Step 2: FenceModel calibration replaces hardcoded ±2.5m

- fence_model.hpp: Line2D + FenceModel with wall positions
- session_mapper: RANSAC yaw fit + percentile boundary + cone clustering
- localizer: setFenceModel(), use calibrated walls in correctWithScanCV
- main.cpp: CALIBRATING→READY state machine (30 scan frames)
- /nav_metrics publisher (5Hz JSON metrics)
- mppi_alive/min_cost exposed via LocalPlanner public members
- endpoint-only path_align (EXACT-MPPI PathFollowCritic style)
- adaptive exploration ratio based on survival rate

CMakeLists.txt

@@ -25,6 +25,7 @@ add_executable(nav_lite_node
   src/global_planner.cpp
   src/local_planner.cpp
   src/bt_executor.cpp
+  src/session_mapper.cpp
 )
 
 target_include_directories(nav_lite_node PRIVATE