fix(lab): show only actionable obstacle boxes

apps/control-station/src/core/laboratory/m4ReplayThreat.ts

@@ -186,7 +186,7 @@ export interface M4ThreatTimeline {
   cameraPointSampleLimit: number;
   worldStateDelivery: "source-paced-latest-wins" | null;
   occupancyProvenanceDelivery: "baseline-versus-additive-component-diff" | null;
-  cameraObstacleProjectionDelivery: "factory-kb4-occupied-voxel-bounds" | null;
+  cameraObstacleProjectionDelivery: "factory-kb4-actionable-added-corridor-bounds" | null;
   worldStateFrameCount: number;
   supersededFrameCount: number;
   sourceRepresentationId: "registered-map-increment-v1";
@@ -747,7 +747,7 @@ export async function fetchM4ThreatTimeline(
       ? null
       : exact(
           payload.camera_obstacle_projection_delivery,
-          "factory-kb4-occupied-voxel-bounds",
+          "factory-kb4-actionable-added-corridor-bounds",
           "M4.8R3 camera obstacle projection delivery",
         ),
     worldStateFrameCount: payload.world_state_frame_count === undefined

apps/control-station/src/workspaces/laboratory/M4ReplayThreatVisual.tsx

@@ -513,7 +513,7 @@ export function M4ReplayThreatVisual({
           shape="pill"
           variant={showStaticObstacles ? "primary" : "secondary"}
           aria-pressed={showStaticObstacles}
-          title="Автоматические рамки занятых LiDAR-компонентов · без ручной разметки"
+          title="Только LOW-STEP LiDAR-препятствия с решением threat · без ручной разметки"
           onClick={() => setShowStaticObstacles((visible) => !visible)}
         >
           OBSTACLES

apps/control-station/src/workspaces/laboratory/m4StaticObstacleBoxes.ts

@@ -2,6 +2,12 @@ import type { RecordedEvidenceBox } from "../../components/laboratory/RecordedEv
 import type { M4ThreatMetricVisual } from "../../core/laboratory/m4ReplayThreat";
 
 const MINIMUM_BOX_SIZE_PX = 12;
+const RETAINED_CURRENT_OVERLAP_LIMIT = 0.5;
+
+interface ObstacleBoxCandidate {
+  readonly box: RecordedEvidenceBox;
+  readonly state: M4ThreatMetricVisual["state"];
+}
 
 function shortComponentId(componentId: string): string {
   const finalSegment = componentId.split(/[-_]/).pop();
@@ -37,26 +43,42 @@ function visibleBox(
   return [left, top, right, bottom];
 }
 
-function tone(obstacle: M4ThreatMetricVisual): RecordedEvidenceBox["tone"] {
+function overlapFractionOfSmaller(
-  if (obstacle.assessment.decision === "threat") return "danger";
+  left: readonly [number, number, number, number],
-  if (obstacle.assessment.decision === "not-threat") return "success";
+  right: readonly [number, number, number, number],
-  return "warning";
+): number {
+  const intersectionWidth = Math.max(
+    0,
+    Math.min(left[2], right[2]) - Math.max(left[0], right[0]),
+  );
+  const intersectionHeight = Math.max(
+    0,
+    Math.min(left[3], right[3]) - Math.max(left[1], right[1]),
+  );
+  const intersectionArea = intersectionWidth * intersectionHeight;
+  const leftArea = (left[2] - left[0]) * (left[3] - left[1]);
+  const rightArea = (right[2] - right[0]) * (right[3] - right[1]);
+  const smallerArea = Math.min(leftArea, rightArea);
+  return smallerArea > 0 ? intersectionArea / smallerArea : 0;
 }
 
