diff --git a/data_only_viz/CLAUDE.md b/data_only_viz/CLAUDE.md
index a4434e0..f252139 100644
--- a/data_only_viz/CLAUDE.md
+++ b/data_only_viz/CLAUDE.md
@@ -35,6 +35,32 @@ Python **3.11+** requis. `pyproject.toml` est la source de vérité — ne jamai
 - Shaders Metal dans `shaders/` (`.metal`), recompilés au runtime ; topologie mesh (SMPL faces) en binaire dans `mesh_topology.py`.
 - OSC out : `osc_listener.py` / `pose_bridge.py` — destination `oscope-of` sur `:57123`.
 
+## action-head (classifier action debout/assise/danse)
+
+Tête de classification d'action streaming au-dessus des j3d SMPL-X (ou body3d MediaPipe en fallback). Implémentée 2026-05-13.
+
+| Fichier | Rôle |
+|---|---|
+| `action_head.py` | `ActionHeadModel` (GRU 1L + MLP, 37 811 params, <2 ms/step M5), `ActionHead.step(pid, j3d) → (label, probs, kin)`, `PerPersonBuffer`, `FeatureExtractor` (201-D : j3d + vel + accel + scalaires) |
+| `action_head_pub.py` | Publisher thread démarré dans `multi.py` `__init__`. Polle `state.persons_smplx` (préféré) ou `state.persons_body3d` (fallback) à 30 Hz, dédup par timestamp, extrait j3d22 via `SMPLX_JOINT_ANCHOR_VERTS` ou `MEDIAPIPE_TO_22`, émet OSC `/pose/action` + `/pose/kin` + `/pose/enter/leave` |
+| `training/{dataset,autolabel,augment,train_action_head,eval,review}.py` | Pipeline complet : jsonl IO + sliding windows + by-session split / règles auto-label + glue CLI / 4 augmentations / training MPS AdamW CE-weighted / confusion matrix + latence micro-bench / TUI textuel pour review manuel |
+| `scripts/capture_actions.py` | Webcam → MP4 + timestamps |
+| `scripts/extract_j3d_offline.py` | MP4 → jsonl j3d22 via `MultiHMRCoreMLBackend.infer()` directement (pas de refactor worker) |
+| `scripts/train_on_studio.sh` | rsync grosmac → bastion electron-server → studio M3 Ultra + uv sync `--extra multihmr` + train MPS + ckpt back |
+
+Pipeline complet de capture à live :
+```bash
+uv run python -m data_only_viz.scripts.capture_actions --session sess01 --duration 600
+uv run python -m data_only_viz.scripts.extract_j3d_offline --session sess01 --video ~/.cache/av-live-action/raw/sess01.mp4
+uv run python -m data_only_viz.training.autolabel --frames ~/.cache/av-live-action/raw/sess01.jsonl --out ~/.cache/av-live-action/dataset/auto.jsonl
+uv run python -m data_only_viz.training.review --in ~/.cache/av-live-action/dataset/auto.jsonl --out ~/.cache/av-live-action/dataset/dataset.jsonl
+./data_only_viz/scripts/train_on_studio.sh --epochs 50
+uv run python -m data_only_viz.training.eval --ckpt ~/.cache/av-live-action/checkpoints/action_head.pt --dataset ~/.cache/av-live-action/dataset/dataset.jsonl
+# Live : publisher déjà câblé dans multi.py, aucune action requise
+```
+
+Checkpoint par défaut : `~/.cache/av-live-action/checkpoints/action_head.pt`. Absent → random init (warmup retourne `debout`).
+
 ## Tests
 
 ```bash
@@ -43,6 +69,8 @@ uv run pytest tests/ -v
 
 Tests TDD-first pour `nlf_worker.py` ; valider avant chaque commit qui touche un worker.
 
+Suite action-head (8 fichiers, 39 tests) : `tests/test_action_head_*.py`, `tests/test_{dataset,autolabel,augment,training_smoke,pose_bridge_action}.py`. Tous doivent rester verts avant chaque commit qui touche `action_head*.py` ou `training/*.py`.
+
 ## Anti-patterns
 
 - Ne pas charger un modèle ML sans guard `try/except ImportError` — les optional-extras peuvent manquer.
diff --git a/data_only_viz/action_head.py b/data_only_viz/action_head.py
new file mode 100644
index 0000000..c455ede
--- /dev/null
+++ b/data_only_viz/action_head.py
@@ -0,0 +1,279 @@
+"""Action classifier head on top of Multi-HMR j3d.
+
+Streaming GRU-1-layer + MLP per-person, with a 16-frame ring buffer.
+Trained windowed (Studio M3 Ultra MPS), inferred streaming (M5 eager CPU).
+
+Output per step: (label_idx, probs (3,), kin (3,)) where kin is
+(speed_m_s, accel_m_s2, symmetry_in_minus1_plus1).
+"""
+from __future__ import annotations
+
+from collections import deque
+from pathlib import Path
+
+import numpy as np
+import torch
+from torch import nn
+
+HIDDEN_DIM: int = 48
+MLP_HIDDEN: int = 32
+WARMUP_FRAMES: int = 3
+NAN_SKIP_BUDGET: int = 5
+
+WINDOW_LEN: int = 16
+J3D_BODY: int = 22
+J3D_FINGERS_PER_HAND: int = 5
+J3D_FINGERS: int = 2 * J3D_FINGERS_PER_HAND  # 10
+J3D_JOINTS: int = J3D_BODY + J3D_FINGERS     # 32
+J3D_DIMS: int = 3
+NUM_CLASSES: int = 3
+LABELS: tuple[str, str, str] = ("debout", "assise", "danse")
+
+EXPR_DIM: int = 10
+EXTRA_SCALARS: int = 4   # hip_y, knee_angle, sym_score, mouth_open
+
+# NEW v3 : MediaPipe Hands keypoints block.
+HANDS_KP_PER_HAND: int = 21
+HANDS_KP_TOTAL: int = 2 * HANDS_KP_PER_HAND       # 42
+HANDS_KP_DIMS: int = 3
+HANDS_KP_FLAT: int = HANDS_KP_TOTAL * HANDS_KP_DIMS  # 126
+
+# Layout per step (v3) :
+#   [0      :  96]  j3d   (32, 3)
+#   [96     : 192]  vel   (32, 3)
+#   [192    : 288]  accel (32, 3)
+#   [288    : 414]  hands_kp (42, 3) zero-padded if absent
+#   [414    : 424]  expression (10,)
+#   [424    : 428]  scalars (hip_y, knee_angle, sym, mouth_open)
+FEATURE_DIM: int = J3D_JOINTS * J3D_DIMS * 3 + HANDS_KP_FLAT + EXPR_DIM + EXTRA_SCALARS  # 428
+
+# Body joint indices (unchanged from v1, indices 0..21).
+HIP_LEFT: int = 1
+HIP_RIGHT: int = 2
+KNEE_LEFT: int = 4
+KNEE_RIGHT: int = 5
+ANKLE_LEFT: int = 7
+ANKLE_RIGHT: int = 8
+SHOULDER_LEFT: int = 16
+SHOULDER_RIGHT: int = 17
+WRIST_LEFT: int = 20
+WRIST_RIGHT: int = 21
+
+# Fingertip indices (new, 22..31), order: L thumb..pinky, R thumb..pinky.
+FINGERTIP_LEFT_BASE: int = 22
+FINGERTIP_RIGHT_BASE: int = 27
+
+
+class FeatureExtractor:
+    """Stateless feature builder over a list of recent j3d frames.
+
+    Vector layout (FEATURE_DIM = 428, v3):
+      [0   :  96]  j3d current frame, flattened (32 joints x 3 dims)
+      [96  : 192]  velocity j3d[t] - j3d[t-1] (32 x 3)
+      [192 : 288]  acceleration vel[t] - vel[t-1] (32 x 3)
+      [288 : 414]  hands_kp (42, 3) MediaPipe Hands, zero-padded if absent
+      [414 : 424]  expression PCA coefficients (10,)
+      [424 : 428]  kinetics scalars (hip_y, knee_angle, symmetry_score, mouth_open)
+    """
+
+    @staticmethod
+    def from_buffer(frames: list[np.ndarray],
+                    expr: np.ndarray | None = None,
+                    mouth_open: float = 0.0,
+                    hands_kp: np.ndarray | None = None) -> np.ndarray:
+        if not frames:
+            return np.zeros(FEATURE_DIM, dtype=np.float32)
+        cur = frames[-1]
+        prev = frames[-2] if len(frames) >= 2 else cur
+        prev2 = frames[-3] if len(frames) >= 3 else prev
+        vel = (cur - prev).astype(np.float32, copy=False)
+        prev_vel = (prev - prev2).astype(np.float32, copy=False)
+        accel = (vel - prev_vel).astype(np.float32, copy=False)
+        hip_y = float((cur[HIP_LEFT, 1] + cur[HIP_RIGHT, 1]) * 0.5)
+        knee_angle = FeatureExtractor._mean_knee_angle(cur)
+        sym = FeatureExtractor._symmetry_score(vel)
+        # hands block (42, 3) -> 126
+        hands_flat = np.zeros(HANDS_KP_FLAT, dtype=np.float32)
+        if hands_kp is not None:
+            hk = np.asarray(hands_kp, dtype=np.float32)
+            if hk.shape == (HANDS_KP_TOTAL, HANDS_KP_DIMS):
+                hands_flat = hk.reshape(-1).astype(np.float32, copy=False)
+        # expression
+        if expr is None:
+            expr_vec = np.zeros(EXPR_DIM, dtype=np.float32)
+        else:
+            expr_vec = np.zeros(EXPR_DIM, dtype=np.float32)
+            n = min(EXPR_DIM, len(expr))
+            expr_vec[:n] = expr[:n]
+        return np.concatenate([
+            cur.reshape(-1),
+            vel.reshape(-1),
+            accel.reshape(-1),
+            hands_flat,
+            expr_vec,
+            np.array([hip_y, knee_angle, sym, float(mouth_open)], dtype=np.float32),
+        ]).astype(np.float32, copy=False)
+
+    @staticmethod
+    def kinetics(frames: list[np.ndarray]) -> np.ndarray:
+        """Return (speed, accel_mag, symmetry) averaged over the buffer."""
+        if len(frames) < 2:
+            return np.zeros(3, dtype=np.float32)
+        arr = np.stack(frames).astype(np.float32, copy=False)
+        diffs = arr[1:] - arr[:-1]
+        speeds = np.linalg.norm(diffs, axis=-1).mean(axis=-1)
+        speed = float(speeds.mean())
+        if len(frames) >= 3:
+            ddiffs = diffs[1:] - diffs[:-1]
+            accel = float(np.linalg.norm(ddiffs, axis=-1).mean())
+        else:
+            accel = 0.0
+        sym = FeatureExtractor._symmetry_score(diffs[-1])
+        return np.array([speed, accel, sym], dtype=np.float32)
+
+    @staticmethod
+    def _mean_knee_angle(j3d: np.ndarray) -> float:
+        """Angle (rad) at left+right knees, averaged."""
+        def _angle(hip: int, knee: int, ankle: int) -> float:
+            v1 = j3d[hip] - j3d[knee]
+            v2 = j3d[ankle] - j3d[knee]
+            n1 = np.linalg.norm(v1) + 1e-6
+            n2 = np.linalg.norm(v2) + 1e-6
+            cos = float(np.dot(v1, v2) / (n1 * n2))
+            return float(np.arccos(np.clip(cos, -1.0, 1.0)))
+        return 0.5 * (_angle(HIP_LEFT, KNEE_LEFT, ANKLE_LEFT)
+                      + _angle(HIP_RIGHT, KNEE_RIGHT, ANKLE_RIGHT))
+
+    @staticmethod
+    def _symmetry_score(vel: np.ndarray) -> float:
+        """Cosine sim between left-arm and mirrored right-arm velocity."""
+        left = vel[WRIST_LEFT].copy()
+        right = vel[WRIST_RIGHT].copy()
+        right_mirror = right.copy()
+        right_mirror[0] = -right_mirror[0]
+        n1 = np.linalg.norm(left) + 1e-6
+        n2 = np.linalg.norm(right_mirror) + 1e-6
+        return float(np.dot(left, right_mirror) / (n1 * n2))
+
+
+class PerPersonBuffer:
+    """Per-pid ring buffer of j3d frames (deque maxlen=WINDOW_LEN)."""
+
+    __slots__ = ("_buffers",)
+
+    def __init__(self) -> None:
+        self._buffers: dict[int, deque[np.ndarray]] = {}
+
+    def append(self, pid: int, j3d: np.ndarray) -> None:
+        if j3d.shape != (J3D_JOINTS, J3D_DIMS):
+            raise ValueError(
+                f"j3d must be ({J3D_JOINTS}, {J3D_DIMS}), got {j3d.shape}"
+            )
+        dq = self._buffers.get(pid)
+        if dq is None:
+            dq = deque(maxlen=WINDOW_LEN)
+            self._buffers[pid] = dq
+        dq.append(j3d.astype(np.float32, copy=False))
+
+    def frames_for(self, pid: int) -> list[np.ndarray]:
+        dq = self._buffers.get(pid)
+        return list(dq) if dq is not None else []
+
+    def forget(self, pid: int) -> None:
+        self._buffers.pop(pid, None)
+
+    def __len__(self) -> int:
+        return len(self._buffers)
+
+    def pids(self) -> list[int]:
+        return list(self._buffers.keys())
+
+
+class ActionHeadModel(nn.Module):
+    """1-layer GRU + small MLP head.
+
+    Input  : (B, FEATURE_DIM) -- single step
+    Hidden : (1, B, HIDDEN_DIM)
+    Output : (B, NUM_CLASSES) logits, new hidden
+    """
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.gru = nn.GRU(input_size=FEATURE_DIM,
+                          hidden_size=HIDDEN_DIM,
+                          num_layers=1,
+                          batch_first=True)
+        self.mlp = nn.Sequential(
+            nn.Linear(HIDDEN_DIM, MLP_HIDDEN),
+            nn.ReLU(inplace=True),
+            nn.Linear(MLP_HIDDEN, NUM_CLASSES),
+        )
+
+    def init_hidden(self, batch: int = 1, device: str = "cpu") -> torch.Tensor:
+        return torch.zeros(1, batch, HIDDEN_DIM, device=device)
+
+    def forward(self, x: torch.Tensor,
+                h: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        out, h_new = self.gru(x.unsqueeze(1), h)
+        logits = self.mlp(out.squeeze(1))
+        return logits, h_new
+
+
+class ActionHead:
+    """Streaming action classifier per person.
+
+    Use:
+        head = ActionHead(ckpt_path=...)
+        label, probs, kin = head.step(pid, j3d)
+        head.forget(pid)
+    """
+
+    def __init__(self,
+                 ckpt_path: Path | None = None,
+                 device: str = "cpu") -> None:
+        self._device = device
+        self._model = ActionHeadModel().to(device).eval()
+        if ckpt_path is not None:
+            payload = torch.load(ckpt_path, map_location=device,
+                                 weights_only=True)
+            state = payload.get("model_state_dict", payload)
+            self._model.load_state_dict(state)
+        self._buffers = PerPersonBuffer()
+        self._hidden: dict[int, torch.Tensor] = {}
+        self._nan_streak: dict[int, int] = {}
+
+    def step(self, pid: int, j3d: np.ndarray,
+             expr: np.ndarray | None = None,
+             mouth_open: float = 0.0,
+             hands_kp: np.ndarray | None = None) -> tuple[str, np.ndarray, np.ndarray]:
+        if np.isnan(j3d).any():
+            streak = self._nan_streak.get(pid, 0) + 1
+            self._nan_streak[pid] = streak
+            if streak > NAN_SKIP_BUDGET:
+                self.forget(pid)
+            probs = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+            return LABELS[0], probs, np.zeros(3, dtype=np.float32)
+        self._nan_streak[pid] = 0
+        self._buffers.append(pid, j3d)
+        frames = self._buffers.frames_for(pid)
+        if len(frames) < WARMUP_FRAMES:
+            probs = np.array([1.0, 0.0, 0.0], dtype=np.float32)
+            return LABELS[0], probs, np.zeros(3, dtype=np.float32)
+        feat = FeatureExtractor.from_buffer(frames, expr=expr, mouth_open=mouth_open,
+                                             hands_kp=hands_kp)
+        kin = FeatureExtractor.kinetics(frames)
+        h = self._hidden.get(pid)
+        if h is None:
+            h = self._model.init_hidden(batch=1, device=self._device)
+        x = torch.from_numpy(feat).unsqueeze(0).to(self._device)
+        with torch.no_grad():
+            logits, h_new = self._model(x, h)
+            probs_t = torch.softmax(logits, dim=-1).squeeze(0)
+        self._hidden[pid] = h_new
+        probs = probs_t.cpu().numpy().astype(np.float32, copy=False)
+        return LABELS[int(np.argmax(probs))], probs, kin
+
+    def forget(self, pid: int) -> None:
+        self._buffers.forget(pid)
+        self._hidden.pop(pid, None)
+        self._nan_streak.pop(pid, None)
diff --git a/data_only_viz/action_head_pub.py b/data_only_viz/action_head_pub.py
new file mode 100644
index 0000000..51bd033
--- /dev/null
+++ b/data_only_viz/action_head_pub.py
@@ -0,0 +1,313 @@
+"""Action-head publisher : reads state.persons_smplx / persons_body3d,
+runs ActionHead per pid, emits /pose/action and /pose/kin via pose_bridge.
+
+Stand-alone thread to avoid touching multi_hmr_worker.py while it
+iterates. Polls state at ~30 Hz, deduplicates by smplx_last_t.
+"""
+from __future__ import annotations
+
+import logging
+import threading
+import time
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+
+from data_only_viz.action_head import (
+    ActionHead,
+    EXPR_DIM,
+    HANDS_KP_DIMS,
+    HANDS_KP_PER_HAND,
+    HANDS_KP_TOTAL,
+    J3D_FINGERS,
+    J3D_FINGERS_PER_HAND,
+    LABELS,
+)
+
+LOG = logging.getLogger("action_head_pub")
+
+DEFAULT_CKPT = (
+    Path.home() / ".cache" / "av-live-action" / "checkpoints" / "action_head.pt"
+)
+
+# Canonical SMPL-X fingertip vertex IDs from smplx.vertex_ids.SMPLX_VERTEX_IDS.
