Stateful heuristic ================== The stateful MILP is exact for small clusters, but its binary count grows with segments × periods × methods. For larger networks Brushline uses the optional `fresh-solutions` local-search heuristic. Adapter ------- `StatefulHeuristicAdapter` exposes one discrete variable per segment-period. Options are `none`, `mowing`, `manual`, and any other declared method. Evaluation replays the same linear risk state, treatment transitions, budget, and crew-hour constraints as the MILP. Search starts from a deterministic risk-per-dollar greedy seed, then uses Fresh Solutions simulated annealing. The result is explicitly marked **feasible**, not optimal. Every incumbent is replayed; the plan records the Fresh Solutions version, seed, iterations, evaluations, and accepted moves. Scaling evidence ---------------- The network demo includes replicated stress cases at roughly 59, 590, 2,950, and 10,030 segments. On the current dev server the 10,030-segment, 50,150-variable case completes in about 11 seconds with 20 local-search iterations after the greedy seed. The exact MILP remains the benchmark for the 59-segment case. Limits ------ - The heuristic does not prove optimality. - The linear risk state remains an abstraction. - The greedy seed may dominate the search if the local-search budget is small. - Access, weather, labour rules, and asset-owner standards are separate constraints.