 /**
- * Build camera evidence from the same world-state components shown in 3D.
+ * Build actionable camera evidence from the same world-state shown in 3D.
- * No image detector or manual review extent participates in these boxes.
+ * Unknown and clear components remain in world-state, but do not compete with
+ * actual corridor threats for the operator's attention.
  */
 export function buildM4StaticObstacleBoxes(
   obstacles: readonly M4ThreatMetricVisual[],
   imageWidth: number,
   imageHeight: number,
 ): readonly RecordedEvidenceBox[] {
-  const result: RecordedEvidenceBox[] = [];
+  const candidates: ObstacleBoxCandidate[] = [];
   for (const obstacle of obstacles) {
     if (
       obstacle.occupancySource === "baseline"
       || (obstacle.state !== "current" && obstacle.state !== "retained")
+      || obstacle.assessment.decision !== "threat"
+      || obstacle.assessment.corridorIntersection !== "intersects"
       || obstacle.cameraProjection === null
     ) continue;
     const projection = obstacle.cameraProjection;
@@ -65,16 +87,33 @@ export function buildM4StaticObstacleBoxes(
     const depth = projection.nearestDepthM.toLocaleString("ru-RU", {
       maximumFractionDigits: 1,
     });
-    result.push({
+    candidates.push({
-      boxXyxy,
+      state: obstacle.state,
-      label: `OBS #${shortComponentId(obstacle.componentId)} · ${depth} м`,
+      box: {
-      tone: tone(obstacle),
+        boxXyxy,
-      dashed: false,
+        label: `OBS #${shortComponentId(obstacle.componentId)} · ${depth} м`,
+        tone: "danger",
+        dashed: false,
+      },
     });
   }
-  return result.sort((left, right) => {
+  const currentBoxes = candidates
-    const leftArea = (left.boxXyxy[2] - left.boxXyxy[0]) * (left.boxXyxy[3] - left.boxXyxy[1]);
+    .filter((candidate) => candidate.state === "current")
-    const rightArea = (right.boxXyxy[2] - right.boxXyxy[0]) * (right.boxXyxy[3] - right.boxXyxy[1]);
+    .map((candidate) => candidate.box.boxXyxy);
-    return rightArea - leftArea;
+  return candidates
-  });
+    .filter((candidate) => (
+      candidate.state !== "retained"
+      || !currentBoxes.some((currentBox) => (
+        overlapFractionOfSmaller(candidate.box.boxXyxy, currentBox)
+        >= RETAINED_CURRENT_OVERLAP_LIMIT
+      ))
+    ))
+    .map((candidate) => candidate.box)
+    .sort((left, right) => {
+      const leftArea = (left.boxXyxy[2] - left.boxXyxy[0])
+        * (left.boxXyxy[3] - left.boxXyxy[1]);
+      const rightArea = (right.boxXyxy[2] - right.boxXyxy[0])
+        * (right.boxXyxy[3] - right.boxXyxy[1]);
+      return rightArea - leftArea;
+    });
 }

apps/control-station/test/m4ReplayThreat.test.mjs

@@ -370,7 +370,7 @@ test("M4.8S timeline binds factory-KB4 camera points through its exact endpoint"
         camera_point_sample_limit: 20000,
         world_state_delivery: "source-paced-latest-wins",
         occupancy_provenance_delivery: "baseline-versus-additive-component-diff",
-        camera_obstacle_projection_delivery: "factory-kb4-occupied-voxel-bounds",
+        camera_obstacle_projection_delivery: "factory-kb4-actionable-added-corridor-bounds",
         world_state_frame_count: 4481,
         superseded_frame_count: 8,
         local_surface_visualization: {
@@ -400,7 +400,7 @@ test("M4.8S timeline binds factory-KB4 camera points through its exact endpoint"
   assert.equal(timeline.supersededFrameCount, 8);
   assert.equal(
     timeline.cameraObstacleProjectionDelivery,
-    "factory-kb4-occupied-voxel-bounds",
+    "factory-kb4-actionable-added-corridor-bounds",
   );
 
