Stateful coordination
The coordination prototype applies Lagrangian prices to annual budget and crew-hour constraints. Cluster subproblems remove those constraints, add non-negative prices to the objective, and are solved exactly by the stateful MILP. A subgradient update raises prices when aggregate spend or hours exceed the global limits. The maximum dual value is a valid lower bound.
Result on the network
The prototype is valid but vacuous on the current stateful model. At zero price, each cluster treats all risk to zero, so the dual value is zero. Raising prices removes treatments and produces negative dual values. On both the 59-segment and 590-segment cases the best dual was 0.0 against a primal objective near 0.029, a 100 percent relative gap.
This is a property of the relaxed integer problem and the linear risk proxy, not an implementation error. The method is retained as a documented negative result; it is not wired into the studio or used for allocation.
What it would take
A useful global coordinator needs a stronger coupling method:
Benders cuts from cluster MILPs;
value functions over annual budget, combined by a master allocation;
or column generation with cluster plans as columns.
A second subgradient implementation is not the answer.