nx_analyst_causal_gate.nx

nx_analyst_causal_gate.nx -- F1005. The role logic is trivial to assert; what this gate really has to prove is WHY it must exist. So the centre of it is two EMPIRICAL demonstrations on real generated data, where the identical operation -- "control for a third column" -- is right once and catastrophic twice: T6 COLLIDER: two INDEPENDENT columns acquire a strong NEGATIVE partial correlation purely by conditioning on their sum. The association is manufactured out of nothing. T7 MEDIATOR: a real, strong x->m->y effect is driven toward ZERO by conditioning on m. T8 CONFOUNDER: the same operation is CORRECT, collapsing a spurious link. F1004's ai_confound_scan picks whichever control weakens the relationship most -- which in T6 and T7 is exactly the wrong column. That is the defect this organ exists to gate. D001-compliant verdict via nx_gate_verdict. license_tier: ORIGINAL expect_exit: 0