   const chunk = await fetchM4ThreatTimelineChunk(replayResultId, 1, 1, {
@@ -455,7 +455,7 @@ test("M4.8S timeline binds factory-KB4 camera points through its exact endpoint"
   assert.match(requested, new RegExp(`^${endpointRoot}/${replayResultId}/timeline/chunk`));
   assert.match(
     requested,
-    /obstacle_projection=factory-kb4-occupied-voxel-bounds/,
+    /obstacle_projection=factory-kb4-actionable-added-corridor-bounds/,
   );
   assert.equal(chunk.frames[0].worldStateAvailable, false);
   assert.equal(chunk.frames[0].terminalOutcome, "superseded");
@@ -525,6 +525,24 @@ test("M4.8R3 turns active low-step components into native fisheye obstacle boxes
     obstacle,
     { ...obstacle, componentId: "baseline", occupancySource: "baseline" },
     { ...obstacle, componentId: "held", state: "held" },
+    {
+      ...obstacle,
+      componentId: "unknown",
+      assessment: {
+        ...obstacle.assessment,
+        decision: "unknown",
+        corridorIntersection: "unknown",
+      },
+    },
+    {
+      ...obstacle,
+      componentId: "clear",
+      assessment: {
+        ...obstacle.assessment,
+        decision: "not-threat",
+        corridorIntersection: "clear",
+      },
+    },
   ], 800, 600);
   assert.equal(boxes.length, 1);
   assert.deepEqual(boxes[0].boxXyxy, [394, 294, 406, 306]);
@@ -533,6 +551,59 @@ test("M4.8R3 turns active low-step components into native fisheye obstacle boxes
   assert.equal(boxes[0].dashed, false);
 });
 
+test("M4.8R3 keeps separate CURRENT threats and suppresses their spanning ROLLING box", () => {
+  const current = {
+    componentId: "temporal-left-1",
+    state: "current",
+    motion: "stationary",
+    centroidBodyXyzM: [3, 0, 0.3],
+    cellCentersBodyXyzM: [[3, 0, 0.3]],
+    occupancySource: "additive-low-step",
+    cameraProjection: {
+      bboxXyxy: [350, 260, 410, 310],
+      nearestDepthM: 3,
+      projectedCellCount: 2,
+      projection: "factory-kb4-occupied-voxel-bounds",
+      authority: "visual-derived",
+    },
+    assessment: {
+      componentId: "temporal-left-1",
+      decision: "threat",
+      corridorIntersection: "intersects",
+      relativeSpeedMps: null,
+      closestApproachM: 2.5,
+      ttcSeconds: null,
+      reasonCodes: ["current-corridor-intersection"],
+    },
+  };
+  const boxes = buildM4StaticObstacleBoxes([
+    current,
+    {
+      ...current,
+      componentId: "temporal-right-2",
+      cameraProjection: {
+        ...current.cameraProjection,
+        bboxXyxy: [405, 265, 445, 320],
+      },
+    },
+    {
+      ...current,
+      componentId: "rolling-spanning",
+      state: "retained",
+      cameraProjection: {
+        ...current.cameraProjection,
+        bboxXyxy: [340, 255, 450, 325],
+      },
+    },
+  ], 800, 600);
+
+  assert.equal(boxes.length, 2);
+  assert.deepEqual(boxes.map((box) => box.label).sort(), [
+    "OBS #1 · 3 м",
+    "OBS #2 · 3 м",
+  ]);
+});
+
 test("M4.6 local SLAM surface reprojects registered increments into the active body frame", () => {
   const frames = [
     timelineFrame(0, 10, {

src/k1link/perception/m48s_replay_timeline.py

@@ -27,6 +27,8 @@ from .spatial_evidence import (
 from .threat import (
     DEFAULT_REPLAY_THREAT_PROFILE_PATH,
     RecordedReplayBodyFrameResolver,
+    ReplayBodyFrame,
+    ReplayThreatProfile,
     load_replay_threat_profile,
 )
 from .threat_timeline import (
@@ -49,8 +51,12 @@ EXPECTED_FRAME_COUNT: Final = 4_489
 CAMERA_ACCUMULATION_WINDOW_SECONDS: Final = 2.0
 CAMERA_ACCUMULATION_POINT_LIMIT: Final = 20_000
 CAMERA_POINT_OVERLAY_SCHEMA: Final = "missioncore.m48s-camera-point-overlay/v1"
+ACTIONABLE_CAMERA_OBSTACLE_PROJECTION: Final = (
+    "factory-kb4-actionable-added-corridor-bounds"
+)
 _SOURCE_ENVELOPE_MARKER: Final = b'"source_envelope":'
 _JSON_DECODER: Final = json.JSONDecoder()
+Cell = tuple[int, int, int]
 
 
 class M48sReplayTimelineError(RuntimeError):
@@ -62,6 +68,12 @@ class _LedgerIndex:
     offsets_by_sequence: dict[int, int]
 
 
+@dataclass(frozen=True, slots=True)
+class _FrameDiff:
+    component_provenance: dict[str, str]
+    added_cells: frozenset[Cell]
+
+
 class M48sReplayTimeline:
     """Read source-indexed chunks while preserving latest-wins world-state gaps."""
 
