feat(perception): stage GSeg3D qualification

docs/23_TRAVERSABILITY_AND_STATIC_OBSTACLE_STACK_DECISION.md

@@ -33,8 +33,13 @@ The sequential evaluation order is:
 
 RF-DETR remains the semantic risk provider for people, children, animals,
 cars, trucks, motorcycles and bicycles. It does not decide the presence or
-shape of an unknown static obstacle. Native `800x600` fisheye input remains
-unchanged.
+shape of an unknown static obstacle. The admitted camera semantic layer owns
+surface/material hints such as low vegetation and woody vegetation. Native
+`800x600` fisheye input remains unchanged.
+
+There is no single globally primary sensor. The camera is primary for semantic
+meaning; LiDAR is primary for metric geometry and positive support. Neither is
+allowed to erase contradictory evidence from the other.
 
 ## Required product representation
 
@@ -55,12 +60,42 @@ robot-footprint collision cost + traversability
         +----> camera projection (diagnostic visualization only)
 
 camera RF-DETR ----> dynamic semantic risk hints ----------------+
+camera semantics --> material / vegetation hint -----------------+
 ```
 
 The canonical safety output is a costmap/elevation/occupancy product. Camera
 boxes may be derived for operator review, but a box must never be the source of
 occupied cells, clearance, passage width or a control action.
 
+## Camera/LiDAR arbitration and vegetation
+
+Tall grass demonstrates why a binary obstacle mask is insufficient. LiDAR may
+correctly report many non-ground returns while being unable to tell whether
+they came from compressible grass, a rigid trunk hidden inside the grass, or a
+drop-off with no positive support. Camera semantics can identify vegetation,
+but a camera label cannot prove that the ground below is load-bearing or free
+of a hidden rigid object.
+
+Mission Core therefore preserves at least these distinct states:
+
+| Evidence | Traversability state |
+| --- | --- |
+| Stable metric non-ground evidence without vegetation support | `RIGID_OR_UNKNOWN_OBSTACLE`; lethal or unknown according to support confidence |
+| Camera vegetation plus diffuse LiDAR returns, but no positive ground support below | `VEGETATION_UNKNOWN`; never silently free |
+| Camera low vegetation plus positive local ground support and no concentrated rigid return | `VEGETATION_POTENTIALLY_TRAVERSABLE`; high-cost terrain, not a wall |
+| Camera vegetation plus concentrated persistent vertical geometry | `VEGETATION_WITH_RIGID_GEOMETRY`; lethal |
+| Positive ground support without blocking geometry | `SUPPORTED_GROUND`; candidate traversable terrain |
+
+`VEGETATION_POTENTIALLY_TRAVERSABLE` is a representation decision, not an
+authorization to drive through it. The later robot profile must still declare
+ground clearance, footprint, traction, admissible slope, vegetation height and
+whether contact with vegetation is allowed. Until that profile exists, the
+planner may prefer a supported hard path and retain vegetation as a higher-cost
+or unknown alternative.
+
+The executable evidence contract is
+[`config/perception/m49-camera-lidar-traversability-v1.json`](../config/perception/m49-camera-lidar-traversability-v1.json).
+
 ## Candidate A — GSeg3D and Ground Consistency
 
 ### GSeg3D
@@ -78,6 +113,11 @@ publishes separate ground and obstacle point clouds and provides:
 Mission Core evaluates it as the first replacement for the current local
 height connected-component heuristic.
 
+The upstream GSeg3D documentation explicitly lists recall degradation in dense
+vegetation as a remaining challenge and semantic-aware refinement as future
+work. Candidate A can therefore qualify the metric ground/non-ground seam, but
+cannot alone close the vegetation product gate.
+
 ### Nav2 Ground Consistency
 
 [DFKI Ground Consistency](https://github.com/dfki-ric/nav2_ground_consistency_costmap_plugin)
@@ -246,6 +286,8 @@ failure.
 - Review the ten mandatory anchors and compute full-run stability/resource
   metrics.
 - Do not draw system camera boxes until the costmap result is sealed.
+- Replay the existing camera semantic vegetation evidence as an independent
+  arbitration input; do not let it clear unsupported or concentrated geometry.
 
 ### T3 — Candidate B challenger