nx_gsplat4d_gate.nx

nx_gsplat4d_gate.nx -- GATE for the 4D spatiotemporal Gaussian core. ★★THIS REFEREE CARRIES NO TUNED THRESHOLDS. An earlier revision judged with hand-picked bands -- "the 1-sigma weight is between 140 and 170", "more than 2000 pixels lit", "the arena is under 4 MiB". Every one of those is a number nobody measured, and a referee that judges with them is measuring the number, not the subject. They are all replaced by one of three honest forms: EXACT IDENTITY -- the envelope must equal the TABLE ENTRY its index formula names (the index arithmetic is the thing that can be wrong; only an exact compare tests it). A MATHEMATICAL LAW -- symmetry in dt, strict monotonicity, antisymmetry of a signed ramp. These hold regardless of table precision, so they need no tolerance at all. A CONTROL -- render the SAME time twice and require ZERO changed pixels, then a different time and require MORE than that. The control replaces the threshold entirely, and is strictly stronger: it proves the difference is caused by TIME, not by noise. The teeth that a plausible-but-wrong implementation FAILS: T4 TEMPORAL LOCALITY is the anti-vacuity tooth. Two Gaussians carry IDENTICAL motion and differ only in t_mu. At one instant one must have moved and the other must not. A deformation that ignored the temporal window -- one that just animated everything -- passes every other tooth and dies here. T1 REST INVARIANCE is the control: with no motion declared, the 4D bake reproduces the static cloud at EVERY t. 4D that cannot reduce to 3D is not a superset, it is a different thing. T9/T10/T11 are deny-guards asked BOTH directions, because A GUARD THAT REFUSES EVERYTHING PASSES EVERY NEGATIVE TEST. license_tier: ORIGINAL expect_exit: 0