@@ -188,7 +200,7 @@ class M48sReplayTimeline:
                 else None
             ),
             "camera_obstacle_projection_delivery": (
-                "factory-kb4-occupied-voxel-bounds"
+                ACTIONABLE_CAMERA_OBSTACLE_PROJECTION
                 if self.frame_diff_path is not None
                 else None
             ),
@@ -435,19 +447,31 @@ class M48sReplayTimeline:
                     body_frame,
                     occupied_voxel_size_m=self.profile.corridor.occupied_voxel_size_m,
                 )
-                provenance = self._component_provenance(sequence)
+                frame_diff = self._frame_diff(sequence)
+                provenance = (
+                    {} if frame_diff is None else frame_diff.component_provenance
+                )
+                actionable_metric_rows = (
+                    []
+                    if frame_diff is None
+                    else _actionable_camera_metric_rows(
+                        metric_rows,
+                        added_cells=frame_diff.added_cells,
+                        body_frame=body_frame,
+                        profile=self.profile,
+                    )
+                )
                 camera_projections = (
                     {}
-                    if frame is None or not provenance
+                    if frame is None or not actionable_metric_rows
                     else project_metric_obstacles_to_camera(
-                        metric_rows,
+                        actionable_metric_rows,
                         position_map_xyz=frame.sensor_position_map,
                         orientation_map_from_lidar_xyzw=frame.sensor_orientation_xyzw,
                         profile=frame.projection,
                         occupied_voxel_size_m=(
                             self.profile.corridor.occupied_voxel_size_m
                         ),
-                        component_ids=set(provenance),
                     )
                 )
                 for visual in metric_visuals:
@@ -537,9 +561,9 @@ class M48sReplayTimeline:
             raise M48sReplayTimelineError("M4.8S frame row is invalid")
         return value
 
-    def _component_provenance(self, sequence: int) -> dict[str, str]:
+    def _frame_diff(self, sequence: int) -> _FrameDiff | None:
         if self.frame_diff_path is None:
-            return {}
+            return None
         offset = self.frame_diff_offsets.get(sequence)
         if offset is None:
             raise M48sReplayTimelineError("M4.8R3 frame diff is incomplete")
@@ -556,7 +580,88 @@ class M48sReplayTimeline:
             for key, item in provenance.items()
         ):
             raise M48sReplayTimelineError("M4.8R3 component provenance changed")
-        return provenance
+        raw_added_cells = value.get("added_cells")
+        if not isinstance(raw_added_cells, list):
+            raise M48sReplayTimelineError("M4.8R3 added cells changed")
+        added_cells: set[Cell] = set()
+        for raw_cell in raw_added_cells:
+            if (
+                not isinstance(raw_cell, list)
+                or len(raw_cell) != 3
+                or any(not isinstance(item, int) or isinstance(item, bool) for item in raw_cell)
+            ):
+                raise M48sReplayTimelineError("M4.8R3 added cell is invalid")
+            added_cells.add((raw_cell[0], raw_cell[1], raw_cell[2]))
+        return _FrameDiff(
+            component_provenance=provenance,
+            added_cells=frozenset(added_cells),
+        )
+
+
+def _actionable_camera_metric_rows(
+    metric_rows: list[dict[str, object]],
+    *,
+    added_cells: frozenset[Cell],
+    body_frame: ReplayBodyFrame,
+    profile: ReplayThreatProfile,
+) -> list[dict[str, object]]:
+    """Keep only newly added voxel cells that cause a current corridor threat.
+
+    A rolling component can join spatially distant baseline and LOW-STEP cells.
+    Projecting its complete envelope produces an honest component bound but not
+    an operator-usable obstacle box.  The camera layer therefore visualizes the
+    exact added cells that participate in the already accepted corridor
+    intersection; threat authority and the complete 3D component stay intact.
+    """
+
+    voxel_size_m = profile.corridor.occupied_voxel_size_m
+    expansion_m = voxel_size_m * math.sqrt(2) / 2
+    minimum_x = -(
+        profile.rig.body_length_m / 2
+        + profile.corridor.rear_margin_m
+        + expansion_m
+    )
+    maximum_x = (
+        profile.rig.body_length_m / 2
+        + profile.corridor.forward_length_m
+        + expansion_m
+    )
+    half_width = (
+        profile.rig.body_width_m / 2
+        + profile.corridor.lateral_clearance_m
+        + expansion_m
+    )
+    actionable: list[dict[str, object]] = []
+    for row in metric_rows:
+        assessment = _object(row.get("assessment"), "metric assessment")
+        if (
+            assessment.get("decision") != "threat"
+            or assessment.get("corridor_intersection") != "intersects"
+        ):
+            continue
+        raw_cells = row.get("cells")
+        if not isinstance(raw_cells, list):
+            raise M48sReplayTimelineError("M4.8R3 metric cells changed")
+        selected_cells: list[dict[str, object]] = []
+        for raw_cell in raw_cells:
+            cell = _object(raw_cell, "metric cell")
+            indices = (
+                _signed_integer(cell.get("x"), "metric cell x"),
+                _signed_integer(cell.get("y"), "metric cell y"),
+                _signed_integer(cell.get("z"), "metric cell z"),
+            )
+            if indices not in added_cells:
+                continue
+            center_map = tuple((index + 0.5) * voxel_size_m for index in indices)
+            center_body = body_frame.map_point_to_body(center_map)
+            if (
+                minimum_x <= center_body[0] <= maximum_x
+                and -half_width <= center_body[1] <= half_width
+            ):
+                selected_cells.append(cell)
+        if selected_cells:
+            actionable.append({**row, "cells": selected_cells})
+    return actionable
 
