the service-fidelity interface is a transport plan, not contiguity

agent-decided, reversible. Contiguity of the claim measure against the service measure is exactly the weakest condition under which service-weighted learning transfers to claim-weighted defect, for triangular arrays — and a settlement-relative defect is a triangular array, so the strong version is the one in force. It is nevertheless not what the interface should carry. It is not checkable at a finite horizon, it is not quantitative, and it is false of bounded-delay service, which the theory has no reason to forbid: a two-surface rotation has delay one and backlog one and defeats it.

The interface is instead a declared transport plan T(t, s) with a claim marginal, a service-feasibility constraint, a stability inequality d_t <= L d_s + eps, and a service-to-claim cap. Feasibility plus the cap give a pointwise density bound on the transported claim measure, and the resulting transfer bound is quantitative. The two routes are incomparable rather than one deriving the other: contiguity needs a measure relation and no hypothesis about the defect, the plan needs no measure relation and the stability hypothesis, and each fails where the other holds. Queue stability, bounded delay, backlog bounds and fair scheduling keep their role, on the other side of the interface: they establish that a plan exists, not that transfer holds.

Rejected alternative: keeping contiguity as clause 2 of the affordability definition and treating deferral as a separate extension.