+# Order : L thumb, L index, L middle, L ring, L pinky,
+#         R thumb, R index, R middle, R ring, R pinky.
+SMPLX_FINGERTIP_VERTS: tuple[int, ...] = (
+    5361, 4933, 5058, 5169, 5286,   # L : lthumb, lindex, lmiddle, lring, lpinky
+    8079, 7669, 7794, 7905, 8022,   # R : rthumb, rindex, rmiddle, rring, rpinky
+)
+
+# 32 vertex indices on the 10475-vertex SMPL-X mesh:
+# 22 body (UNCHANGED from v1) + 10 fingertips.
+# NOTE: approximate vertex anchors -- real SMPL-X joints come from
+# J_regressor @ v3d, but loading the regressor here is avoided for
+# live OSC performance. Action-head training must use the same anchors.
+SMPLX_JOINT_ANCHOR_VERTS: tuple[int, ...] = (
+    # 22 body (UNCHANGED indices, same vertex IDs as before)
+    8204, 3992, 6677, 3500, 3469, 6394, 3279, 3327, 6736, 3074,
+    8846, 8889, 8848, 1300, 4660, 8964, 3013, 6470, 1602, 5083,
+    2114, 5559,
+    # 10 fingertips
+    *SMPLX_FINGERTIP_VERTS,
+)
+assert len(SMPLX_JOINT_ANCHOR_VERTS) == 32
+
+# Mouth-open: distance between two lip vertices on SMPL-X mesh.
+# vert 8970 (upper outer lip), 8855 (lower outer lip) -- approximate.
+SMPLX_UPPER_LIP_VERT: int = 8970
+SMPLX_LOWER_LIP_VERT: int = 8855
+
+# MediaPipe FaceMesh inner-mouth landmark indices.
+# 13 = upper inner mid, 14 = lower inner mid.
+MEDIAPIPE_LIP_UPPER_INNER: int = 13
+MEDIAPIPE_LIP_LOWER_INNER: int = 14
+
+# MediaPipe HAND fingertip indices (21-kp hand model).
+MEDIAPIPE_HAND_FINGERTIPS: tuple[int, ...] = (4, 8, 12, 16, 20)
+
+# MediaPipe 33-landmark indices mapped into the 22-joint slot order.
+# NOTE: approximate mapping -- spine joints reuse hip/shoulder anchors.
+# https://developers.google.com/mediapipe/solutions/vision/pose_landmarker
+MEDIAPIPE_TO_22: tuple[int, ...] = (
+    24, 23, 24, 23, 25, 26, 11, 27, 28, 11,
+    31, 32, 0, 11, 12, 0, 11, 12, 13, 14, 15, 16,
+)
+
+
+class ActionHeadPublisher(threading.Thread):
+    """Thread that polls state, runs ActionHead per pid, emits OSC."""
+
+    def __init__(self, state: Any, bridge: Any,
+                 ckpt_path: Path | None = DEFAULT_CKPT,
+                 period_s: float = 1.0 / 30.0) -> None:
+        super().__init__(daemon=True, name="action-head-pub")
+        self.state = state
+        self.bridge = bridge
+        self.period = period_s
+        try:
+            ckpt = ckpt_path if (ckpt_path and ckpt_path.exists()) else None
+            self.head = ActionHead(ckpt_path=ckpt, device="cpu")
+            LOG.info("action_head loaded ckpt=%s",
+                     ckpt if ckpt else "<random init>")
+        except Exception as e:
+            LOG.warning("action_head init failed: %s", e)
+            self.head = None
+        self._stop = threading.Event()
+        self._last_smplx_t = 0.0
+        self._last_body_t = 0.0
+        self._last_pids: set[int] = set()
+
+    def stop(self) -> None:
+        self._stop.set()
+
+    def run(self) -> None:
+        if self.head is None:
+            LOG.warning("publisher exiting: no action_head")
+            return
+        LOG.info("publisher started")
+        while not self._stop.is_set():
+            t0 = time.perf_counter()
+            try:
+                self._tick(t0)
+            except Exception:
+                LOG.exception("publisher tick failed")
+            dt = time.perf_counter() - t0
+            if dt < self.period:
+                time.sleep(self.period - dt)
+        LOG.info("publisher stopped")
+
+    def _tick(self, t_now: float) -> None:
+        persons32, source_t, source_tag, is_new = self._read_sources()
+        if not is_new:
+            return
+        if "smplx" in source_tag:
+            self._last_smplx_t = source_t
+        else:
+            self._last_body_t = source_t
+        current_pids: set[int] = set()
+        if persons32:
+            for pid, j3d, expr_np, mouth, hands_kp42 in persons32:
+                current_pids.add(pid)
+                label, probs, kin = self.head.step(pid, j3d, expr=expr_np,
+                                                    mouth_open=mouth,
+                                                    hands_kp=hands_kp42)
+                idx = LABELS.index(label)
+                self.bridge.send_action(pid, idx, probs, t_now, force=True)
+                self.bridge.send_kin(pid, kin, t_now, force=True)
+                if pid not in self._last_pids:
+                    self.bridge.send_enter(pid=pid)
+        for gone in self._last_pids - current_pids:
+            self.head.forget(gone)
+            self.bridge.send_leave(pid=gone)
+        self._last_pids = current_pids
+
+    def _read_sources(self) -> tuple[
+        list[tuple[int, np.ndarray, np.ndarray, float, np.ndarray]] | None,
+        float, str, bool,
+    ]:
+        """Return (persons32, source_t, source_tag, is_new).
+
+        Each person entry is (pid, j3d32, expr10, mouth_open, hands_kp42x3).
+        is_new is True when the timestamp advanced (even if person list
+        is empty), so _tick can still run the purge loop.
+        """
+        with self.state.lock():
+            persons_smplx = getattr(self.state, "persons_smplx", None)
+            t_smplx = getattr(self.state, "smplx_last_t", 0.0)
+            persons_b3d = getattr(self.state, "persons_body3d", None)
+            ids_b3d = getattr(self.state, "persons_body_ids", None)
+            persons_face = getattr(self.state, "persons_face", None)
+            ids_face = getattr(self.state, "persons_face_ids", None)
+            persons_hands = getattr(self.state, "persons_hands", None)
+            ids_hands = getattr(self.state, "persons_hands_ids", None)
+            t_body = getattr(self.state, "pose_last_t", 0.0)
+
+        # Build pid -> hands_kp(42, 3) map from MediaPipe persons_hands.
+        hands_by_pid: dict[int, np.ndarray] = self._build_hands_map(
+            persons_hands or [], ids_hands or [],
+        )
+        # Build pid -> mouth_open scalar from MediaPipe persons_face lips.
+        face_mouth_by_pid: dict[int, float] = self._build_face_mouth_map(
+            persons_face or [], ids_face or [],
+        )
+
+        # SMPL-X path (preferred)
+        if t_smplx > self._last_smplx_t:
+            out: list[tuple[int, np.ndarray, np.ndarray, float, np.ndarray]] = []
+            for i, p in enumerate(persons_smplx or []):
+                pid = int(p.get("pid", i))
+                v3d = p.get("v3d")
+                if v3d is None:
+                    continue
+                # CoreMLArray wraps a numpy array but has no __array__
+                # protocol; unwrap via .numpy() before np.asarray.
+                if hasattr(v3d, "numpy") and not isinstance(v3d, np.ndarray):
+                    v3d = v3d.numpy()
+                v3d_np = np.asarray(v3d, dtype=np.float32)
+                if v3d_np.shape[0] < max(SMPLX_JOINT_ANCHOR_VERTS) + 1:
+                    continue
+                j3d32 = v3d_np[list(SMPLX_JOINT_ANCHOR_VERTS)].astype(np.float32)
+                # expression
+                expr = p.get("expression")
+                if expr is not None:
+                    if hasattr(expr, "numpy") and not isinstance(expr, np.ndarray):
+                        expr = expr.numpy()
+                    expr_np = np.asarray(expr, dtype=np.float32).flatten()
+                else:
+                    expr_np = np.zeros(EXPR_DIM, dtype=np.float32)
+                # mouth_open: prefer MediaPipe face lips, fallback SMPL-X v3d.
+                if pid in face_mouth_by_pid:
+                    mouth = face_mouth_by_pid[pid]
+                elif v3d_np.shape[0] > max(SMPLX_UPPER_LIP_VERT, SMPLX_LOWER_LIP_VERT):
+                    mouth = float(np.linalg.norm(
+                        v3d_np[SMPLX_UPPER_LIP_VERT] - v3d_np[SMPLX_LOWER_LIP_VERT]
+                    ))
+                else:
+                    mouth = 0.0
+                hands_kp42 = hands_by_pid.get(
+                    pid, np.zeros((HANDS_KP_TOTAL, HANDS_KP_DIMS), dtype=np.float32)
+                )
+                out.append((pid, j3d32, expr_np, mouth, hands_kp42))
+            return out or None, t_smplx, "smplx", True
+
+        # MediaPipe body3d fallback
+        if t_body > self._last_body_t:
+            ids = ids_b3d or list(range(len(persons_b3d or [])))
+            out = []
+            for i, body in enumerate(persons_b3d or []):
+                pid = int(ids[i]) if i < len(ids) else i
+                arr = self._kp_list_to_array(body)
+                if arr is None or arr.shape[0] < 33:
+                    continue
+                body22 = arr[list(MEDIAPIPE_TO_22)].astype(np.float32)
+                # fingertips from persons_hands if available
+                tips = np.zeros((J3D_FINGERS, 3), dtype=np.float32)
+                hands_kp42 = hands_by_pid.get(
+                    pid, np.zeros((HANDS_KP_TOTAL, HANDS_KP_DIMS), dtype=np.float32)
+                )
+                # extract fingertips from hands_kp42 (idx 4,8,12,16,20 each side)
+                for side_idx in (0, 1):
+                    base = side_idx * HANDS_KP_PER_HAND
+                    for k, mp_tip in enumerate(MEDIAPIPE_HAND_FINGERTIPS):
+                        if base + mp_tip < hands_kp42.shape[0]:
+                            tips[side_idx * J3D_FINGERS_PER_HAND + k] = \
+                                hands_kp42[base + mp_tip]
+                j3d32 = np.concatenate([body22, tips], axis=0)
+                mouth = face_mouth_by_pid.get(pid, 0.0)
+                expr_np = np.zeros(EXPR_DIM, dtype=np.float32)
+                out.append((pid, j3d32, expr_np, mouth, hands_kp42))
+            return out or None, t_body, "body3d", True
+        return None, 0.0, "", False
+
+    def _build_hands_map(self, persons_hands: list,
+                          ids_hands: list) -> dict[int, np.ndarray]:
+        """Combine left+right hand kp arrays per pid into a single (42, 3) array.
+
+        persons_hands is a flat list ; ids_hands maps each hand-list entry to a
+        pid (and odd/even index indicates which side). When the user's pipeline
+        keeps a different convention, this helper makes the best effort and
+        pads zeros for missing sides.
+        """
+        out: dict[int, np.ndarray] = {}
+        for hi, hkp in enumerate(persons_hands):
+            if hkp is None:
+                continue
+            pid_raw = ids_hands[hi] if hi < len(ids_hands) else hi
+            try:
+                pid = int(pid_raw)
+            except (TypeError, ValueError):
+                pid = hi
+            side = hi % 2  # 0 = L, 1 = R
+            arr = self._kp_list_to_array(hkp)
+            if arr is None or arr.shape[0] < HANDS_KP_PER_HAND:
+                continue
+            slot = out.setdefault(
+                pid, np.zeros((HANDS_KP_TOTAL, HANDS_KP_DIMS), dtype=np.float32)
+            )
+            base = side * HANDS_KP_PER_HAND
+            slot[base:base + HANDS_KP_PER_HAND] = arr[:HANDS_KP_PER_HAND]
+        return out
+
+    def _build_face_mouth_map(self, persons_face: list,
+                               ids_face: list) -> dict[int, float]:
+        """Compute mouth_open = norm(upper_inner_lip - lower_inner_lip) per pid."""
+        out: dict[int, float] = {}
+        for fi, fkp in enumerate(persons_face):
+            if fkp is None:
+                continue
+            arr = self._kp_list_to_array(fkp)
+            if arr is None or arr.shape[0] <= MEDIAPIPE_LIP_LOWER_INNER:
+                continue
+            upper = arr[MEDIAPIPE_LIP_UPPER_INNER]
+            lower = arr[MEDIAPIPE_LIP_LOWER_INNER]
+            mouth = float(np.linalg.norm(upper - lower))
+            try:
+                pid = int(ids_face[fi]) if fi < len(ids_face) else fi
+            except (TypeError, ValueError):
+                pid = fi
+            out[pid] = mouth
+        return out
+
+    @staticmethod
+    def _kp_list_to_array(body: Any) -> np.ndarray | None:
+        """Best-effort conversion of a body keypoint list to (N, 3) array."""
+        if body is None:
+            return None
+        if isinstance(body, np.ndarray):
+            return body
+        try:
+            return np.asarray(
+                [
+                    (
+                        getattr(kp, "x", kp[0]),
+                        getattr(kp, "y", kp[1]),
+                        getattr(kp, "z", kp[2] if len(kp) > 2 else 0.0),
+                    )
+                    for kp in body
+                ],
+                dtype=np.float32,
+            )
+        except (TypeError, IndexError, AttributeError):
+            return None
diff --git a/data_only_viz/apple_vision_pose.py b/data_only_viz/apple_vision_pose.py
index f94041f..2ccb3f5 100644
--- a/data_only_viz/apple_vision_pose.py
+++ b/data_only_viz/apple_vision_pose.py
@@ -552,32 +552,13 @@ class AppleVisionPoseWorker:
             if not hasattr(self, "_logged_face_ok_" + region_name):
                 LOG.info("face: region %s count=%d", region_name, count)
                 setattr(self, "_logged_face_ok_" + region_name, True)
-            # API stable : pointAtIndex_(k) retourne un CGPoint struct.
-            n = min(count, end - start)
-            n_written = 0
-            for k in range(n):
-                try:
-                    pt = region.pointAtIndex_(k)
-                    # CGPoint en pyobjc : tuple (x, y) ou struct
-                    try:
-                        nx_bb = float(pt.x); ny_bb = float(pt.y)
-                    except (AttributeError, TypeError):
-                        nx_bb = float(pt[0]); ny_bb = float(pt[1])
-                    fx = bx + nx_bb * bw
-                    fy_bl = by + ny_bb * bh
-                    kps[start + k] = PoseKp(
-                        x=fx, y=1.0 - fy_bl, z=0.0, c=1.0)
-                    n_written += 1
-                except Exception as e:
-                    if not hasattr(self, "_logged_face_pt_err"):
-                        LOG.info("face: pt %s[%d] err: %s (pt=%r)",
-                                 region_name, k, e, type(pt).__name__
-                                 if 'pt' in dir() else "??")
-                        self._logged_face_pt_err = True
-                    continue
-            if n_written > 0 and not hasattr(self, "_logged_face_write_" + region_name):
-                LOG.info("face: %s wrote %d points", region_name, n_written)
-                setattr(self, "_logged_face_write_" + region_name, True)
+            # pyobjc 11 ne sait pas que pointAtIndex_ prend 1 arg, et
+            # pointsInImageOfSize_ retourne un PyObjCPointer C-array sans
+            # API d'acces simple. Face parsing depuis Apple Vision est
+            # actuellement bloque ; on garde MediaPipe (CPU XNNPACK) pour
+            # face/hand fin tandis que Vision sert body 2D sur ANE.
+            # Skip pour eviter le spam ObjCPointerWarning a 30 fps.
+            return
 
         # faceContour
         fill("faceContour", *FACE_OFFSETS["contour"])
@@ -590,13 +571,19 @@ class AppleVisionPoseWorker:
         fill("nose",        *FACE_OFFSETS["nose"])
         fill("medianLine",  *FACE_OFFSETS["median"])
 
-        # Pupilles : VNFaceLandmarkRegion2D simple (1 point chacune).
+        # Pupilles : single-point regions ; meme workaround pyobjc.
         for region_name, idx in (("leftPupil", 81), ("rightPupil", 82)):
             try:
                 region = getattr(landmarks, region_name)()
                 if region is None or region.pointCount() < 1:
                     continue
-                pt = region.pointAtIndex_(0)
+                try:
+                    pts = region.pointsInImageOfSize_((1.0, 1.0))
+                except Exception:
+                    pts = region.normalizedPoints()
+                if not pts:
+                    continue
+                pt = pts[0]
                 try:
                     px, py = float(pt.x), float(pt.y)
                 except (AttributeError, TypeError):
diff --git a/data_only_viz/main.py b/data_only_viz/main.py
index a81423b..0be7474 100644
--- a/data_only_viz/main.py
+++ b/data_only_viz/main.py
@@ -16,6 +16,7 @@ from __future__ import annotations
 
 import argparse
 import logging
+import os
 import signal
 import sys
 
@@ -265,9 +266,55 @@ class AppDelegate(NSObject):
                             self._opts, "motion_gate", 5.0),
                         camera_index=getattr(self._opts, "camera_index", -1))
                     self._pose_worker.start()
-                    self._smplx_tcp = SMPLXTCPSender(self._state)
+                    # MESH_RIG=0 disables the 30 fps rigid translation
+                    # rigger from mesh_rigger.py (used to debug deformation
+                    # issues introduced by the hybrid rigging path).
+                    self._smplx_tcp = SMPLXTCPSender(
+                        self._state,
+                        enable_rigging=os.environ.get("MESH_RIG", "1") != "0",
+                    )
                     self._smplx_tcp.start()
                     LOG.info("worker: Multi-HMR + SMPL-X (mesh dense)")
+                    # Secondary body-pose worker in parallel: AVLiveBody
+                    # gets body keypoints on UDP :57126 alongside the mesh
+                    # on TCP :57130. Default: Apple Vision (ANE-accel,
+                    # body only 19 joints). Set AV_LIVE_PARALLEL_POSE=
+                    # mediapipe to swap to MediaPipe Holistic (CPU
+                    # XNNPACK but provides face + hand + 3D world).
