Publication-grade audit of the DS18 exact off-(L) basin and empirical transfer

Independent adversarial audit of DS-NONCENTERED-GLOBAL-BASIN-TRANSFER. Verdict: verified with hardened assumptions. The population optimum, its uniqueness a.s. up to labels and null sets, quantitative strict isolation in decision distance, the almost-sure transfer of every sequence of exact finite global optimizers on one selection-independent probability-one event, finite ordinary one-point exchange stability, and the fixed margins (1/4, 1/4, 1/2) all survive. Five hardenings: (H1) the proof no longer imports the (L)-conditioned DS15 Lemma 3 and Proposition 5, which DS18 cited outside their registered hypothesis set, and proves the scalar uniqueness, rigidity and consistency steps for this law directly; (H2) the in-bin DS9 feasibility convention is named, and a zero binned nuisance block is shown to force sum_i S_lambda_i = 0 and hence to be null under the law, so regularity is a.s. vacuous while remaining load-bearing off that null set; (H3) uniqueness is a.s. up to labels and null sets, with the half-open reference labeling stated; (H4) the probability-one event is exhibited explicitly and the transfer is given a computable finite-N form P_N(z != q*) <= 3 Delta_N / eta + boundary band mass with Delta_N = v_hat_3 - Phi_hat(z*); (H5) DS14's (M3) is lambda_min(I_hat) >= kappa, not det/tr, which the margin triple needs. Two prior-art repairs: six direct antecedents already in the project bibliography were unlinked, and the uniqueness citation was mis-scoped to a strict-log-concavity result that Uniform[-1,1] fails.