perf: avoid duplicate soft weight clearing

Let the model update hook own clearing and assigning per-knot soft weights. The core dispatcher now only calls the hook for structural soft models, avoiding a second L1/L2 vector clear before generated soft models run.

Harness: core5+specialist

Scope: solver-core

Evidence: pytest -q tests/minimodel/test_constraints.py tests/minimodel/test_model_safety.py; cmake --build build -j16; ctest --test-dir build --output-on-failure -j16; nmpc-bench race_cars fixed-binary A/B in /tmp/nmpc-bench-commit3-race-repeat/summary.csv; git diff --check

Author: EdwardOptimization

include/minisolver/core/constraint_semantics.h

@@ -124,8 +124,6 @@ template <typename Model, typename Knot> void update_soft_constraint_weights(Kno
 {
     if constexpr (model_may_have_l1_constraints<Model>()
         || model_may_have_l2_constraints<Model>()) {
-        kp.l1_weight.setZero();
-        kp.l2_weight.setZero();
         if constexpr (has_update_soft_constraint_weights<Model, Knot>::value) {
             Model::template update_soft_constraint_weights<double>(kp);
         }