+                    # Defaut: lance BOTH Apple Vision (body 19 joints sur
+                    # ANE, ~30 fps) ET MediaPipe Multi (face 468 + hands 21
+                    # + pose 3D world sur CPU XNNPACK). Set
+                    # AV_LIVE_PARALLEL_POSE=apple_vision pour ne garder que
+                    # le path ANE (face/hand fin disparait), ou =mediapipe
+                    # pour ne garder que CPU.
+                    parallel = _os.environ.get(
+                        "AV_LIVE_PARALLEL_POSE", "both")
+                    if parallel in ("apple_vision", "both"):
+                        try:
+                            from .apple_vision_pose import AppleVisionPoseWorker
+                            if AppleVisionPoseWorker.is_available():
+                                self._av_worker = AppleVisionPoseWorker(
+                                    self._state, target_fps=30.0,
+                                    num_persons=4)
+                                self._av_worker.start()
+                                LOG.info("worker: + Apple Vision body pose "
+                                         "(ANE) in parallel")
+                            else:
+                                raise RuntimeError("apple_vision unavailable")
+                        except Exception as e:  # noqa: BLE001
+                            LOG.warning("Apple Vision parallel start failed "
+                                        "(%s)", e)
+                    if parallel in ("mediapipe", "both"):
+                        try:
+                            from .multi import MultiWorker
+                            self._mp_worker = MultiWorker(
+                                self._state, num_persons=4)
+                            self._mp_worker.start()
+                            LOG.info("worker: + MediaPipe Multi (3D pose "
+                                     "+ face + hand) in parallel")
+                        except Exception as e:  # noqa: BLE001
+                            LOG.warning("MediaPipe parallel start failed "
+                                        "(%s)", e)
                     return
                 LOG.info("Multi-HMR indisponible (checkpoints manquants) "
                          "— voir scripts/setup_multihmr.sh")
diff --git a/data_only_viz/mesh_rigger.py b/data_only_viz/mesh_rigger.py
new file mode 100644
index 0000000..3eda2fc
--- /dev/null
+++ b/data_only_viz/mesh_rigger.py
@@ -0,0 +1,215 @@
+"""Mesh rigging hybride keyframe (Multi-HMR) + delta Apple Vision.
+
+Multi-HMR produit un mesh SMPL-X dense (10475 verts) tous les ~300 ms
+sur M5 (PyTorch MPS ~3.5 fps). Entre deux keyframes, Apple Vision sur
+ANE produit 30 fps de body keypoints 2D. On exploite le pelvis 2D de
+Vision pour translater rigidement le mesh keyframe et donner une
+perception fluide a 30 fps cote launcher RealityKit.
+
+Limitations connues (premiere iteration) :
+  - Translation rigide uniquement (pas de rotation, pas de LBS articule)
+  - Pelvis 2D delta projete en X/Y a profondeur constante (z keyframe)
+  - Pas de matching d'identite Vision <-> Multi-HMR : on prend la
+    personne Vision la plus proche du pelvis projete keyframe
+"""
+from __future__ import annotations
+
+import math
+import threading
+import time
+from dataclasses import dataclass, field
+
+import numpy as np
+
+from .state import PoseKp, SMPLXPerson, State
+
+
+# Indices MediaPipe POSE_LANDMARKS pour les hanches (pelvis 2D = midpoint).
+_LEFT_HIP = 23
+_RIGHT_HIP = 24
+
+# Focale par defaut Multi-HMR (camera intrinsics typiques utilisees
+# dans multi_hmr_worker : focal = IMG_SIZE).
+_IMG_SIZE = 672
+_FOCAL = float(_IMG_SIZE)
+
+
+@dataclass
+class _Keyframe:
+    """Snapshot d'un mesh Multi-HMR + reference Vision au moment T."""
+    pid: int
+    t: float
+    # Mesh world coords (10475, 3) float32 incluant la translation
+    vertices_3d: np.ndarray
+    translation: np.ndarray            # (3,) world pelvis
+    vision_pelvis_2d: tuple[float, float] | None  # (cx, cy) normalises 0..1
+
+
+def _pelvis_2d_from_body(body: list[PoseKp]) -> tuple[float, float] | None:
+    """Midpoint des deux hanches MediaPipe si confidence > 0."""
+    if not body or len(body) <= _RIGHT_HIP:
+        return None
+    lh, rh = body[_LEFT_HIP], body[_RIGHT_HIP]
+    if lh.c <= 0.1 or rh.c <= 0.1:
+        return None
+    return (0.5 * (lh.x + rh.x), 0.5 * (lh.y + rh.y))
+
+
+def _vision_pid_match(
+    keyframe_pelvis_2d: tuple[float, float] | None,
+    vision_bodies: list[list[PoseKp]],
+    vision_ids: list[int],
+) -> int | None:
+    """Retourne le pid Vision dont le pelvis 2D est le plus proche du
+    keyframe pelvis projete. None si rien."""
+    if keyframe_pelvis_2d is None or not vision_bodies:
+        return None
+    kx, ky = keyframe_pelvis_2d
+    best_pid: int | None = None
+    best_d2 = float("inf")
+    for body, vpid in zip(vision_bodies, vision_ids):
+        p = _pelvis_2d_from_body(body)
+        if p is None:
+            continue
+        d2 = (p[0] - kx) ** 2 + (p[1] - ky) ** 2
+        if d2 < best_d2:
+            best_d2 = d2
+            best_pid = int(vpid)
+    return best_pid
+
+
+class MeshRigger:
+    """Rig le mesh SMPL-X keyframe via le delta pelvis Vision.
+
+    Usage :
+        rigger = MeshRigger(state)
+        rigged_persons = rigger.apply(state.persons_smplx,
+                                      state.persons_body,
+                                      t_now)
+    Thread-safe : ne mute pas le state, retourne une nouvelle liste.
+    """
+
+    def __init__(self, state: State, hold_window_s: float = 1.5) -> None:
+        self.state = state
+        self.hold_window_s = hold_window_s
+        self._lock = threading.Lock()
+        # pid Multi-HMR -> keyframe
+        self._keyframes: dict[int, _Keyframe] = {}
+        # pid Multi-HMR -> pid Vision matched (sticky across frames)
+        self._vision_pid_map: dict[int, int] = {}
+
+    def apply(
+        self,
+        persons_smplx: list[SMPLXPerson],
+        persons_body: list[list[PoseKp]],
+        persons_body_ids: list[int],
+        t_now: float,
+    ) -> list[SMPLXPerson]:
+        """Retourne une liste SMPLXPerson translatee par delta Vision."""
+        # 1) Detect new keyframes (timestamp tracked via state.smplx_last_t)
+        with self._lock:
+            current_pids = {p.pid for p in persons_smplx}
+            # Drop stale keyframes (person disparue)
+            for old_pid in list(self._keyframes):
+                if old_pid not in current_pids:
+                    self._keyframes.pop(old_pid, None)
+                    self._vision_pid_map.pop(old_pid, None)
+
+            out: list[SMPLXPerson] = []
+            for person in persons_smplx:
+                kf = self._keyframes.get(person.pid)
+                # Detect keyframe refresh : translation differs from kf
+                is_new_kf = (kf is None or not np.allclose(
+                    kf.translation, person.translation, atol=1e-4))
+                if is_new_kf:
+                    # Trouver le pid Vision le plus proche pour ce mesh.
+                    # On projette le pelvis world en 2D image-normalized :
+                    # x_img = (X / Z) * focal / IMG_SIZE + 0.5
+                    pelvis_2d = self._project_pelvis(person.translation)
+                    matched = _vision_pid_match(
+                        pelvis_2d, persons_body, persons_body_ids)
+                    if matched is None:
+                        matched = self._vision_pid_map.get(person.pid)
+                    if matched is not None:
+                        self._vision_pid_map[person.pid] = matched
+                    # Capture du pelvis 2D Vision au moment du keyframe
+                    vp = None
+                    if matched is not None:
+                        try:
+                            i = persons_body_ids.index(matched)
+                            vp = _pelvis_2d_from_body(persons_body[i])
+                        except (ValueError, IndexError):
+                            vp = None
+                    self._keyframes[person.pid] = _Keyframe(
+                        pid=person.pid,
+                        t=t_now,
+                        vertices_3d=person.vertices_3d.copy(),
+                        translation=person.translation.copy(),
+                        vision_pelvis_2d=vp,
+                    )
+                    out.append(person)
+                    continue
+
+                # Entre keyframes : applique delta translation depuis
+                # Vision pelvis 2D actuel vs keyframe pelvis 2D.
+                if t_now - kf.t > self.hold_window_s:
+                    # Trop ancien, on lache le rig (mesh statique)
+                    out.append(person)
+                    continue
+                matched_pid = self._vision_pid_map.get(person.pid)
+                if matched_pid is None or kf.vision_pelvis_2d is None:
+                    out.append(person)
+                    continue
+                try:
+                    i = persons_body_ids.index(matched_pid)
+                except ValueError:
+                    out.append(person)
+                    continue
+                current_vp = _pelvis_2d_from_body(persons_body[i])
+                if current_vp is None:
+                    out.append(person)
+                    continue
+
+                # Image-normalized 2D delta -> world XY delta a depth z_kf.
+                # Pour un pelvis aux coords image (px in [0,1] centre 0.5),
+                # X_world = (px - 0.5) * IMG_SIZE * Z / focal = (px-0.5)*Z
+                # (focal=IMG_SIZE). Delta image -> Delta world a Z fixe.
+                z_kf = float(kf.translation[2]) if abs(
+                    kf.translation[2]) > 1e-3 else 1.0
+                dx_img = current_vp[0] - kf.vision_pelvis_2d[0]
+                dy_img = current_vp[1] - kf.vision_pelvis_2d[1]
+                dx_world = dx_img * _IMG_SIZE * z_kf / _FOCAL
+                dy_world = dy_img * _IMG_SIZE * z_kf / _FOCAL
+
+                # Applique a tous les vertices + a translation.
+                new_verts = kf.vertices_3d.copy()
+                new_verts[:, 0] += np.float32(dx_world)
+                new_verts[:, 1] += np.float32(dy_world)
+                new_transl = kf.translation.copy()
+                new_transl[0] += np.float32(dx_world)
+                new_transl[1] += np.float32(dy_world)
+
+                out.append(SMPLXPerson(
+                    pid=person.pid,
+                    vertices_3d=new_verts,
+                    translation=new_transl,
+                    confidence=person.confidence,
+                    betas=person.betas,
+                    expression=person.expression,
+                ))
+            return out
+
+    @staticmethod
+    def _project_pelvis(
+        translation: np.ndarray,
+    ) -> tuple[float, float] | None:
+        """World pelvis (X,Y,Z) -> image-normalized 2D pelvis."""
+        z = float(translation[2])
+        if abs(z) < 1e-3:
+            return None
+        x_img = (float(translation[0]) * _FOCAL / z) / _IMG_SIZE + 0.5
+        y_img = (float(translation[1]) * _FOCAL / z) / _IMG_SIZE + 0.5
+        # Clamp en [0,1]
+        if not (0.0 <= x_img <= 1.0 and 0.0 <= y_img <= 1.0):
+            return None
+        return (x_img, y_img)
diff --git a/data_only_viz/multi_hmr_worker.py b/data_only_viz/multi_hmr_worker.py
index 415f851..a00b6dc 100644
--- a/data_only_viz/multi_hmr_worker.py
+++ b/data_only_viz/multi_hmr_worker.py
@@ -30,6 +30,7 @@ CACHE = Path.home() / ".cache" / "av-live-multihmr"
 CKPT = CACHE / "checkpoints" / "multiHMR_672_S.pt"
 SMPLX_PATH = CACHE / "models" / "smplx" / "SMPLX_NEUTRAL.npz"
 MULTIHMR_REPO = CACHE / "multi-hmr"
+COREML_MLPACKAGE = CACHE / "multihmr_full_672_s.mlpackage"
 
 IMG_SIZE = 672
 N_VERTS = 10475
@@ -41,7 +42,8 @@ class MultiHMRWorker:
                  det_thresh: float = 0.3,
                  nms_kernel_size: int = 5,
                  motion_gate: float = 5.0,
-                 camera_index: int = -1) -> None:
+                 camera_index: int = -1,
+                 backend: str | None = None) -> None:
         self.state = state
         self.num_persons = num_persons
         self.period = 1.0 / max(1.0, target_fps)
@@ -55,6 +57,12 @@ class MultiHMRWorker:
         self.motion_gate = motion_gate
         # -1 = auto-select Mac BuiltInWideAngleCamera (cf _camera_select)
         self.camera_index = camera_index
+        # backend: 'pytorch' (default) or 'coreml'. CoreML uses the
+        # .mlpackage at COREML_MLPACKAGE, bypasses MPS torch, and runs
+        # on ANE/GPU/CPU via CoreML.framework natively (3-4x faster).
+        self.backend = (backend
+                        or os.environ.get("MULTIHMR_BACKEND", "pytorch")
+                        ).strip().lower()
         self._stop = threading.Event()
         self._thread: threading.Thread | None = None
         self._smooth_shape = [
@@ -72,6 +80,9 @@ class MultiHMRWorker:
 
     @staticmethod
     def is_available() -> bool:
+        backend = os.environ.get("MULTIHMR_BACKEND", "pytorch").strip().lower()
+        if backend == "coreml":
+            return COREML_MLPACKAGE.exists()
         return CKPT.exists() and SMPLX_PATH.exists() and MULTIHMR_REPO.exists()
 
     def start(self) -> None:
@@ -83,6 +94,12 @@ class MultiHMRWorker:
         self._stop.set()
 
     def _run(self) -> None:
+        if self.backend == "coreml":
+            self._run_coreml()
+            return
+        self._run_pytorch()
+
+    def _run_pytorch(self) -> None:
         if str(MULTIHMR_REPO) not in sys.path:
             sys.path.insert(0, str(MULTIHMR_REPO))
         # Multi-HMR demo.py tire pyrender / pyvista (OpenGL offscreen) et
@@ -415,3 +432,223 @@ class MultiHMRWorker:
 
         cap.stop()
         LOG.info("multi_hmr worker stopped")
+
+    # ------------------------------------------------------------------
+    # CoreML backend
+    # ------------------------------------------------------------------
+    def _run_coreml(self) -> None:
+        """CoreML inference path (ANE+GPU+CPU via Apple's framework).
+
+        Mirrors _run_pytorch but loads the .mlpackage via pyobjc + the
+        CoreML.framework, bypassing torch/MPS entirely. ~3-4x faster
+        on M5 (28.8ms median vs ~100ms with MPS)."""
+        try:
+            import cv2
+        except ImportError as e:
+            LOG.error("opencv-python missing: %s", e)
+            return
+        try:
+            from .multihmr_coreml import MultiHMRCoreMLBackend
+            backend = MultiHMRCoreMLBackend(COREML_MLPACKAGE)
+        except Exception as e:  # noqa: BLE001
+            LOG.error("CoreML backend init failed: %s", e)
+            return
+
+        focal = float(IMG_SIZE)
+        K_np = np.array([[focal, 0.0, IMG_SIZE / 2.0],
+                         [0.0, focal, IMG_SIZE / 2.0],
+                         [0.0, 0.0, 1.0]], dtype=np.float32)
+
+        from ._av_capture import (
+            AVCapture, find_builtin_device, enumerate_devices)
+        if self.camera_index >= 0:
+            devs = enumerate_devices()
+            if self.camera_index >= len(devs):
+                LOG.error("camera_index %d hors de %d devices",
+                          self.camera_index, len(devs))
+                return
+            info = devs[self.camera_index]
+        else:
+            info = find_builtin_device()
+            if info is None:
+                LOG.error("aucune BuiltInWideAngleCamera trouvee")
+                return
+        cap = AVCapture(info)
+        if not cap.start():
+            LOG.error("AVCapture start failed pour %s", info["name"])
+            return
+        LOG.info("camera ouverte %s (%s) [coreml backend]",
+                 info["name"], info["type"])
+
+        frame_count = 0
+        persons_count = 0
+        skipped_static = 0
+        next_heartbeat = time.monotonic() + 5.0
+        prev_thumb: np.ndarray | None = None
+
+        while not self._stop.is_set():
+            t_cap_start = time.monotonic()
+            ok, frame_bgr = cap.read(timeout_s=0.5)
+            if not ok or frame_bgr is None:
+                time.sleep(self.period)
+                continue
+
+            t_pre_start = time.monotonic()
+            h, w = frame_bgr.shape[:2]
+            if (h, w) != (IMG_SIZE, IMG_SIZE):
+                side = min(h, w)
+                y0 = (h - side) // 2
+                x0 = (w - side) // 2
+                frame_bgr = frame_bgr[y0:y0 + side, x0:x0 + side]
+                frame_bgr = cv2.resize(frame_bgr, (IMG_SIZE, IMG_SIZE))
+
+            if self.motion_gate > 0:
+                thumb = cv2.cvtColor(
+                    cv2.resize(frame_bgr, (112, 112)),
+                    cv2.COLOR_BGR2GRAY)
+                if prev_thumb is not None:
+                    diff_mean = float(np.mean(
+                        cv2.absdiff(thumb, prev_thumb)))
+                    if diff_mean < self.motion_gate:
+                        prev_thumb = thumb
+                        skipped_static += 1
+                        time.sleep(self.period)
+                        continue
+                prev_thumb = thumb
+
+            frame_rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
+            img = frame_rgb.transpose(2, 0, 1).astype(np.float32) / 255.0
+
+            t_inf_start = time.monotonic()
+            try:
+                humans = backend.infer(img, K_np, det_thresh=self.det_thresh)
+            except Exception as e:  # noqa: BLE001
+                LOG.warning("coreml inference failed: %s", e)
+                time.sleep(self.period)
+                continue
+
+            t_post_start = time.monotonic()
+            t_now = time.monotonic()
+            frame_count += 1
+            persons_count += len(humans) if humans else 0
+            if t_now >= next_heartbeat:
+                fps = frame_count / 5.0
+                avg = persons_count / max(1, frame_count)
+                LOG.info(
+                    "hb[coreml]: %.1f fps, %.2f persons/frame, %d skipped",
+                    fps, avg, skipped_static)
+                frame_count = 0
+                persons_count = 0
+                skipped_static = 0
+                next_heartbeat = t_now + 5.0
+
+            if not humans:
+                with self.state.lock():
+                    self.state.persons_smplx = []
+                time.sleep(self.period)
+                continue
+
+            # Dedup intra-frame (same logic as pytorch path).
