nx_converge_gate.nx

nx_converge_gate.nx -- proves the convergence controller actually CONTROLS, and refuses rather than spins. It imports the same lib the shipping tool imports, so it grades the code that ships. T1 CONVERGES : lands inside tolerance on targets across the whole reachable range. T2 BOUNDED COST : never exceeds the declared render budget. The budget is the PRODUCT CLAIM -- every step is 25-50s of GPU, so an unbounded search is a broken product even if it eventually converges. T3 REFUSES FREE : an unreachable target costs ZERO renders. Refusing after burning a render is not refusing. T4 EARLY EXIT : a first probe already inside tolerance stops immediately (1 render, not a full bisection) -- proves it does not do fixed work regardless of input. T5 BRACKET SANITY : every proposed probe stays inside [CV_MIN,CV_MAX] and the bracket never inverts. A controller that proposes denoise 1300 would silently clamp at the engine and report a wrong cause. T6 DIRECTION : the inversion is correct -- measuring ABOVE target must raise the floor (more denoise), measuring BELOW must lower the ceiling. Getting this backwards still terminates, so only a direction tooth catches it. T7 DETERMINISM : identical inputs produce an identical trajectory (replayable evidence). T8 ORACLE MONOTONE : the grounded oracle is genuinely decreasing across the range -- if it were not, T1 would be proving convergence against a shape the real problem does not have, and the guarantee would be vacuous. expect_exit: 0 license_tier: ORIGINAL