nx_refbench_gate.nx

nx_refbench_gate.nx -- proves the reference-retention ruler is a RULER and not a rubber stamp. It imports the SAME lib the shipping tool imports, so it grades the code that ships. T1 IDENTITY : an image against ITSELF must score 1000 on BOTH channels. T2 DISCRIMINATION : a MIRRORED copy (identical histogram, different structure) must score materially LOWER. This is the anti-vacuity tooth -- a ruler that has never been seen to say NO is not a ruler. T3 SCALE INVARIANCE: the same content at a DIFFERENT SIZE must still score high. This is the real production case (2550x3300 reference vs 768x1024 render); if it failed, every retention number would be measuring resolution, not content. T4 MONOTONICITY : increasing corruption must produce MONOTONICALLY FALLING retention. The whole L3 denoise curve is worthless if the ruler does not ORDER correctly -- this tooth is what licenses reading the curve as a curve. T5 MIN-HEADLINE : retention must equal the WEAKEST channel, never the mean. Proven with a planted split (strong one side, weak the other) so a good channel can never flatter a bad one. T6 FAIL-CLOSED : an unreadable/absent image must return an ERROR, never a score. Two failed decodes must NOT both become zeroed planes and score a perfect 1000 against each other -- the blank-decoy trap that already fooled an artifact tooth in this ecosystem (a failing producer wrote a valid PNG of an empty render). expect_exit: 0 license_tier: ORIGINAL