 
 def _index_ledger(
@@ -708,7 +813,14 @@ def _text(value: object, label: str) -> str:
     return value
 
 
+def _signed_integer(value: object, label: str) -> int:
+    if not isinstance(value, int) or isinstance(value, bool):
+        raise M48sReplayTimelineError(f"M4.8S {label} is invalid")
+    return value
+
+
 __all__ = [
+    "ACTIONABLE_CAMERA_OBSTACLE_PROJECTION",
     "CAMERA_ACCUMULATION_POINT_LIMIT",
     "CAMERA_ACCUMULATION_WINDOW_SECONDS",
     "CAMERA_POINT_OVERLAY_SCHEMA",

tests/test_m48s_replay_timeline.py

@@ -0,0 +1,53 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+from k1link.perception.m48s_replay_timeline import _actionable_camera_metric_rows
+from k1link.perception.threat import ReplayBodyFrame, load_replay_threat_profile
+
+REPOSITORY_ROOT = Path(__file__).resolve().parents[1]
+
+
+def test_camera_obstacles_keep_only_added_cells_inside_an_accepted_threat() -> None:
+    profile = load_replay_threat_profile(
+        REPOSITORY_ROOT / "config/perception/m4-replay-threat-v3.json"
+    )
+    body_frame = ReplayBodyFrame(
+        frame_id="frame-000001",
+        origin_map_xyz_m=(0.0, 0.0, 0.0),
+        basis_map_from_body=((1.0, 0.0, 0.0), (0.0, 1.0, 0.0), (0.0, 0.0, 1.0)),
+        sensor_height_m=1.25,
+        surface_slope_deg=0.0,
+        forward_source="test",
+        camera_forward_alignment_deg=0.0,
+    )
+    threat = {
+        "component_id": "rolling-actionable",
+        "cells": [
+            {"x": 0, "y": 0, "z": 0},
+            {"x": 1, "y": 4, "z": 0},
+            {"x": 2, "y": 0, "z": 0},
+        ],
+        "assessment": {
+            "decision": "threat",
+            "corridor_intersection": "intersects",
+        },
+    }
+    unknown = {
+        **threat,
+        "component_id": "rolling-unknown",
+        "assessment": {
+            "decision": "unknown",
+            "corridor_intersection": "unknown",
+        },
+    }
+
+    selected = _actionable_camera_metric_rows(
+        [threat, unknown],
+        added_cells=frozenset({(0, 0, 0), (1, 4, 0)}),
+        body_frame=body_frame,
+        profile=profile,
+    )
+
+    assert [row["component_id"] for row in selected] == ["rolling-actionable"]
+    assert selected[0]["cells"] == [{"x": 0, "y": 0, "z": 0}]