+            cand: list[tuple[
+                float, float, float, float, float,
+                np.ndarray, int]] = []
+            for i, hh in enumerate(humans):
+                v = hh["v3d"].detach().cpu().numpy()
+                xmin = float(v[:, 0].min()); ymin = float(v[:, 1].min())
+                xmax = float(v[:, 0].max()); ymax = float(v[:, 1].max())
+                score = float(hh["scores"].item())
+                pelv = hh["transl_pelvis"].detach().cpu().numpy(
+                    ).flatten()[:3]
+                cand.append((score, xmin, ymin, xmax, ymax, pelv, i))
+            cand.sort(key=lambda c: -c[0])
+            keep_idx: list[int] = []
+            kept: list[tuple[float, float, float, float, np.ndarray]] = []
+            for sc, x0, y0, x1, y1, pelv, src_i in cand:
+                a_area = max(0.0, x1 - x0) * max(0.0, y1 - y0)
+                drop = False
+                for (kx0, ky0, kx1, ky1, kpelv) in kept:
+                    ix0 = max(x0, kx0); iy0 = max(y0, ky0)
+                    ix1 = min(x1, kx1); iy1 = min(y1, ky1)
+                    iw = max(0.0, ix1 - ix0); ih = max(0.0, iy1 - iy0)
+                    inter = iw * ih
+                    if a_area <= 0 or inter <= 0:
+                        continue
+                    k_area = (kx1 - kx0) * (ky1 - ky0)
+                    iou = inter / (a_area + k_area - inter + 1e-9)
+                    pelv_d = float(np.linalg.norm(pelv - kpelv))
+                    if iou > 0.55 and pelv_d < 0.20:
+                        drop = True
+                        break
+                if not drop:
+                    keep_idx.append(src_i)
+                    kept.append((x0, y0, x1, y1, pelv))
+                    if len(keep_idx) >= self.num_persons:
+                        break
+            humans = [humans[i] for i in keep_idx]
+            n_keep = len(humans)
+
+            bboxes = []
+            for hh in humans:
+                v = hh["v3d"].detach().cpu().numpy()
+                xmin, ymin = float(v[:, 0].min()), float(v[:, 1].min())
+                xmax, ymax = float(v[:, 0].max()), float(v[:, 1].max())
+                bboxes.append([PoseKp(x=xmin, y=ymin, c=1.0),
+                               PoseKp(x=xmax, y=ymax, c=1.0)])
+            ids = self._tracker.update(bboxes)
+
+            persons: list[SMPLXPerson] = []
+            for i, hh in enumerate(humans[:n_keep]):
+                pid = ids[i] if i < len(ids) else i
+                if pid < 0:
+                    continue
+                v3d = hh["v3d"].detach().cpu().numpy()
+                transl_np = hh["transl_pelvis"].detach().cpu().numpy().flatten()
+                shape_raw = hh["shape"].detach().cpu().numpy().flatten()
+                expr_raw = hh["expression"].detach().cpu().numpy().flatten()
+
+                pid_c = pid % self.num_persons
+                shape_n = min(10, len(shape_raw))
+                expr_n = min(10, len(expr_raw))
+                shape_smooth = np.zeros(10, dtype=np.float32)
+                expr_smooth = np.zeros(10, dtype=np.float32)
+                for k in range(shape_n):
+                    shape_smooth[k] = self._smooth_shape[pid_c][k](
+                        float(shape_raw[k]), t_now)
+                for k in range(expr_n):
+                    expr_smooth[k] = self._smooth_expr[pid_c][k](
+                        float(expr_raw[k]), t_now)
+
+                persons.append(SMPLXPerson(
+                    pid=int(pid),
+                    vertices_3d=np.ascontiguousarray(v3d, dtype=np.float32),
+                    translation=np.ascontiguousarray(
+                        transl_np[:3], dtype=np.float32),
+                    confidence=float(hh["scores"].item()),
+                    betas=np.ascontiguousarray(shape_smooth, dtype=np.float32),
+                    expression=np.ascontiguousarray(expr_smooth, dtype=np.float32),
+                ))
+
+            with self.state.lock():
+                self.state.persons_smplx = persons
+                self.state.smplx_last_t = t_now
+
+            t_end = time.monotonic()
+            dt_total = (t_end - t_cap_start) * 1e3
+            if LOG.isEnabledFor(logging.DEBUG) or dt_total > 100.0:
+                LOG.log(
+                    logging.DEBUG if dt_total <= 100.0 else logging.WARNING,
+                    "frame[coreml]: cap=%.1f pre=%.1f inf=%.1f "
+                    "post=%.1fms total=%.1fms",
+                    (t_pre_start - t_cap_start) * 1e3,
+                    (t_inf_start - t_pre_start) * 1e3,
+                    (t_post_start - t_inf_start) * 1e3,
+                    (t_end - t_post_start) * 1e3,
+                    dt_total,
+                )
+
+            dt = time.monotonic() - t_cap_start
+            if dt < self.period:
+                time.sleep(self.period - dt)
+
+        cap.stop()
+        LOG.info("multi_hmr coreml worker stopped")
diff --git a/data_only_viz/multihmr_coreml.py b/data_only_viz/multihmr_coreml.py
new file mode 100644
index 0000000..3992800
--- /dev/null
+++ b/data_only_viz/multihmr_coreml.py
@@ -0,0 +1,276 @@
+"""Multi-HMR CoreML backend (ANE/GPU/CPU via Apple's CoreML framework).
+
+Python 3.14 cannot use `coremltools.MLModel` because `libcoremlpython`
+and `libmilstoragepython` native extensions are not distributed for
+3.14. We load CoreML.framework directly via `objc.loadBundle()` —
+same pattern as `coreml_pose.py`.
+
+Unlike `coreml_pose.py`, this backend does NOT use Vision: Vision is
+limited to image inputs and cannot feed a second MLMultiArray (cam_K).
+We invoke `MLModel.predictionFromFeatures:error:` directly with a
+`MLDictionaryFeatureProvider` wrapping two `MLMultiArray`s.
+
+Public API:
+    backend = MultiHMRCoreMLBackend(mlpackage_path)
+    humans = backend.infer(image_chw_f32, K_33_f32, det_thresh=0.3)
+    # humans is a list[dict] with the same keys as the PyTorch model
+    # output. Values are CoreMLArray instances that quack like torch
+    # tensors (.detach().cpu().numpy() / .item()).
+"""
+from __future__ import annotations
+
+import logging
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import objc
+from Foundation import NSURL
+
+LOG = logging.getLogger("multihmr_coreml")
+
+DEFAULT_MLPACKAGE = (
+    Path.home() / ".cache" / "av-live-multihmr"
+    / "multihmr_full_672_s.mlpackage"
+)
+
+# Multi-HMR exported with apply_topk(K=4): outputs are fixed shape.
+N_PERSONS_FIXED = 4
+N_VERTS = 10475
+
+# CoreML output names from the exported .mlpackage.
+OUT_V3D = "var_2412"          # (4, 10475, 3)
+OUT_TRANSL = "var_2415"       # (4, 1, 3)
+OUT_SCORES = "var_2428"       # (4,)
+OUT_BETAS = "var_2431"        # (4, 10)
+OUT_EXPR = "var_2434"         # (4, 10)
+
+# MLMultiArrayDataType raw values (from CoreML headers).
+ML_DTYPE_FLOAT32 = 65568
+ML_DTYPE_FLOAT16 = 65552
+ML_DTYPE_DOUBLE = 65600
+ML_DTYPE_INT32 = 131104
+
+
+_NS: dict[str, Any] = {}
+_FRAMEWORKS_LOADED = False
+
+
+def _load_frameworks() -> dict[str, Any]:
+    global _FRAMEWORKS_LOADED
+    if _FRAMEWORKS_LOADED:
+        return _NS
+    objc.loadBundle("CoreML", _NS,
+                    "/System/Library/Frameworks/CoreML.framework")
+    _FRAMEWORKS_LOADED = True
+    return _NS
+
+
+class CoreMLArray:
+    """Tiny tensor-like adapter so the existing worker hot path can
+    treat CoreML outputs the same way it treats torch tensors.
+
+    Supports `.detach().cpu().numpy()` and `.item()`. The wrapper is
+    a no-op around a numpy array; we keep the chain so callers don't
+    need any conditional branch."""
+
+    __slots__ = ("_arr",)
+
+    def __init__(self, arr: np.ndarray) -> None:
+        self._arr = arr
+
+    def detach(self) -> "CoreMLArray":
+        return self
+
+    def cpu(self) -> "CoreMLArray":
+        return self
+
+    def numpy(self) -> np.ndarray:
+        return self._arr
+
+    def item(self) -> float:
+        return float(self._arr.reshape(-1)[0])
+
+    @property
+    def shape(self) -> tuple[int, ...]:
+        return tuple(self._arr.shape)
+
+
+def _np_to_mlarray(arr: np.ndarray):
+    """Create a contiguous float32 MLMultiArray from a numpy array.
+
+    We always feed FLOAT32 — even though outputs are FLOAT16, CoreML
+    will auto-cast on the input side."""
+    ns = _load_frameworks()
+    MLMultiArray = ns["MLMultiArray"]
+    arr = np.ascontiguousarray(arr, dtype=np.float32)
+    shape = [int(s) for s in arr.shape]
+    ml = MLMultiArray.alloc().initWithShape_dataType_error_(
+        shape, ML_DTYPE_FLOAT32, None)
+    if ml is None:
+        raise RuntimeError("MLMultiArray alloc failed")
+    # Copy bytes through dataPointer (raw void*). pyobjc exposes it as
+    # a memoryview-like opaque; we use ctypes to memcpy.
+    import ctypes
+    ptr = ml.dataPointer()
+    n_bytes = arr.nbytes
+    # pyobjc returns either an objc.varlist or a Python int pointer.
+    addr = int(ptr) if isinstance(ptr, int) else ctypes.cast(
+        ptr, ctypes.c_void_p).value
+    if addr is None:
+        raise RuntimeError("MLMultiArray dataPointer null")
+    ctypes.memmove(addr, arr.ctypes.data, n_bytes)
+    return ml
+
+
+def _mlarray_to_np(ml) -> np.ndarray:
+    """Copy an MLMultiArray (FLOAT16 or FLOAT32) into a numpy float32."""
+    import ctypes
+    shape = tuple(int(s) for s in ml.shape())
+    dtype_id = int(ml.dataType())
+    count = 1
+    for s in shape:
+        count *= s
+    ptr = ml.dataPointer()
+    addr = int(ptr) if isinstance(ptr, int) else ctypes.cast(
+        ptr, ctypes.c_void_p).value
+    if addr is None:
+        raise RuntimeError("MLMultiArray dataPointer null")
+    if dtype_id == ML_DTYPE_FLOAT16:
+        raw = (ctypes.c_uint16 * count).from_address(addr)
+        arr = np.ctypeslib.as_array(raw).view(np.float16).astype(np.float32)
+    elif dtype_id == ML_DTYPE_FLOAT32:
+        raw = (ctypes.c_float * count).from_address(addr)
+        arr = np.ctypeslib.as_array(raw).copy()
+    elif dtype_id == ML_DTYPE_DOUBLE:
+        raw = (ctypes.c_double * count).from_address(addr)
+        arr = np.ctypeslib.as_array(raw).astype(np.float32)
+    else:
+        raise RuntimeError(f"unsupported MLMultiArray dtype {dtype_id}")
+    return arr.reshape(shape)
+
+
+class MultiHMRCoreMLBackend:
+    """CoreML inference wrapper for Multi-HMR (full_672_s)."""
+
+    def __init__(self, mlpackage_path: Path | None = None) -> None:
+        self.path = Path(mlpackage_path) if mlpackage_path else DEFAULT_MLPACKAGE
+        if not self.path.exists():
+            raise FileNotFoundError(f"mlpackage missing: {self.path}")
+        ns = _load_frameworks()
+        MLModel = ns["MLModel"]
+        MLModelConfiguration = ns["MLModelConfiguration"]
+        cfg = MLModelConfiguration.alloc().init()
+        try:
+            # MLComputeUnits: 0=CPUOnly, 1=CPUAndGPU, 2=All (ANE+GPU+CPU),
+            # 3=CPUAndNeuralEngine. Multi-HMR's ANEF compile fails
+            # (validated 2026-05-13 on M5), and 'All' falls back to a
+            # slow path (~146ms). CPU+GPU = 28ms = ~35fps on M5.
+            cfg.setComputeUnits_(1)
+        except Exception:  # noqa: BLE001
+            pass
+        url = NSURL.fileURLWithPath_(str(self.path))
+        # .mlpackage must be compiled to .mlmodelc before MLModel can
+        # load it. compileModelAtURL_error_ returns an NSURL to a
+        # temp .mlmodelc bundle.
+        compiled_url = MLModel.compileModelAtURL_error_(url, None)
+        if compiled_url is None:
+            raise RuntimeError(f"compileModelAtURL failed for {self.path}")
+        model = MLModel.modelWithContentsOfURL_configuration_error_(
+            compiled_url, cfg, None)
+        if model is None:
+            raise RuntimeError(f"MLModel load failed for {compiled_url}")
+        self._model = model
+        self._ns = ns
+        LOG.info("Multi-HMR CoreML model loaded (%s, computeUnits=CPU+GPU)",
+                 self.path.name)
+
+    @staticmethod
+    def is_available(mlpackage_path: Path | None = None) -> bool:
+        p = Path(mlpackage_path) if mlpackage_path else DEFAULT_MLPACKAGE
+        if not p.exists():
+            return False
+        try:
+            _load_frameworks()
+            return True
+        except Exception:  # noqa: BLE001
+            return False
+
+    def _predict(self, image_4d: np.ndarray, K_33: np.ndarray) -> dict:
+        ns = self._ns
+        MLDictionaryFeatureProvider = ns["MLDictionaryFeatureProvider"]
+        MLFeatureValue = ns["MLFeatureValue"]
+        img_ml = _np_to_mlarray(image_4d)
+        k_ml = _np_to_mlarray(K_33)
+        feats = {
+            "image": MLFeatureValue.featureValueWithMultiArray_(img_ml),
+            "cam_K": MLFeatureValue.featureValueWithMultiArray_(k_ml),
+        }
+        provider = MLDictionaryFeatureProvider.alloc(
+            ).initWithDictionary_error_(feats, None)
+        if provider is None:
+            raise RuntimeError("MLDictionaryFeatureProvider alloc failed")
+        out = self._model.predictionFromFeatures_error_(provider, None)
+        if out is None:
+            raise RuntimeError("MLModel predict failed")
+        names = [str(n) for n in out.featureNames()]
+        result = {}
+        for n in names:
+            fv = out.featureValueForName_(n)
+            ml = fv.multiArrayValue()
+            if ml is None:
+                continue
+            result[n] = _mlarray_to_np(ml)
+        return result
+
+    def infer(
+        self,
+        image_chw_float32: np.ndarray,
+        K_33: np.ndarray,
+        det_thresh: float = 0.3,
+    ) -> list[dict]:
+        """Run a forward pass and return list of humans dicts.
+
+        Args:
+            image_chw_float32: (3, 672, 672) or (1, 3, 672, 672) in [0,1].
+            K_33: (3, 3) or (1, 3, 3) camera intrinsics.
+            det_thresh: scores threshold; CoreML forwards K=4 always.
+
+        Returns:
+            list[dict] with keys v3d, transl_pelvis, scores, shape,
+            expression. Values are CoreMLArray wrappers.
+        """
+        img = np.asarray(image_chw_float32, dtype=np.float32)
+        if img.ndim == 3:
+            img = img[np.newaxis, ...]
+        if img.shape != (1, 3, 672, 672):
+            raise ValueError(f"image shape {img.shape}, expected (1,3,672,672)")
+        K = np.asarray(K_33, dtype=np.float32)
+        if K.ndim == 2:
+            K = K[np.newaxis, ...]
+        if K.shape != (1, 3, 3):
+            raise ValueError(f"K shape {K.shape}, expected (1,3,3)")
+
+        raw = self._predict(img, K)
+        v3d = raw.get(OUT_V3D)
+        transl = raw.get(OUT_TRANSL)
+        scores = raw.get(OUT_SCORES)
+        betas = raw.get(OUT_BETAS)
+        expr = raw.get(OUT_EXPR)
+        if any(x is None for x in (v3d, transl, scores, betas, expr)):
+            raise RuntimeError(
+                "missing outputs; got keys=" + ",".join(raw.keys()))
+
+        humans: list[dict] = []
+        for k in range(N_PERSONS_FIXED):
+            sc = float(scores[k])
+            if sc < det_thresh:
+                continue
+            humans.append({
+                "v3d": CoreMLArray(v3d[k]),                # (10475, 3)
+                "transl_pelvis": CoreMLArray(transl[k]),   # (1, 3)
+                "scores": CoreMLArray(np.array([sc], dtype=np.float32)),
+                "shape": CoreMLArray(betas[k]),            # (10,)
+                "expression": CoreMLArray(expr[k]),        # (10,)
+            })
+        return humans
diff --git a/data_only_viz/scripts/capture_actions.py b/data_only_viz/scripts/capture_actions.py
new file mode 100644
index 0000000..331858d
--- /dev/null
+++ b/data_only_viz/scripts/capture_actions.py
@@ -0,0 +1,74 @@
+"""Record webcam frames + timestamps for action-head training.
