sum_distances takes start and end, not a tour

ext/MathOptVRPTestExt.jl

@@ -326,11 +326,17 @@ function MathOptVRP.Tests.test_tsp(
 )
     inst = _tsp_instance(; seed, n)
     model = _model(optimizer_factory; time_limit)
-    JuMP.@variable(model, nodes[1:n] in MathOptVRP.Permutation(n))
-    JuMP.@objective(model, Min, MathOptVRP.op_sum_distances(inst.dist, nodes))
+    # Fix node `n` as the start/end of the tour. This breaks rotational
+    # symmetry and expresses TSP as the one-vehicle VRP special case.
+    JuMP.@variable(model, nodes[1:(n-1)] in MathOptVRP.Permutation(n - 1))
+    JuMP.@objective(
+        model,
+        Min,
+        MathOptVRP.op_sum_distances(inst.dist, [n; nodes; n]),
+    )
     JuMP.optimize!(model)
     @test JuMP.termination_status(model) in _OPTIMAL_STATUSES
-    tour = [round(Int, JuMP.value(v)) for v in nodes]
+    tour = [n; [round(Int, JuMP.value(v)) for v in nodes]]
     @test sort(tour) == collect(1:n)
     expected = sum(inst.dist[tour[k], tour[mod1(k + 1, n)]] for k = 1:n)
     @test round(Int, JuMP.objective_value(model)) == expected

src/operators.jl

@@ -1,13 +1,15 @@
-# Solver-agnostic JuMP nonlinear operator for closed-tour distance.
+# Solver-agnostic JuMP nonlinear operator for sequence distance.
 # A solver wrapper recognises `MOI.ScalarNonlinearFunction` with head
 # `:sum_distances` and lowers it into its own modelling primitives.
 
 """
     sum_distances(dist_matrix, nodes)
 
-Build a JuMP nonlinear expression representing the closed-tour cost over
-`nodes` using `dist_matrix` as the edge weights. Has no scalar fallback
-method — solvers that consume `:sum_distances` provide the lowering.
+Build a JuMP nonlinear expression equal to
+`sum(dist_matrix[nodes[i], nodes[i+1]])` over consecutive entries of `nodes`.
+Closure is explicit: pass `[first; route; first]` for a closed tour. Has no
+scalar fallback method — solvers that consume `:sum_distances` provide the
+lowering.
 """
 function sum_distances end