE + an artifact that does NOT exist -> GAP, flagged liar-killed. THE anti-vacuity tooth. Without this, liar_killed=0 proves nothing. T3 UNGROUNDED : HAVE + no artifact named at all -> capped to PARTIAL, flagged ungrounded. T4 KILL-BEATS-RANK: PARTIAL + absent artifact -> GAP too. A false SPECIFIC claim is killed regardless of how modest the declared level was. T5 SILENCE-OK : PARTIAL + no artifact -> stays PARTIAL, unflagged. Declining to claim is not a lie; only asserting an unbacked WIN is. T6 ONE-WAY : GAP + a present artifact -> stays GAP. Evidence may only ever DEMOTE. A gate that let evidence PROMOTE could launder any gap into a win. T7/T8 PROBE : have() is truthful in both directions (present->1, absent->0). T9 STREQ : verb matching is exact -- a prefix must NOT match, or 'json' would fire on 'jsonx' and silently change which report is published. T10 MONOTONIC : a liar-kill must LOWER the computed headline. Proves the demotion actually reaches the arithmetic instead of just printing a warning next to it. expect_exit: 0 license_tier: ORIGINAL No hw writes (Rule 26).