+
+Usage:
+    uv run python -m data_only_viz.scripts.capture_actions \
+        --session sess03 --duration 600
+"""
+from __future__ import annotations
+
+import argparse
+import logging
+import time
+from pathlib import Path
+
+import cv2
+
+LOG = logging.getLogger("capture_actions")
+RAW_DIR = Path("~/.cache/av-live-action/raw").expanduser()
+
+
+def capture(session: str, duration_s: float,
+            cam_index: int = 0, fps: int = 30,
+            size: int = 672) -> Path:
+    RAW_DIR.mkdir(parents=True, exist_ok=True)
+    out = RAW_DIR / f"{session}.mp4"
+    ts_out = RAW_DIR / f"{session}.ts.txt"
+    cap = cv2.VideoCapture(cam_index)
+    if not cap.isOpened():
+        raise RuntimeError(f"cannot open camera {cam_index}")
+    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+    writer = cv2.VideoWriter(str(out), fourcc, fps, (size, size))
+    try:
+        t_start = time.perf_counter()
+        with ts_out.open("w") as ts_f:
+            n = 0
+            while time.perf_counter() - t_start < duration_s:
+                ok, frame = cap.read()
+                if not ok:
+                    LOG.warning("frame read failed")
+                    break
+                h, w = frame.shape[:2]
+                side = min(h, w)
+                y0 = (h - side) // 2
+                x0 = (w - side) // 2
+                crop = frame[y0:y0 + side, x0:x0 + side]
+                resized = cv2.resize(crop, (size, size))
+                writer.write(resized)
+                ts_f.write(f"{n} {time.perf_counter() - t_start:.6f}\n")
+                n += 1
+                cv2.imshow("capture (q=quit)", resized)
+                if cv2.waitKey(1) & 0xFF == ord("q"):
+                    break
+        LOG.info("wrote %s (%d frames)", out, n)
+        return out
+    finally:
+        cap.release()
+        writer.release()
+        cv2.destroyAllWindows()
+
+
+def _cli() -> None:
+    p = argparse.ArgumentParser()
+    p.add_argument("--session", required=True)
+    p.add_argument("--duration", type=float, default=600.0)
+    p.add_argument("--cam-index", type=int, default=0)
+    p.add_argument("--fps", type=int, default=30)
+    args = p.parse_args()
+    logging.basicConfig(level=logging.INFO,
+                        format="%(asctime)s [%(name)s] %(message)s")
+    capture(args.session, args.duration,
+            cam_index=args.cam_index, fps=args.fps)
+
+
+if __name__ == "__main__":
+    _cli()
diff --git a/data_only_viz/scripts/coreml_full_probe.py b/data_only_viz/scripts/coreml_full_probe.py
index 241ce15..f1f4387 100644
--- a/data_only_viz/scripts/coreml_full_probe.py
+++ b/data_only_viz/scripts/coreml_full_probe.py
@@ -157,6 +157,32 @@ if hasattr(model.backbone, "encoder") and hasattr(model.backbone.encoder,
 # torch.inverse(K) plante coremltools (op non implementee). Comme K est
 # fixe (camera intrinsics avec focal=IMG_SIZE), on pre-calcule K_inv
 # en closed-form et on l'utilise comme buffer module-level.
+print("==> Patching roma.rotmat_to_rotvec (branchless atan2)")
+# roma.rotmat_to_rotvec utilise torch.empty + 8 index_put_ qui se
+# traduisent en CoreML par scatter_nd successifs sur un buffer
+# garbage-initialise. Resultat : cellules non touchees restent NaN,
+# propagees via quat normalization -> v3d/transl all-NaN.
+# Remplacement branchless via atan2 : pas de torch.empty, pas
+# d'index_put_, juste des stack/clamp/norm/atan2 stables CoreML.
+# Precision vs roma original : 2.26e-6 L_inf sur batch random.
+import roma as _roma
+
+def _rotmat_to_rotvec_branchless(R, eps=1e-6):
+    w = torch.stack([
+        R[..., 2, 1] - R[..., 1, 2],
+        R[..., 0, 2] - R[..., 2, 0],
+        R[..., 1, 0] - R[..., 0, 1],
+    ], dim=-1) * 0.5
+    trace = R[..., 0, 0] + R[..., 1, 1] + R[..., 2, 2]
+    cos_theta = ((trace - 1.0) * 0.5).clamp(-1.0, 1.0)
+    sin_theta = torch.norm(w, dim=-1)
+    theta = torch.atan2(sin_theta, cos_theta)
+    sin_theta_safe = sin_theta.clamp(min=eps)
+    return w * (theta / sin_theta_safe).unsqueeze(-1)
+
+_roma.rotmat_to_rotvec = _rotmat_to_rotvec_branchless
+
+
 print("==> Patching utils.camera.inverse_perspective_projection")
 import utils.camera as _camera
 
@@ -170,11 +196,28 @@ _K_INV_PRE = torch.tensor([
 ])
 
 def inverse_perspective_projection_fixed(points, K, distance):
-    """Bypass torch.inverse : utilise K_inv pre-calcule en closed-form
-    (notre K est connu et fixe). Le K argument est ignore."""
-    K_inv = _K_INV_PRE.to(points.device).to(points.dtype)
-    points = torch.cat([points, torch.ones_like(points[..., :1])], -1)
-    points = torch.einsum('bij,bkj->bki', K_inv, points)
+    """Bypass torch.inverse + einsum + matmul pour eviter le bug
+    coremltools de broadcast batch 1->K sur ces ops. K_inv etant
+    fixe et structure (diag + translate), on ecrit les composantes
+    explicitement en ops elementaires.
+
+    K_inv = [[1/f, 0, -cx/f], [0, 1/f, -cy/f], [0, 0, 1]]
+    Pour points (b, N, 3) : out = points @ K_inv.T donne :
+      out[..., 0] = points[..., 0]/f - (cx/f) * points[..., 2]
+      out[..., 1] = points[..., 1]/f - (cy/f) * points[..., 2]
+      out[..., 2] = points[..., 2]
+    """
+    points_hom = torch.cat([points, torch.ones_like(points[..., :1])], -1)
+    inv_f = 1.0 / focal_val
+    cx_over_f = cx / focal_val
+    cy_over_f = cy / focal_val
+    x = points_hom[..., 0:1]
+    y = points_hom[..., 1:2]
+    z = points_hom[..., 2:3]
+    out0 = x * inv_f - z * cx_over_f
+    out1 = y * inv_f - z * cy_over_f
+    out2 = z
+    points = torch.cat([out0, out1, out2], dim=-1)
     if distance is None:
         return points
     points = points * distance
@@ -190,6 +233,26 @@ model_mod.inverse_perspective_projection = inverse_perspective_projection_fixed
 import blocks.smpl_layer as _smpl_layer
 _smpl_layer.inverse_perspective_projection = inverse_perspective_projection_fixed
 
+# Aussi perspective_projection (utilise dans smpl_layer.py:143-144 pour
+# j2d et v2d) -> rewrite einsum en matmul pour le meme broadcast bug.
+def perspective_projection_fixed(x, K):
+    """Element-wise rewrite de la projection perspective avec K fixe
+    (focal=IMG_SIZE, cx=cy=IMG_SIZE/2). Bypass matmul/einsum pour eviter
+    les bugs broadcast coremltools.
+    K = [[f, 0, cx], [0, f, cy], [0, 0, 1]]
+    out[..., 0] = f * x_norm + cx * z_norm (mais on veut [..., :2])
+                = f * (x/z) + cx
+    out[..., 1] = f * (y/z) + cy
+    """
+    z = x[..., 2:3]
+    px = x[..., 0:1] / z * focal_val + cx
+    py = x[..., 1:2] / z * focal_val + cy
+    return torch.cat([px, py], dim=-1)
+
+_camera.perspective_projection = perspective_projection_fixed
+_utils_pkg.perspective_projection = perspective_projection_fixed
+_smpl_layer.perspective_projection = perspective_projection_fixed
+
 
 # === Wrapper qui produit tuple fixe ===
 class TracedMHMR(nn.Module):
@@ -222,6 +285,12 @@ class TracedMHMR(nn.Module):
         ]).squeeze(-1)
         shape = torch.stack([h["shape"] for h in humans])
         expr = torch.stack([h["expression"] for h in humans])
+        # NOTE: CoreML mlprogram conversion currently produces all-NaN
+        # outputs for v3d and transl while PyTorch eager produces valid
+        # finite values from the same trace. nan_to_num here masks the
+        # symptom but yields all-zero meshes (no information). Leave
+        # raw outputs and let downstream decide; investigation tracked
+        # in task #2 (op-by-op bisection needed).
         return v3d, transl, scores, shape, expr
 
 
@@ -422,6 +491,28 @@ def _diagonal_general(context, node):
 
 _TORCH_OPS_REGISTRY.name_to_func_mapping["diagonal"] = _diagonal_general
 
+
+# Instrument reshape pour logger node source au moment de l'erreur.
+from coremltools.converters.mil.mil.ops.defs.iOS15 import tensor_transformation as _tt
+_orig_reshape_ti = _tt.reshape.type_inference
+
+
+def _reshape_ti_logged(self):
+    try:
+        return _orig_reshape_ti(self)
+    except ValueError as e:
+        if "Invalid target shape" in str(e):
+            try:
+                from_shape = list(self.x.shape)
+                target = list(self.shape.val) if hasattr(self.shape, "val") else "?"
+                print(f"  >>> RESHAPE FAIL : name={self.name} from={from_shape} target={target}")
+            except Exception:
+                pass
+        raise
+
+
+_tt.reshape.type_inference = _reshape_ti_logged
+
 try:
     mlmodel = ct.convert(
         traced,
@@ -433,6 +524,10 @@ try:
         compute_units=ct.ComputeUnit.CPU_AND_GPU,
         minimum_deployment_target=ct.target.macOS15,
         convert_to="mlprogram",
+        # FP16 OK depuis le patch roma branchless (cf rapport bisection
+        # 2026-05-13) : la source du NaN etait torch.empty + index_put_
+        # dans roma.rotmat_to_rotvec, pas la precision.
+        compute_precision=ct.precision.FLOAT16,
     )
     out_path = "/tmp/multihmr_full_672_s.mlpackage"
     mlmodel.save(out_path)
diff --git a/data_only_viz/scripts/extract_j3d_offline.py b/data_only_viz/scripts/extract_j3d_offline.py
index 1ad9254..6fb5b09 100644
--- a/data_only_viz/scripts/extract_j3d_offline.py
+++ b/data_only_viz/scripts/extract_j3d_offline.py
@@ -1,4 +1,4 @@
-"""Extract j3d (22 SMPL-X joint anchors) from a recorded MP4 using the
+"""Extract j3d (32 SMPL-X joint anchors) from a recorded MP4 using the
 Multi-HMR CoreML backend, write per-frame per-person jsonl rows.
 
 Usage:
@@ -17,7 +17,12 @@ from pathlib import Path
 import cv2
 import numpy as np
 
-from data_only_viz.action_head_pub import SMPLX_JOINT_ANCHOR_VERTS
+from data_only_viz.action_head import EXPR_DIM, HANDS_KP_DIMS, HANDS_KP_TOTAL
+from data_only_viz.action_head_pub import (
+    SMPLX_JOINT_ANCHOR_VERTS,
+    SMPLX_UPPER_LIP_VERT,
+    SMPLX_LOWER_LIP_VERT,
+)
 from data_only_viz.multihmr_coreml import MultiHMRCoreMLBackend
 
 LOG = logging.getLogger("extract_j3d_offline")
@@ -47,14 +52,37 @@ def _frame_to_chw(frame_bgr: np.ndarray, size: int = IMG_SIZE) -> np.ndarray:
     return rgb.transpose(2, 0, 1)  # CHW
 
 
-def _person_to_j3d22(person: dict, anchors: tuple[int, ...]) -> np.ndarray | None:
+def _person_to_j3d32(
+    person: dict,
+    anchors: tuple[int, ...],
+) -> tuple[np.ndarray, np.ndarray, float] | None:
+    """Return (j3d32, expression, mouth_open) or None if v3d absent/too small."""
     v3d = person.get("v3d")
     if v3d is None:
         return None
+    # CoreMLArray wraps numpy but lacks __array__; unwrap before asarray.
+    if hasattr(v3d, "numpy") and not isinstance(v3d, np.ndarray):
+        v3d = v3d.numpy()
     v3d_np = np.asarray(v3d, dtype=np.float32)
     if v3d_np.shape[0] < max(anchors) + 1:
         return None
-    return v3d_np[list(anchors)].astype(np.float32)
+    j3d32 = v3d_np[list(anchors)].astype(np.float32)
+    # expression
+    expr = person.get("expression")
+    if expr is not None:
+        if hasattr(expr, "numpy") and not isinstance(expr, np.ndarray):
+            expr = expr.numpy()
+        expr_np = np.asarray(expr, dtype=np.float32).flatten()
+    else:
+        expr_np = np.zeros(EXPR_DIM, dtype=np.float32)
+    # mouth_open
+    if v3d_np.shape[0] > max(SMPLX_UPPER_LIP_VERT, SMPLX_LOWER_LIP_VERT):
+        mouth = float(np.linalg.norm(
+            v3d_np[SMPLX_UPPER_LIP_VERT] - v3d_np[SMPLX_LOWER_LIP_VERT]
+        ))
+    else:
+        mouth = 0.0
+    return j3d32, expr_np, mouth
 
 
 def extract(session: str, video: Path, out: Path,
@@ -86,14 +114,20 @@ def extract(session: str, video: Path, out: Path,
                 continue
             ts = n_frames / fps
             for i, person in enumerate(persons):
-                j3d = _person_to_j3d22(person, anchors)
-                if j3d is None:
+                result = _person_to_j3d32(person, anchors)
+                if result is None:
                     continue
+                j3d32, expr_np, mouth = result
                 f.write(json.dumps({
                     "ts": ts,
                     "session": session,
                     "pid": int(person.get("pid", i)),
-                    "j3d": j3d.tolist(),
+                    "j3d": j3d32.tolist(),
+                    "expression": expr_np.tolist(),
+                    "mouth_open": mouth,
+                    "hands_kp": np.zeros(
+                        (HANDS_KP_TOTAL, HANDS_KP_DIMS), dtype=np.float32
+                    ).tolist(),
                 }) + "\n")
                 n_rows += 1
             n_frames += 1
diff --git a/data_only_viz/scripts/extract_mediapipe_offline.py b/data_only_viz/scripts/extract_mediapipe_offline.py
new file mode 100644
index 0000000..18fc108
--- /dev/null
+++ b/data_only_viz/scripts/extract_mediapipe_offline.py
@@ -0,0 +1,208 @@
+"""Extract action-head v3 jsonl rows from a recorded MP4 using MediaPipe
+Holistic. Populates real hands_kp (42, 3) and mouth_open (face lips
+distance), unlike extract_j3d_offline.py (SMPL-X path) which writes zeros
+for hands_kp.
+
+Output jsonl row format (matches dataset.py load_frames_jsonl) :
+
+    {
+      "ts":         float seconds,
+      "session":    str,
+      "pid":        int (always 0 — Holistic is single-person),
+      "j3d":        [[32, 3]] floats (body22 + 10 fingertips),
+      "expression": [10] zeros (MediaPipe has no SMPL-X PCA),
+      "mouth_open": float (lips inner distance),
+      "hands_kp":   [[42, 3]] floats (21 L + 21 R, zero-padded if absent),
+    }
+
+Usage :
+
+    uv run python -m data_only_viz.scripts.extract_mediapipe_offline \
+        --session sess03 \
+        --video ~/.cache/av-live-action/raw/sess03.mp4 \
+        --out   ~/.cache/av-live-action/raw/sess03_mp.jsonl
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+from data_only_viz.action_head import (
+    EXPR_DIM,
+    HANDS_KP_DIMS,
+    HANDS_KP_PER_HAND,
+    HANDS_KP_TOTAL,
+    J3D_BODY,
+    J3D_FINGERS,
+    J3D_FINGERS_PER_HAND,
+    J3D_JOINTS,
+)
+from data_only_viz.action_head_pub import (
+    MEDIAPIPE_HAND_FINGERTIPS,
+    MEDIAPIPE_LIP_LOWER_INNER,
+    MEDIAPIPE_LIP_UPPER_INNER,
+    MEDIAPIPE_TO_22,
+)
+
+LOG = logging.getLogger("extract_mediapipe_offline")
+DEFAULT_OUT_DIR = Path("~/.cache/av-live-action/raw").expanduser()
+
+
+def _build_landmarker():
+    """Build a MediaPipe HolisticLandmarker in VIDEO running mode."""
+    from mediapipe.tasks.python import vision
+    from mediapipe.tasks.python.core.base_options import BaseOptions
+    from data_only_viz.holistic import _ensure_model
+    model_path = _ensure_model()
+    opts = vision.HolisticLandmarkerOptions(
+        base_options=BaseOptions(model_asset_path=str(model_path)),
+        running_mode=vision.RunningMode.VIDEO,
+        min_pose_detection_confidence=0.3,
+        min_pose_landmarks_confidence=0.3,
+        min_face_detection_confidence=0.3,
+        min_face_landmarks_confidence=0.3,
+        min_hand_landmarks_confidence=0.3,
+    )
+    return vision.HolisticLandmarker.create_from_options(opts)
+
+
+def _lmk_list_to_array(lmks) -> np.ndarray | None:
+    """Convert MediaPipe NormalizedLandmark / Landmark list to (N, 3) array."""
+    if lmks is None:
+        return None
+    try:
+        return np.asarray(
+            [(lm.x, lm.y, getattr(lm, "z", 0.0)) for lm in lmks],
+            dtype=np.float32,
+        )
+    except (AttributeError, TypeError):
+        return None
+
+
+def _build_j3d32(body3d_arr: np.ndarray | None,
+                 hands_kp42: np.ndarray) -> np.ndarray | None:
+    """Map MediaPipe body3d (33, 3) + hands_kp (42, 3) -> j3d (32, 3).
+
+    body22 indices via MEDIAPIPE_TO_22, fingertips from hands_kp idx
+    MEDIAPIPE_HAND_FINGERTIPS (4, 8, 12, 16, 20) for each side.
