diff --git a/data_only_viz/icp_fusion.py b/data_only_viz/icp_fusion.py
index b1ff14f..f668aca 100644
--- a/data_only_viz/icp_fusion.py
+++ b/data_only_viz/icp_fusion.py
@@ -158,3 +158,58 @@ def partition_lidar_by_pid(
             continue
         out[pid] = pts[sel]
     return out
+
+
+PELVIS_VERT_INDEX = 5559  # SMPL-X canonical pelvis vertex
+
+
+@dataclass
+class FusionMetadata:
+    applied: set[int]
+    fitness: dict[int, float]
+    rmse_m: dict[int, float]
+    n_lidar_points_used: int
+
+
+class FusionWorker:
+    """Per-frame ICP fusion orchestrator (caller-driven, no internal thread)."""
+
+    def __init__(self, extrinsic, config: IcpConfig | None = None) -> None:
+        self._extrinsic = extrinsic
+        self._config = config or IcpConfig()
+
+    def set_extrinsic(self, extrinsic) -> None:
+        self._extrinsic = extrinsic
+
+    def run_once(self, state) -> FusionMetadata:
+        applied: set[int] = set()
+        fitness: dict[int, float] = {}
+        rmse: dict[int, float] = {}
+
+        lidar = getattr(state, "lidar_points", None)
+        if lidar is None or getattr(lidar, "size", 0) == 0 or not state.persons_smplx:
+            return FusionMetadata(applied, fitness, rmse, 0)
+
+        T = np.asarray(self._extrinsic.T_arkit_to_cam, dtype=np.float32)
+        homog = np.concatenate([lidar, np.ones((lidar.shape[0], 1), dtype=np.float32)], axis=1)
+        lidar_cam = (homog @ T.T)[:, :3]
+
+        pelvises = {
+            p.pid: p.vertices_3d[PELVIS_VERT_INDEX]
+            for p in state.persons_smplx
+            if p.vertices_3d is not None
+        }
+        parts = partition_lidar_by_pid(lidar_cam, pelvises, max_dist_m=1.0)
+
+        for person in state.persons_smplx:
+            pts = parts.get(person.pid)
+            if pts is None:
+                continue
+            result = register_mesh_to_lidar(person.vertices_3d, pts, self._config)
+            fitness[person.pid] = result.fitness
+            rmse[person.pid] = result.rmse_m
+            if result.accepted:
+                person.vertices_3d = result.vertices_registered
+                applied.add(person.pid)
+
+        return FusionMetadata(applied, fitness, rmse, lidar_cam.shape[0])
diff --git a/data_only_viz/state.py b/data_only_viz/state.py
index 21b99fc..af64d60 100644
--- a/data_only_viz/state.py
+++ b/data_only_viz/state.py
@@ -139,6 +139,14 @@ class State:
     persons_arkit_joints: dict = field(default_factory=dict)
     persons_arkit_last_t: dict = field(default_factory=dict)
 
+    # ---- LiDAR / ICP mesh fusion (Task 8 - 2026-05-14) ----
+    # Set by the LidarTCPReader poller; consumed by FusionWorker.run_once.
+    # The mesh-level fusion is complementary to the ARKit *joint* fusion
+    # above: joints are sparse + 60 Hz, LiDAR is dense + 5-10 Hz.
+    lidar_points: object = None          # np.ndarray (N, 3) float32 ARKit world; None if no frame
+    lidar_timestamp_ns: int = 0
+    icp_metadata: object = None          # FusionMetadata from icp_fusion or None
+
     # v1.3: centralised webcam source. WebcamSource owns the single
     # cv2.VideoCapture on the host and writes BGR frames here so all
     # consumers (MediaPipe Multi, Apple Vision, Multi-HMR worker,
diff --git a/data_only_viz/tests/test_icp_fusion.py b/data_only_viz/tests/test_icp_fusion.py
index 0babf8e..ed48a00 100644
--- a/data_only_viz/tests/test_icp_fusion.py
+++ b/data_only_viz/tests/test_icp_fusion.py
@@ -96,3 +96,49 @@ def test_partition_returns_empty_dict_when_no_pelvises() -> None:
 
     out = partition_lidar_by_pid(np.zeros((100, 3), dtype=np.float32), pelvises={}, max_dist_m=1.0)
     assert out == {}
+
+
+def test_fusion_worker_in_place_update(monkeypatch) -> None:
+    from data_only_viz.icp_fusion import FusionWorker, IcpConfig
+    from data_only_viz.lidar_calib import Extrinsic
+    from data_only_viz.state import SMPLXPerson, State
+
+    src = _synthetic_smplx_torso(seed=30)
+    verts = np.zeros((10475, 3), dtype=np.float32)
+    verts[: src.shape[0]] = src
+    verts[5559] = src.mean(axis=0)
+
+    person = SMPLXPerson(pid=0, vertices_3d=verts.copy())
+    state = State()
+    state.persons_smplx = [person]
+
+    lidar_pts = src + np.array([0.0, 0.04, 0.0], dtype=np.float32)
+    state.lidar_points = lidar_pts
+    state.lidar_timestamp_ns = 1
+
+    worker = FusionWorker(
+        extrinsic=Extrinsic.identity(),
+        config=IcpConfig(),
+    )
+    metadata = worker.run_once(state)
+
+    assert metadata.applied == {0}
+    delta = state.persons_smplx[0].vertices_3d[5559] - verts[5559]
+    assert 0.02 <= delta[1] <= 0.06
+
+
+def test_fusion_worker_skips_when_no_lidar() -> None:
+    from data_only_viz.icp_fusion import FusionWorker, IcpConfig
+    from data_only_viz.lidar_calib import Extrinsic
+    from data_only_viz.state import SMPLXPerson, State
+
+    verts = np.zeros((10475, 3), dtype=np.float32)
+    verts[5559] = [0.0, 1.0, 2.0]
+    state = State()
+    state.persons_smplx = [SMPLXPerson(pid=0, vertices_3d=verts.copy())]
+    state.lidar_points = None
+
+    worker = FusionWorker(extrinsic=Extrinsic.identity(), config=IcpConfig())
+    metadata = worker.run_once(state)
+    assert metadata.applied == set()
+    np.testing.assert_array_equal(state.persons_smplx[0].vertices_3d, verts)