+    """
+    if body3d_arr is None or body3d_arr.shape[0] < 33:
+        return None
+    body22 = body3d_arr[list(MEDIAPIPE_TO_22)].astype(np.float32)
+    tips = np.zeros((J3D_FINGERS, 3), dtype=np.float32)
+    for side_idx in (0, 1):
+        base = side_idx * HANDS_KP_PER_HAND
+        for k, mp_tip in enumerate(MEDIAPIPE_HAND_FINGERTIPS):
+            if base + mp_tip < hands_kp42.shape[0]:
+                tips[side_idx * J3D_FINGERS_PER_HAND + k] = hands_kp42[base + mp_tip]
+    return np.concatenate([body22, tips], axis=0)
+
+
+def _mouth_open(face_arr: np.ndarray | None) -> float:
+    if face_arr is None or face_arr.shape[0] <= MEDIAPIPE_LIP_LOWER_INNER:
+        return 0.0
+    upper = face_arr[MEDIAPIPE_LIP_UPPER_INNER]
+    lower = face_arr[MEDIAPIPE_LIP_LOWER_INNER]
+    return float(np.linalg.norm(upper - lower))
+
+
+def _hands_kp42(left_arr: np.ndarray | None,
+                right_arr: np.ndarray | None) -> np.ndarray:
+    out = np.zeros((HANDS_KP_TOTAL, HANDS_KP_DIMS), dtype=np.float32)
+    if left_arr is not None and left_arr.shape[0] >= HANDS_KP_PER_HAND:
+        out[:HANDS_KP_PER_HAND] = left_arr[:HANDS_KP_PER_HAND]
+    if right_arr is not None and right_arr.shape[0] >= HANDS_KP_PER_HAND:
+        out[HANDS_KP_PER_HAND:] = right_arr[:HANDS_KP_PER_HAND]
+    return out
+
+
+def extract(session: str, video: Path, out: Path) -> int:
+    """Run MediaPipe Holistic on every frame of video, write jsonl rows.
+
+    Returns the number of frames where at least body3d was detected
+    (rows written). Frames with no person are silently skipped.
+    """
+    import mediapipe as mp
+    out.parent.mkdir(parents=True, exist_ok=True)
+    cap = cv2.VideoCapture(str(video))
+    if not cap.isOpened():
+        raise RuntimeError(f"cannot open {video}")
+    fps = cap.get(cv2.CAP_PROP_FPS) or 30.0
+    landmarker = _build_landmarker()
+    n_frames = 0
+    n_rows = 0
+    expr_zeros_list = np.zeros(EXPR_DIM, dtype=np.float32).tolist()
+    try:
+        with out.open("w") as f:
+            while True:
+                ok, frame = cap.read()
+                if not ok:
+                    break
+                rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+                mp_img = mp.Image(image_format=mp.ImageFormat.SRGB, data=rgb)
+                ts_ms = int(n_frames * 1000 / fps)
+                try:
+                    res = landmarker.detect_for_video(mp_img, ts_ms)
+                except Exception:
+                    LOG.exception("detect failed at frame=%d", n_frames)
+                    n_frames += 1
+                    continue
+                body3d = _lmk_list_to_array(
+                    getattr(res, "pose_world_landmarks", None)
+                )
+                face_arr = _lmk_list_to_array(
+                    getattr(res, "face_landmarks", None)
+                )
+                left_arr = _lmk_list_to_array(
+                    getattr(res, "left_hand_landmarks", None)
+                )
+                right_arr = _lmk_list_to_array(
+                    getattr(res, "right_hand_landmarks", None)
+                )
+                hands_kp42 = _hands_kp42(left_arr, right_arr)
+                j3d32 = _build_j3d32(body3d, hands_kp42)
+                if j3d32 is None:
+                    n_frames += 1
+                    continue
+                ts = n_frames / fps
+                mouth = _mouth_open(face_arr)
+                f.write(json.dumps({
+                    "ts": ts,
+                    "session": session,
+                    "pid": 0,
+                    "j3d": j3d32.tolist(),
+                    "expression": expr_zeros_list,
+                    "mouth_open": mouth,
+                    "hands_kp": hands_kp42.tolist(),
+                }) + "\n")
+                n_rows += 1
+                n_frames += 1
+                if n_frames % 100 == 0:
+                    LOG.info("frame=%d rows=%d", n_frames, n_rows)
+    finally:
+        cap.release()
+        landmarker.close()
+    LOG.info("done : %d frames, %d rows -> %s", n_frames, n_rows, out)
+    return n_rows
+
+
+def _cli() -> None:
+    p = argparse.ArgumentParser()
+    p.add_argument("--session", required=True)
+    p.add_argument("--video", required=True, type=Path)
+    p.add_argument("--out", type=Path)
+    args = p.parse_args()
+    logging.basicConfig(level=logging.INFO,
+                        format="%(asctime)s [%(name)s] %(message)s")
+    DEFAULT_OUT_DIR.mkdir(parents=True, exist_ok=True)
+    out = args.out or (DEFAULT_OUT_DIR / f"{args.session}_mp.jsonl")
+    extract(args.session, args.video, out)
+
+
+if __name__ == "__main__":
+    _cli()
diff --git a/data_only_viz/scripts/setup_multihmr.sh b/data_only_viz/scripts/setup_multihmr.sh
index b51df54..fcf96cb 100755
--- a/data_only_viz/scripts/setup_multihmr.sh
+++ b/data_only_viz/scripts/setup_multihmr.sh
@@ -49,4 +49,71 @@ if [ ! -e "$CACHE/multi-hmr/models" ]; then
     ln -sfn ../models "$CACHE/multi-hmr/models"
 fi
 
+# CoreML conversion patches : remplace les torch.einsum dans utils/camera.py
+# par des ops element-wise (broadcast-friendly). Sans ca, ct.convert echoue
+# avec "Invalid target shape in reshape op ([1, N, 3] to [K*N, 3, 1])"
+# quand batch K detections != 1. Idempotent.
+CAM="$CACHE/multi-hmr/utils/camera.py"
+if [ -f "$CAM" ] && ! grep -q "_apply_intrinsics_componentwise" "$CAM"; then
+    echo "==> Patch utils/camera.py (einsum -> componentwise)"
+    python3 - "$CAM" <<'PYEOF'
+import sys, pathlib
+p = pathlib.Path(sys.argv[1])
+src = p.read_text()
+helper = '''
+def _apply_intrinsics_componentwise(K, y):
+    """CoreML-friendly: out[b,k,i] = sum_j K[b,i,j] * y[b,k,j]
+    Replaces torch.einsum('bij,bkj->bki', K, y) with pure broadcast ops.
+    """
+    K00 = K[:, 0:1, 0:1]; K01 = K[:, 0:1, 1:2]; K02 = K[:, 0:1, 2:3]
+    K10 = K[:, 1:2, 0:1]; K11 = K[:, 1:2, 1:2]; K12 = K[:, 1:2, 2:3]
+    K20 = K[:, 2:3, 0:1]; K21 = K[:, 2:3, 1:2]; K22 = K[:, 2:3, 2:3]
+    y0 = y[:, :, 0:1]; y1 = y[:, :, 1:2]; y2 = y[:, :, 2:3]
+    out0 = K00 * y0 + K01 * y1 + K02 * y2
+    out1 = K10 * y0 + K11 * y1 + K12 * y2
+    out2 = K20 * y0 + K21 * y1 + K22 * y2
+    return torch.cat([out0, out1, out2], dim=-1)
+
+
+'''
+src = src.replace(
+    "def perspective_projection(x, K):",
+    helper + "def perspective_projection(x, K):",
+)
+src = src.replace(
+    "y = torch.einsum('bij,bkj->bki', K, y)  # (bs, N, 3)",
+    "y = _apply_intrinsics_componentwise(K, y)",
+)
+src = src.replace(
+    "points = torch.einsum('bij,bkj->bki', torch.inverse(K), points)",
+    "points = _apply_intrinsics_componentwise(torch.inverse(K), points)",
+)
+p.write_text(src)
+print("  camera.py patched")
+PYEOF
+fi
+
+# CoreML conversion patch : smplx/lbs.py landmarks einsum (mеme bug broadcast)
+# Patch best-effort sur tous les venvs presents (data_only_viz + /tmp/coreml312).
+for VENV in \
+    "$(dirname "$(dirname "$(readlink -f "$0")")")/.venv" \
+    "/tmp/coreml312"; do
+    LBS="$VENV/lib/python3.14/site-packages/smplx/lbs.py"
+    [ -f "$LBS" ] || LBS="$VENV/lib/python3.12/site-packages/smplx/lbs.py"
+    if [ -f "$LBS" ] && grep -q "torch.einsum('blfi,blf->bli'" "$LBS"; then
+        echo "==> Patch $LBS (landmarks einsum)"
+        python3 - "$LBS" <<'PYEOF'
+import sys, pathlib
+p = pathlib.Path(sys.argv[1])
+s = p.read_text()
+s = s.replace(
+    "landmarks = torch.einsum('blfi,blf->bli', [lmk_vertices, lmk_bary_coords])\n    return landmarks",
+    "# CoreML-friendly: replace einsum('blfi,blf->bli', ...) with broadcast+sum\n    landmarks = (lmk_vertices * lmk_bary_coords.unsqueeze(-1)).sum(dim=2)\n    return landmarks",
+)
+p.write_text(s)
+print("  smplx/lbs.py patched")
+PYEOF
+    fi
+done
+
 echo "Setup OK. Cache : $CACHE"
diff --git a/data_only_viz/scripts/train_on_studio.sh b/data_only_viz/scripts/train_on_studio.sh
new file mode 100755
index 0000000..e4ec3fe
--- /dev/null
+++ b/data_only_viz/scripts/train_on_studio.sh
@@ -0,0 +1,111 @@
+#!/usr/bin/env bash
+# Train action-head on MacStudio M3 Ultra (Tailscale 100.116.92.12).
+#
+# SSH direct grosmac→studio is broken since reboot 2026-05-12 ;
+# we route via electron-server bastion (cf. CLAUDE.md root).
+#
+# Usage:
+#   ./train_on_studio.sh                    # uses defaults
+#   ./train_on_studio.sh --epochs 80 --lr 5e-4
+#
+# Local layout :
+#   ~/.cache/av-live-action/dataset/dataset.jsonl   (input)
+#   ~/.cache/av-live-action/checkpoints/            (output, after rsync back)
+#
+# Remote layout :
+#   studio:~/av-live-action/repo/                   (rsynced code subset)
+#   studio:~/av-live-action/dataset/                (rsynced dataset)
+#   studio:~/av-live-action/checkpoints/            (training output)
+
+set -euo pipefail
+
+BASTION_USER_HOST="${BASTION_USER_HOST:-electron-server}"
+STUDIO_USER_HOST="${STUDIO_USER_HOST:-clems@100.116.92.12}"
+STUDIO_USER="${STUDIO_USER:-clems}"
+STUDIO_UV="${STUDIO_UV:-/opt/homebrew/bin/uv}"
+
+REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")"/../.. && pwd)"
+LOCAL_CACHE="$HOME/.cache/av-live-action"
+LOCAL_DATASET="$LOCAL_CACHE/dataset"
+LOCAL_CKPT="$LOCAL_CACHE/checkpoints"
+
+REMOTE_ROOT="/Users/${STUDIO_USER}/av-live-action"
+REMOTE_REPO="$REMOTE_ROOT/repo"
+REMOTE_DATASET="$REMOTE_ROOT/dataset"
+REMOTE_CKPT="$REMOTE_ROOT/checkpoints"
+
+DATASET_FILE="${DATASET_FILE:-$LOCAL_DATASET/dataset.jsonl}"
+CKPT_NAME="${CKPT_NAME:-action_head.pt}"
+
+# Quote train args defensively before forwarding through bastion ssh +
+# studio ssh (each layer reparses). Reject single quotes — they break
+# the single-quoted payload in bastion_ssh and could allow injection.
+for a in "$@"; do
+  if [[ "$a" == *"'"* ]]; then
+    printf '[train_on_studio] forbidden single quote in arg: %s\n' "$a" >&2
+    exit 3
+  fi
+done
+TRAIN_ARGS="$(printf '%q ' "$@")"
+
+log() { printf '[train_on_studio] %s\n' "$*" >&2; }
+
+[[ -f "$DATASET_FILE" ]] || { log "missing dataset: $DATASET_FILE"; exit 2; }
+mkdir -p "$LOCAL_CKPT"
+
+bastion_ssh() {
+  # The remote shell on the bastion must receive the studio command
+  # as a single argument, otherwise `;` and `&&` are parsed
+  # bastion-side instead of studio-side.
+  # All paths in commands MUST be absolute (no $HOME, no ~) since
+  # we use single-quotes for the studio-side payload.
+  ssh -o ConnectTimeout=5 "$BASTION_USER_HOST" \
+      "ssh -o ConnectTimeout=5 $STUDIO_USER_HOST '$*'"
+}
+
+bastion_rsync() {
+  # rsync via ssh ProxyJump through bastion. Direct grosmac->studio
+  # known_hosts entry may be stale (SSH direct broken since reboot
+  # 2026-05-12). accept-new lets us add the key on first use.
+  local src="$1" dst="$2"
+  rsync -avz --delete \
+    -e "ssh -o ConnectTimeout=5 -o StrictHostKeyChecking=accept-new -A -J $BASTION_USER_HOST" \
+    "$src" "$dst"
+}
+
+log "== Studio reachability =="
+bastion_ssh "echo studio OK ; $STUDIO_UV --version"
+
+log "== Push code subset =="
+bastion_ssh "mkdir -p $REMOTE_REPO/data_only_viz $REMOTE_DATASET $REMOTE_CKPT"
+rsync -avz --delete \
+  --exclude='.venv/' --exclude='__pycache__/' --exclude='.pytest_cache/' \
+  --exclude='.ruff_cache/' --exclude='*.pyc' --exclude='.DS_Store' \
+  --exclude='web/' --exclude='shaders/' \
+  -e "ssh -o ConnectTimeout=5 -o StrictHostKeyChecking=accept-new -A -J $BASTION_USER_HOST" \
+  "$REPO_ROOT/data_only_viz/" \
+  "$STUDIO_USER_HOST:av-live-action/repo/data_only_viz/"
+
+log "== Push dataset =="
+bastion_rsync "$LOCAL_DATASET/" "$STUDIO_USER_HOST:av-live-action/dataset/"
+
+log "== Remote uv sync =="
+# multihmr extra pulls torch (action-head training needs torch but no pyobjc).
+# We piggy-back on the multihmr extras since torch is the main thing we need.
+bastion_ssh "cd $REMOTE_REPO && $STUDIO_UV sync --no-progress --project data_only_viz --extra multihmr"
+
+log "== Remote train (MPS) =="
+# cwd must be the PARENT of data_only_viz/ so the package is importable as
+# top-level. uv resolves the env via --project data_only_viz.
+bastion_ssh "cd $REMOTE_REPO && \
+             $STUDIO_UV run --project data_only_viz python -m data_only_viz.training.train_action_head \
+                 --dataset $REMOTE_DATASET/$(basename "$DATASET_FILE") \
+                 --ckpt-out $REMOTE_CKPT/$CKPT_NAME \
+                 --device mps \
+                 $TRAIN_ARGS"
+
+log "== Pull checkpoint back =="
+bastion_rsync "$STUDIO_USER_HOST:av-live-action/checkpoints/" "$LOCAL_CKPT/"
+
+log "== Done. Checkpoint: $LOCAL_CKPT/$CKPT_NAME =="
+ls -la "$LOCAL_CKPT/$CKPT_NAME"
diff --git a/data_only_viz/smplx_osc_sender.py b/data_only_viz/smplx_osc_sender.py
index aa7b379..18dbdd8 100644
--- a/data_only_viz/smplx_osc_sender.py
+++ b/data_only_viz/smplx_osc_sender.py
@@ -25,6 +25,7 @@ from typing import Sequence
 
 import numpy as np
 
+from .mesh_rigger import MeshRigger
 from .state import SMPLXPerson, State
 
 LOG = logging.getLogger("smplx_tcp")
@@ -35,7 +36,8 @@ PORT = 57130
 
 class SMPLXTCPSender:
     def __init__(self, state: State, host: str = "127.0.0.1",
-                 port: int = PORT, target_fps: float = 12.0) -> None:
+                 port: int = PORT, target_fps: float = 30.0,
+                 enable_rigging: bool = True) -> None:
         self.state = state
         self.host = host
         self.port = port
@@ -43,6 +45,9 @@ class SMPLXTCPSender:
         self._stop = threading.Event()
         self._thread: threading.Thread | None = None
         self._sock: socket.socket | None = None
+        # Hybrid keyframe rigging : entre deux keyframes Multi-HMR (~3 fps),
+        # on translate le mesh via le delta pelvis Apple Vision (30 fps).
+        self._rigger = MeshRigger(state) if enable_rigging else None
 
     def start(self) -> None:
         self._thread = threading.Thread(
@@ -124,6 +129,9 @@ class SMPLXTCPSender:
 
     def _run(self) -> None:
         last_warn = 0.0
+        n_sent = 0
+        n_rigged = 0
+        next_hb = time.monotonic() + 5.0
         while not self._stop.is_set():
             t0 = time.monotonic()
             if not self._ensure_connected():
@@ -136,8 +144,30 @@ class SMPLXTCPSender:
 
             with self.state.lock():
                 persons = list(self.state.persons_smplx)
+                body_kp = list(self.state.persons_body) if hasattr(
+                    self.state, "persons_body") else []
+                body_ids = list(self.state.persons_body_ids) if hasattr(
+                    self.state, "persons_body_ids") else (
+                    list(range(len(body_kp))) if body_kp else [])
+
+            if persons and self._rigger is not None:
+                rigged = self._rigger.apply(
+                    persons, body_kp, body_ids, t0)
+                if rigged is not persons:
+                    n_rigged += 1
+                persons = rigged
+
+            if t0 >= next_hb:
+                fps = n_sent / 5.0
+                rig_pct = (n_rigged / n_sent * 100.0) if n_sent else 0.0
+                LOG.info("hb: %.1f fps tcp, %.0f%% rigged",
+                         fps, rig_pct)
+                n_sent = 0
+                n_rigged = 0
+                next_hb = t0 + 5.0
 
             if persons:
+                n_sent += 1
                 t_ser_start = time.monotonic()
                 payload = self._serialize_persons(persons)
                 t_send_start = time.monotonic()
diff --git a/data_only_viz/tests/test_action_head_features.py b/data_only_viz/tests/test_action_head_features.py
new file mode 100644
index 0000000..b6df78d
--- /dev/null
+++ b/data_only_viz/tests/test_action_head_features.py
@@ -0,0 +1,116 @@
+"""Unit tests for ActionHead feature extraction and buffers."""
+from __future__ import annotations
+
+import numpy as np
+import pytest
+
+
+def test_module_imports() -> None:
+    from data_only_viz import action_head
+    assert hasattr(action_head, "FeatureExtractor")
+    assert hasattr(action_head, "PerPersonBuffer")
+    assert hasattr(action_head, "ActionHead")
+    assert action_head.WINDOW_LEN == 16
+    assert action_head.J3D_JOINTS == 32
+    assert action_head.FEATURE_DIM == 428
+    assert action_head.HANDS_KP_TOTAL == 42
+    assert action_head.HANDS_KP_FLAT == 126
+    assert action_head.NUM_CLASSES == 3
+    assert action_head.LABELS == ("debout", "assise", "danse")
+
+
+def _rand_j3d(seed: int = 0) -> np.ndarray:
+    rng = np.random.default_rng(seed)
+    return rng.normal(size=(32, 3)).astype(np.float32)
+
+
+def test_buffer_starts_empty() -> None:
+    from data_only_viz.action_head import PerPersonBuffer
+    buf = PerPersonBuffer()
+    assert len(buf) == 0
+    assert buf.frames_for(7) == []
+
+
+def test_buffer_append_grows_per_pid() -> None:
+    from data_only_viz.action_head import PerPersonBuffer
+    buf = PerPersonBuffer()
+    buf.append(pid=1, j3d=_rand_j3d(1))
+    buf.append(pid=1, j3d=_rand_j3d(2))
+    buf.append(pid=2, j3d=_rand_j3d(3))
+    assert len(buf.frames_for(1)) == 2
+    assert len(buf.frames_for(2)) == 1
+
+
+def test_buffer_max_len_16() -> None:
+    from data_only_viz.action_head import PerPersonBuffer, WINDOW_LEN
+    buf = PerPersonBuffer()
+    for i in range(WINDOW_LEN + 5):
+        buf.append(pid=1, j3d=_rand_j3d(i))
+    assert len(buf.frames_for(1)) == WINDOW_LEN
+
+
+def test_buffer_forget_releases_pid() -> None:
+    from data_only_viz.action_head import PerPersonBuffer
+    buf = PerPersonBuffer()
+    buf.append(pid=1, j3d=_rand_j3d(0))
+    buf.forget(1)
+    assert buf.frames_for(1) == []
+    assert len(buf) == 0
+
+
+def test_buffer_rejects_bad_shape() -> None:
+    from data_only_viz.action_head import PerPersonBuffer
+    buf = PerPersonBuffer()
+    with pytest.raises(ValueError, match="32"):
+        buf.append(pid=1, j3d=np.zeros((22, 3), dtype=np.float32))
+
+
+def test_feature_extractor_shape_full_buffer() -> None:
+    from data_only_viz.action_head import FeatureExtractor, WINDOW_LEN, FEATURE_DIM
+    frames = [_rand_j3d(i) for i in range(WINDOW_LEN)]
+    feat = FeatureExtractor.from_buffer(frames)
+    assert feat.shape == (428,)
+    assert feat.dtype == np.float32
+    assert not np.isnan(feat).any()
+
+
+def test_feature_extractor_short_buffer_pads() -> None:
+    from data_only_viz.action_head import FeatureExtractor, FEATURE_DIM
+    frames = [_rand_j3d(0), _rand_j3d(1), _rand_j3d(2)]
+    feat = FeatureExtractor.from_buffer(frames)
+    assert feat.shape == (FEATURE_DIM,)
+
+
+def test_feature_extractor_static_buffer_zero_velocity() -> None:
+    from data_only_viz.action_head import FeatureExtractor, WINDOW_LEN, J3D_JOINTS
+    static = _rand_j3d(42)
+    frames = [static.copy() for _ in range(WINDOW_LEN)]
+    feat = FeatureExtractor.from_buffer(frames)
+    vel_block = feat[J3D_JOINTS * 3 : J3D_JOINTS * 3 * 2]
+    assert np.allclose(vel_block, 0.0, atol=1e-6)
+
+
+def test_feature_extractor_kinetics_speed_and_accel() -> None:
+    from data_only_viz.action_head import FeatureExtractor, WINDOW_LEN
+    frames = []
+    for t in range(WINDOW_LEN):
+        f = np.zeros((32, 3), dtype=np.float32)
+        f[0, 0] = 0.1 * t
+        frames.append(f)
+    kin = FeatureExtractor.kinetics(frames)
+    assert kin.shape == (3,)
+    assert kin[0] > 0
+    assert abs(kin[0] - 0.1 / 32) < 1e-4
+    assert abs(kin[1]) < 1e-4
+
+
+def test_feature_extractor_symmetry_sign() -> None:
+    from data_only_viz.action_head import FeatureExtractor, WINDOW_LEN, WRIST_LEFT, WRIST_RIGHT
+    frames = []
+    for t in range(WINDOW_LEN):
+        f = np.zeros((32, 3), dtype=np.float32)
+        f[WRIST_LEFT, 0] = 0.05 * t
+        f[WRIST_RIGHT, 0] = -0.05 * t
+        frames.append(f)
+    kin = FeatureExtractor.kinetics(frames)
+    assert kin[2] > 0.9
diff --git a/data_only_viz/tests/test_action_head_model.py b/data_only_viz/tests/test_action_head_model.py
new file mode 100644
index 0000000..939d8c4
--- /dev/null
+++ b/data_only_viz/tests/test_action_head_model.py
@@ -0,0 +1,83 @@
+"""Tests for ActionHead model (forward, step, checkpoint roundtrip)."""
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+torch = pytest.importorskip("torch")
+
+
+def _rand_j3d(seed: int = 0) -> np.ndarray:
+    rng = np.random.default_rng(seed)
+    return rng.normal(size=(32, 3)).astype(np.float32)
+
+
+def test_model_forward_shape() -> None:
+    from data_only_viz.action_head import ActionHeadModel, FEATURE_DIM, NUM_CLASSES
+    model = ActionHeadModel()
+    x = torch.zeros(1, FEATURE_DIM)
+    h = model.init_hidden(batch=1)
+    logits, h_new = model(x, h)
+    assert logits.shape == (1, NUM_CLASSES)
+    assert h_new.shape == h.shape
+
+
+def test_model_param_count_under_100k() -> None:
+    from data_only_viz.action_head import ActionHeadModel
+    model = ActionHeadModel()
+    n = sum(p.numel() for p in model.parameters())
+    assert n < 100_000, f"too many params: {n}"
+
+
+def test_action_head_step_warmup_returns_debout() -> None:
+    from data_only_viz.action_head import ActionHead, LABELS
+    head = ActionHead(ckpt_path=None)
+    label, probs, kin = head.step(pid=1, j3d=_rand_j3d(0))
+    assert label == LABELS[0]
+    assert probs.shape == (3,)
+    assert pytest.approx(float(probs[0]), abs=1e-6) == 1.0
+    assert kin.shape == (3,)
+    assert float(kin[0]) == 0.0
+
+
+def test_action_head_step_after_warmup_returns_some_label() -> None:
+    from data_only_viz.action_head import ActionHead, LABELS
+    head = ActionHead(ckpt_path=None)
+    for i in range(5):
+        label, probs, kin = head.step(pid=1, j3d=_rand_j3d(i))
+    assert label in LABELS
+    assert abs(float(probs.sum()) - 1.0) < 1e-5
+
+
+def test_action_head_forget_resets_hidden_state(tmp_path: Path) -> None:
+    from data_only_viz.action_head import ActionHead
+    head = ActionHead(ckpt_path=None)
+    for i in range(5):
+        head.step(pid=1, j3d=_rand_j3d(i))
+    assert 1 in head._hidden
+    head.forget(1)
+    assert 1 not in head._hidden
+    assert head._buffers.frames_for(1) == []
+
+
+def test_action_head_checkpoint_roundtrip(tmp_path: Path) -> None:
+    from data_only_viz.action_head import ActionHead, ActionHeadModel
+    model = ActionHeadModel()
+    ckpt = tmp_path / "ah.pt"
+    torch.save({"model_state_dict": model.state_dict(),
+                "version": 1}, ckpt)
+    head = ActionHead(ckpt_path=ckpt)
+    for k, v in head._model.state_dict().items():
+        assert torch.allclose(v, model.state_dict()[k])
+
+
+def test_action_head_step_handles_nan() -> None:
+    from data_only_viz.action_head import ActionHead, LABELS
+    head = ActionHead(ckpt_path=None)
+    j = _rand_j3d(0)
+    j[5, 1] = float("nan")
+    label, probs, _kin = head.step(pid=1, j3d=j)
+    assert label in LABELS
+    assert not np.isnan(probs).any()
diff --git a/data_only_viz/tests/test_action_head_pub.py b/data_only_viz/tests/test_action_head_pub.py
new file mode 100644
index 0000000..540e5ad
--- /dev/null
+++ b/data_only_viz/tests/test_action_head_pub.py
@@ -0,0 +1,154 @@
+"""Tests for ActionHeadPublisher."""
+from __future__ import annotations
+
+import threading
+from unittest.mock import MagicMock
+
+import numpy as np
+import pytest
+
+torch = pytest.importorskip("torch")
+
+
+class _FakeState:
+    def __init__(self) -> None:
+        self.persons_smplx = []
+        self.smplx_last_t = 0.0
+        self.persons_body3d = []
+        self.persons_body_ids = []
+        self.pose_last_t = 0.0
+        self.persons_hands = []
+        self.persons_hands_ids = []
+        self.persons_face = []
+        self.persons_face_ids = []
+        self._lock = threading.RLock()
+
+    def lock(self):
+        return self._lock
+
+
+def _make_smplx_person(pid: int, seed: int = 0) -> dict:
+    rng = np.random.default_rng(seed)
+    return {"pid": pid, "v3d": rng.normal(size=(10475, 3)).astype(np.float32)}
+
+
+def test_publisher_smplx_source_emits_osc() -> None:
+    from data_only_viz.action_head_pub import ActionHeadPublisher
+    state = _FakeState()
+    bridge = MagicMock()
+    pub = ActionHeadPublisher(state, bridge, ckpt_path=None)
+    state.persons_smplx = [_make_smplx_person(7)]
+    state.smplx_last_t = 1.0
+    pub._tick(t_now=0.0)
+    actions = [c for c in bridge.send_action.call_args_list]
+    assert len(actions) == 1
+    assert actions[0].kwargs.get("pid", actions[0].args[0]) == 7
+    bridge.send_enter.assert_called_with(pid=7)
+
+
+def test_publisher_falls_back_to_mediapipe_body3d() -> None:
+    from data_only_viz.action_head_pub import ActionHeadPublisher
+    state = _FakeState()
+    bridge = MagicMock()
+    pub = ActionHeadPublisher(state, bridge, ckpt_path=None)
+    state.persons_body3d = [[(0.1 * i, 0.2 * i, 0.3 * i) for i in range(33)]]
+    state.persons_body_ids = [42]
+    state.pose_last_t = 1.0
+    pub._tick(t_now=0.0)
+    bridge.send_action.assert_called_once()
+    bridge.send_enter.assert_called_with(pid=42)
+
+
+def test_publisher_purges_lost_pid() -> None:
+    from data_only_viz.action_head_pub import ActionHeadPublisher
+    state = _FakeState()
+    bridge = MagicMock()
+    pub = ActionHeadPublisher(state, bridge, ckpt_path=None)
+    state.persons_smplx = [_make_smplx_person(1)]
+    state.smplx_last_t = 1.0
+    pub._tick(t_now=0.0)
+    bridge.reset_mock()
+    state.persons_smplx = []
+    state.smplx_last_t = 2.0
+    state.persons_body3d = []
+    pub._tick(t_now=1.0)
+    bridge.send_leave.assert_called_with(pid=1)
+
+
+def test_publisher_no_double_emit_same_timestamp() -> None:
+    from data_only_viz.action_head_pub import ActionHeadPublisher
+    state = _FakeState()
+    bridge = MagicMock()
+    pub = ActionHeadPublisher(state, bridge, ckpt_path=None)
+    state.persons_smplx = [_make_smplx_person(1)]
+    state.smplx_last_t = 1.0
+    pub._tick(t_now=0.0)
+    bridge.reset_mock()
+    pub._tick(t_now=1.0)  # same smplx_last_t
+    bridge.send_action.assert_not_called()
+
+
+def test_publisher_uses_face_lips_for_mouth_open() -> None:
+    """mouth_open from MediaPipe lip landmarks (idx 13 and 14) must be ~1.0."""
+    from unittest.mock import patch
+    from data_only_viz.action_head_pub import ActionHeadPublisher, MEDIAPIPE_LIP_UPPER_INNER, MEDIAPIPE_LIP_LOWER_INNER
+    state = _FakeState()
+    bridge = MagicMock()
+    pub = ActionHeadPublisher(state, bridge, ckpt_path=None)
+
+    # Build a fake face landmark list: at least 15 landmarks.
+    # idx 13 = upper inner (y=0), idx 14 = lower inner (y=1), rest zeros.
+    face_kps = [(0.0, 0.0, 0.0)] * 15
+    face_kps[MEDIAPIPE_LIP_UPPER_INNER] = (0.0, 0.0, 0.0)
+    face_kps[MEDIAPIPE_LIP_LOWER_INNER] = (1.0, 0.0, 0.0)  # 1m apart in x
+    state.persons_face = [face_kps]
+    state.persons_face_ids = [0]
+
+    captured_mouth: list[float] = []
+    original_step = pub.head.step
+
+    def spy_step(pid, j3d, expr=None, mouth_open=0.0, hands_kp=None):
+        captured_mouth.append(mouth_open)
+        return original_step(pid, j3d, expr=expr, mouth_open=mouth_open, hands_kp=hands_kp)
+
+    pub.head.step = spy_step  # type: ignore[method-assign]
+
+    state.persons_smplx = [_make_smplx_person(0)]
+    state.smplx_last_t = 1.0
+    pub._tick(t_now=0.0)
+
+    assert len(captured_mouth) == 1
+    assert abs(captured_mouth[0] - 1.0) < 1e-5
+
+
+def test_publisher_passes_hands_kp_to_step() -> None:
+    """hands_kp of shape (42, 3) must be passed to head.step."""
+    from data_only_viz.action_head_pub import ActionHeadPublisher
+    state = _FakeState()
+    bridge = MagicMock()
+    pub = ActionHeadPublisher(state, bridge, ckpt_path=None)
+
+    # Two 21-kp hand arrays (left + right) for pid=0.
+    rng = np.random.default_rng(7)
+    left_kps = rng.normal(size=(21, 3)).astype(np.float32)
+    right_kps = rng.normal(size=(21, 3)).astype(np.float32)
+    # persons_hands flat list: [left, right], ids both 0 (same pid).
+    state.persons_hands = [left_kps, right_kps]
+    state.persons_hands_ids = [0, 0]
+
+    captured_hands: list = []
+    original_step = pub.head.step
+
+    def spy_step(pid, j3d, expr=None, mouth_open=0.0, hands_kp=None):
+        captured_hands.append(hands_kp)
+        return original_step(pid, j3d, expr=expr, mouth_open=mouth_open, hands_kp=hands_kp)
+
+    pub.head.step = spy_step  # type: ignore[method-assign]
+
+    state.persons_smplx = [_make_smplx_person(0)]
+    state.smplx_last_t = 1.0
+    pub._tick(t_now=0.0)
+
+    assert len(captured_hands) == 1
+    assert captured_hands[0] is not None
+    assert captured_hands[0].shape == (42, 3)
diff --git a/data_only_viz/tests/test_augment.py b/data_only_viz/tests/test_augment.py
new file mode 100644
index 0000000..c6b9a5a
--- /dev/null
+++ b/data_only_viz/tests/test_augment.py
@@ -0,0 +1,46 @@
+"""Tests for j3d augmentations."""
+from __future__ import annotations
+
+import numpy as np
+
+WINDOW_LEN = 16
+
+
+def _sample_stack(seed: int = 0) -> np.ndarray:
+    rng = np.random.default_rng(seed)
+    return rng.normal(size=(WINDOW_LEN, 32, 3)).astype(np.float32)
+
+
+def test_mirror_swap_left_right_joints() -> None:
+    from data_only_viz.training.augment import mirror_x, MIRROR_MAP
+    x = _sample_stack(0)
+    y = mirror_x(x)
+    # Check output shape
+    assert y.shape == (WINDOW_LEN, 32, 3)
+    # x-coords are negated after reindexing
+    assert np.allclose(y[..., 0], -x[:, list(MIRROR_MAP), :][:, :, 0], atol=1e-6)
+
+
+def test_noise_within_sigma() -> None:
+    from data_only_viz.training.augment import add_noise
+    rng = np.random.default_rng(0)
+    x = _sample_stack(0)
+    y = add_noise(x, sigma=0.01, rng=rng)
+    diff = y - x
+    assert np.allclose(diff.std(), 0.01, atol=2e-3)
+
+
+def test_time_stretch_keeps_shape() -> None:
+    from data_only_viz.training.augment import time_stretch
+    x = _sample_stack(0)
+    y = time_stretch(x, factor=0.9, rng=None)
+    assert y.shape == x.shape
+
+
+def test_rotate_y_preserves_distances() -> None:
+    from data_only_viz.training.augment import rotate_y
+    x = _sample_stack(0)
+    y = rotate_y(x, angle_rad=0.3)
+    d_x = np.linalg.norm(x[0, 0] - x[0, 1])
+    d_y = np.linalg.norm(y[0, 0] - y[0, 1])
+    assert abs(d_x - d_y) < 1e-5
diff --git a/data_only_viz/tests/test_autolabel.py b/data_only_viz/tests/test_autolabel.py
new file mode 100644
index 0000000..a529b1e
--- /dev/null
+++ b/data_only_viz/tests/test_autolabel.py
@@ -0,0 +1,85 @@
+"""Tests for rule-based auto-labeler."""
+from __future__ import annotations
+
+import numpy as np
+
+from data_only_viz.action_head import WINDOW_LEN
+
+
+def _static_seated(frame_count: int = WINDOW_LEN) -> list[np.ndarray]:
+    """Hip low (y small), knee bent ~80°."""
+    frames = []
+    for _ in range(frame_count):
+        f = np.zeros((32, 3), dtype=np.float32)
+        f[1] = [-0.1, 0.4, 0.0]
+        f[2] = [0.1, 0.4, 0.0]
+        f[4] = [-0.1, 0.4, 0.3]
+        f[5] = [0.1, 0.4, 0.3]
+        f[7] = [-0.1, 0.1, 0.3]
+        f[8] = [0.1, 0.1, 0.3]
+        frames.append(f)
+    return frames
+
+
+def _static_standing(frame_count: int = WINDOW_LEN) -> list[np.ndarray]:
+    """Hip high, knees ~180°."""
+    frames = []
+    for _ in range(frame_count):
+        f = np.zeros((32, 3), dtype=np.float32)
+        f[1] = [-0.1, 0.9, 0.0]
+        f[2] = [0.1, 0.9, 0.0]
+        f[4] = [-0.1, 0.5, 0.0]
+        f[5] = [0.1, 0.5, 0.0]
+        f[7] = [-0.1, 0.1, 0.0]
+        f[8] = [0.1, 0.1, 0.0]
+        frames.append(f)
+    return frames
+
+
+def _dancing(frame_count: int = WINDOW_LEN) -> list[np.ndarray]:
+    """Standing pose with high wrist velocity."""
+    base = _static_standing(1)[0]
+    frames = []
+    for t in range(frame_count):
+        f = base.copy()
+        phase = 2 * np.pi * t * 0.125  # 0.125 = 1/8, slower oscillation
+        f[20] = base[20] + np.array([np.sin(phase) * 0.5, np.cos(phase) * 0.5, 0])
+        f[21] = base[21] + np.array(
+            [-np.sin(phase) * 0.5, np.cos(phase) * 0.5, 0]
+        )
+        frames.append(f.astype(np.float32))
+    return frames
+
+
+def test_autolabel_static_standing_is_debout() -> None:
+    from data_only_viz.training.autolabel import autolabel_window
+
+    label, conf = autolabel_window(_static_standing())
+    assert label == "debout"
+    assert conf >= 0.5
+
+
+def test_autolabel_static_seated_is_assise() -> None:
+    from data_only_viz.training.autolabel import autolabel_window
+
+    label, conf = autolabel_window(_static_seated())
+    assert label == "assise"
+    assert conf >= 0.5
+
+
+def test_autolabel_dancing_is_danse() -> None:
+    from data_only_viz.training.autolabel import autolabel_window
+
+    label, conf = autolabel_window(_dancing())
+    assert label == "danse"
+    assert conf >= 0.5
+
+
+def test_autolabel_ambiguous_is_none() -> None:
+    from data_only_viz.training.autolabel import autolabel_window
+
+    base = _static_standing(WINDOW_LEN)
+    for t, f in enumerate(base):
+        f[20, 0] += 0.01 * np.sin(t)
+    label, _conf = autolabel_window(base)
+    assert label in ("debout", None)
diff --git a/data_only_viz/tests/test_dataset.py b/data_only_viz/tests/test_dataset.py
index 679cf66..64f3e08 100644
--- a/data_only_viz/tests/test_dataset.py
+++ b/data_only_viz/tests/test_dataset.py
@@ -15,7 +15,7 @@ def _make_session_jsonl(path: Path, n_frames: int = 64) -> None:
             row = {"ts": t / 30.0,
                    "session": "sess01",
                    "pid": 1,
-                   "j3d": rng.normal(size=(22, 3)).tolist()}
+                   "j3d": rng.normal(size=(32, 3)).tolist()}
             f.write(json.dumps(row) + "\n")
 
 
@@ -25,7 +25,7 @@ def test_load_frames_jsonl(tmp_path: Path) -> None:
     _make_session_jsonl(p)
     frames = load_frames_jsonl(p)
     assert len(frames) == 64
-    assert frames[0].j3d.shape == (22, 3)
+    assert frames[0].j3d.shape == (32, 3)
     assert frames[0].pid == 1
     assert frames[0].session == "sess01"
 
@@ -40,7 +40,7 @@ def test_sliding_windows(tmp_path: Path) -> None:
     frames = load_frames_jsonl(p)
     windows = list(sliding_windows(frames, window_len=16, stride=4))
     assert len(windows) == 13
-    assert windows[0].j3d_stack.shape == (16, 22, 3)
+    assert windows[0].j3d_stack.shape == (16, 32, 3)
     assert windows[0].session == "sess01"
 
 
@@ -55,7 +55,7 @@ def test_write_and_load_dataset_jsonl(tmp_path: Path) -> None:
         DatasetRow(
             window_id=f"sess01_pid1_w{i:04d}",
             label="debout" if i % 2 == 0 else "danse",
-            j3d_stack=rng.normal(size=(16, 22, 3)).astype(np.float32),
+            j3d_stack=rng.normal(size=(16, 32, 3)).astype(np.float32),
             session="sess01",
             pid_local=1,
             auto_label_confidence=0.8,
@@ -68,10 +68,57 @@ def test_write_and_load_dataset_jsonl(tmp_path: Path) -> None:
     loaded = load_dataset_jsonl(out)
     assert len(loaded) == 5
     assert loaded[0].label == "debout"
-    assert loaded[0].j3d_stack.shape == (16, 22, 3)
+    assert loaded[0].j3d_stack.shape == (16, 32, 3)
     assert np.allclose(loaded[0].j3d_stack, rows[0].j3d_stack, atol=1e-6)
 
 
+def test_write_and_load_dataset_jsonl_with_hands_kp(tmp_path: Path) -> None:
+    from data_only_viz.training.dataset import (
+        DatasetRow,
+        load_dataset_jsonl,
+        write_dataset_jsonl,
+    )
+    rng = np.random.default_rng(1)
+    hands_kp = rng.normal(size=(16, 42, 3)).astype(np.float32)
+    row = DatasetRow(
+        window_id="sess01_pid1_w0000",
+        label="danse",
+        j3d_stack=rng.normal(size=(16, 32, 3)).astype(np.float32),
+        session="sess01",
+        pid_local=1,
+        auto_label_confidence=0.9,
+        manually_validated=True,
+        hands_kp_stack=hands_kp,
+    )
+    out = tmp_path / "with_hands.jsonl"
+    write_dataset_jsonl([row], out)
+    loaded = load_dataset_jsonl(out)
+    assert loaded[0].hands_kp_stack is not None
+    assert loaded[0].hands_kp_stack.shape == (16, 42, 3)
+    assert np.allclose(loaded[0].hands_kp_stack, hands_kp, atol=1e-6)
+
+
+def test_load_dataset_jsonl_without_hands_kp_is_ok(tmp_path: Path) -> None:
+    """Legacy v2 rows without hands_kp field should load with hands_kp_stack=None."""
+    import json
+    from data_only_viz.training.dataset import load_dataset_jsonl
+    rng = np.random.default_rng(2)
+    row = {
+        "window_id": "sess01_pid1_w0000",
+        "label": "debout",
+        "j3d": rng.normal(size=(16, 32, 3)).tolist(),
+        "session": "sess01",
+        "pid_local": 1,
+        "auto_label_confidence": 0.8,
+        "manually_validated": False,
+    }
+    out = tmp_path / "legacy.jsonl"
+    out.write_text(json.dumps(row) + "\n")
+    loaded = load_dataset_jsonl(out)
+    assert len(loaded) == 1
+    assert loaded[0].hands_kp_stack is None
+
+
 def test_split_by_session(tmp_path: Path) -> None:
     from data_only_viz.training.dataset import DatasetRow, split_by_session
     rng = np.random.default_rng(0)
@@ -79,7 +126,7 @@ def test_split_by_session(tmp_path: Path) -> None:
     for sess in ("s01", "s02", "s03", "s04", "s05", "s06", "s07"):
         rows.append(DatasetRow(
             window_id=f"{sess}_w0", label="debout",
-            j3d_stack=rng.normal(size=(16, 22, 3)).astype(np.float32),
+            j3d_stack=rng.normal(size=(16, 32, 3)).astype(np.float32),
             session=sess, pid_local=1, auto_label_confidence=0.7,
             manually_validated=False,
         ))
diff --git a/data_only_viz/tests/test_extract_mediapipe_offline.py b/data_only_viz/tests/test_extract_mediapipe_offline.py
new file mode 100644
index 0000000..0ba6e38
--- /dev/null
+++ b/data_only_viz/tests/test_extract_mediapipe_offline.py
@@ -0,0 +1,82 @@
+"""Sanity tests for MediaPipe offline extractor (no MediaPipe runtime -- we
+mock the landmarker and feed synthetic landmarks)."""
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+
+
+def test_build_j3d32_combines_body_and_fingertips() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _build_j3d32
+    from data_only_viz.action_head import J3D_JOINTS
+    body3d = np.linspace(0, 1, 33 * 3).reshape(33, 3).astype(np.float32)
+    hands_kp42 = np.linspace(2, 3, 42 * 3).reshape(42, 3).astype(np.float32)
+    j3d = _build_j3d32(body3d, hands_kp42)
+    assert j3d is not None
+    assert j3d.shape == (J3D_JOINTS, 3)
+    # The body22 portion comes from body3d via MEDIAPIPE_TO_22.
+    # The fingertip portion (indices 22..31) comes from hands_kp at idx 4,8,12,16,20.
+    assert np.allclose(j3d[22], hands_kp42[4])
+    assert np.allclose(j3d[26], hands_kp42[20])
+    assert np.allclose(j3d[27], hands_kp42[21 + 4])
+    assert np.allclose(j3d[31], hands_kp42[21 + 20])
+
+
+def test_build_j3d32_returns_none_when_no_body() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _build_j3d32
+    j3d = _build_j3d32(None, np.zeros((42, 3), dtype=np.float32))
+    assert j3d is None
+
+
+def test_hands_kp42_combines_left_right_sides() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _hands_kp42
+    left = np.linspace(0, 1, 21 * 3).reshape(21, 3).astype(np.float32)
+    right = np.linspace(2, 3, 21 * 3).reshape(21, 3).astype(np.float32)
+    out = _hands_kp42(left, right)
+    assert out.shape == (42, 3)
+    assert np.allclose(out[:21], left)
+    assert np.allclose(out[21:], right)
+
+
+def test_hands_kp42_zero_pads_when_missing() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _hands_kp42
+    left = np.ones((21, 3), dtype=np.float32)
+    out = _hands_kp42(left, None)
+    assert np.allclose(out[:21], left)
+    assert np.allclose(out[21:], 0.0)
+
+
+def test_mouth_open_from_face_lips() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _mouth_open
+    # MediaPipe FaceMesh has 478 landmarks. Build a sparse array : zero
+    # everywhere except idx 13 (upper inner) and idx 14 (lower inner),
+    # 1 metre apart on the y axis.
+    face = np.zeros((478, 3), dtype=np.float32)
+    face[13] = [0.0, 1.0, 0.0]
+    face[14] = [0.0, 0.0, 0.0]
+    assert abs(_mouth_open(face) - 1.0) < 1e-6
+
+
+def test_mouth_open_returns_zero_on_empty_face() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _mouth_open
+    assert _mouth_open(None) == 0.0
+    assert _mouth_open(np.zeros((10, 3), dtype=np.float32)) == 0.0
+
+
+def test_lmk_list_to_array_round_trip() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _lmk_list_to_array
+    class _Lmk:
+        def __init__(self, x: float, y: float, z: float) -> None:
+            self.x = x; self.y = y; self.z = z
+    lmks = [_Lmk(i, 2 * i, 3 * i) for i in range(5)]
+    arr = _lmk_list_to_array(lmks)
+    assert arr is not None
+    assert arr.shape == (5, 3)
+    assert np.allclose(arr[2], [2.0, 4.0, 6.0])
+
+
+def test_lmk_list_to_array_none_input() -> None:
+    from data_only_viz.scripts.extract_mediapipe_offline import _lmk_list_to_array
+    assert _lmk_list_to_array(None) is None
diff --git a/data_only_viz/tests/test_training_smoke.py b/data_only_viz/tests/test_training_smoke.py
new file mode 100644
index 0000000..7b9c74c
--- /dev/null
+++ b/data_only_viz/tests/test_training_smoke.py
@@ -0,0 +1,63 @@
+"""Smoke test for action-head training (2 epochs, tiny dataset, CPU)."""
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+torch = pytest.importorskip("torch")
+
+
+def _make_tiny_dataset(tmp_path: Path) -> Path:
+    from data_only_viz.training.dataset import DatasetRow, write_dataset_jsonl
+    rng = np.random.default_rng(0)
+    rows = []
+    for sess_i, sess in enumerate(("s01", "s02", "s03")):
+        for w in range(30):
+            label = ("debout", "assise", "danse")[w % 3]
+            rows.append(DatasetRow(
+                window_id=f"{sess}_w{w:03d}",
+                label=label,
+                j3d_stack=rng.normal(size=(16, 32, 3)).astype(np.float32),
+                session=sess, pid_local=1,
+                auto_label_confidence=0.8,
+                manually_validated=True,
+                expr_stack=np.zeros((16, 10), dtype=np.float32),
+                mouth_open_stack=np.zeros(16, dtype=np.float32),
+            ))
+    out = tmp_path / "tiny.jsonl"
+    write_dataset_jsonl(rows, out)
+    return out
+
+
+def test_train_2_epochs_no_crash(tmp_path: Path) -> None:
+    from data_only_viz.training.train_action_head import train
+    ds = _make_tiny_dataset(tmp_path)
+    ckpt = tmp_path / "ckpt.pt"
+    history = train(
+        dataset_path=ds,
+        ckpt_out=ckpt,
+        epochs=2,
+        batch_size=8,
+        lr=1e-3,
+        device="cpu",
+        seed=0,
+        log_every=10_000,
+    )
+    assert ckpt.exists()
+    assert len(history["train_loss"]) == 2
+    assert all(np.isfinite(history["train_loss"]))
+
+
+def test_trained_checkpoint_loadable(tmp_path: Path) -> None:
+    from data_only_viz.action_head import ActionHead
+    from data_only_viz.training.train_action_head import train
+    ds = _make_tiny_dataset(tmp_path)
+    ckpt = tmp_path / "ckpt.pt"
+    train(dataset_path=ds, ckpt_out=ckpt, epochs=1, batch_size=8,
+          lr=1e-3, device="cpu", seed=0, log_every=10_000)
+    head = ActionHead(ckpt_path=ckpt)
+    for i in range(5):
+        label, probs, _ = head.step(pid=1, j3d=np.zeros((32, 3), dtype=np.float32))
+    assert abs(float(probs.sum()) - 1.0) < 1e-5
diff --git a/data_only_viz/training/augment.py b/data_only_viz/training/augment.py
new file mode 100644
index 0000000..0c3cff0
--- /dev/null
+++ b/data_only_viz/training/augment.py
@@ -0,0 +1,85 @@
+"""On-the-fly augmentations for j3d windows."""
+from __future__ import annotations
+
+import numpy as np
+
+# SMPL-X left/right joint mirror map for 32-joint layout.
+# Body joints 0..21 (unchanged), fingertips 22..31 (L 22..26 <-> R 27..31).
+MIRROR_MAP: tuple[int, ...] = (
+    # 22 body (unchanged)
+    0,
+    2, 1,
+    3,
+    5, 4,
+    6,
+    8, 7,
+    9,
+    11, 10,
+    12,
+    14, 13,
+    15,
+    17, 16,
+    19, 18,
+    21, 20,
+    # 10 fingertips: L (22..26) <-> R (27..31)
+    27, 28, 29, 30, 31, 22, 23, 24, 25, 26,
+)
+assert len(MIRROR_MAP) == 32
+
+
+def mirror_x(stack: np.ndarray) -> np.ndarray:
+    """Mirror across the YZ plane: flip x and swap left↔right joints."""
+    out = stack[:, list(MIRROR_MAP), :].copy()
+    out[..., 0] = -out[..., 0]
+    return out
+
+
+def add_noise(stack: np.ndarray, sigma: float, rng: np.random.Generator) -> np.ndarray:
+    noise = rng.normal(scale=sigma, size=stack.shape).astype(np.float32)
+    return (stack + noise).astype(np.float32, copy=False)
+
+
+def time_stretch(stack: np.ndarray, factor: float,
+                 rng: np.random.Generator | None = None) -> np.ndarray:
+    """Resample the time axis with linear interpolation, keep window_len fixed."""
+    T = stack.shape[0]
+    new_T = int(round(T * factor))
+    new_T = max(2, new_T)
+    src = np.linspace(0.0, T - 1, num=new_T)
+    interp = np.empty((new_T, *stack.shape[1:]), dtype=np.float32)
+    lo = np.floor(src).astype(int)
+    hi = np.minimum(lo + 1, T - 1)
+    frac = (src - lo).astype(np.float32)
+    interp = (1 - frac[:, None, None]) * stack[lo] + frac[:, None, None] * stack[hi]
+    if new_T >= T:
+        start = (new_T - T) // 2
+        return interp[start:start + T].astype(np.float32, copy=False)
+    pad_before = (T - new_T) // 2
+    pad_after = T - new_T - pad_before
+    return np.concatenate([
+        np.repeat(interp[:1], pad_before, axis=0),
+        interp,
+        np.repeat(interp[-1:], pad_after, axis=0),
+    ]).astype(np.float32, copy=False)
+
+
+def rotate_y(stack: np.ndarray, angle_rad: float) -> np.ndarray:
+    """Rotate around Y (vertical) axis."""
+    c, s = np.cos(angle_rad), np.sin(angle_rad)
+    R = np.array([[c, 0, s], [0, 1, 0], [-s, 0, c]], dtype=np.float32)
+    return (stack @ R.T).astype(np.float32, copy=False)
+
+
+def random_augment(stack: np.ndarray, rng: np.random.Generator) -> np.ndarray:
+    out = stack
+    if rng.random() < 0.5:
+        out = mirror_x(out)
+    if rng.random() < 0.8:
+        out = add_noise(out, sigma=0.01, rng=rng)
+    if rng.random() < 0.5:
+        factor = float(rng.uniform(0.9, 1.1))
+        out = time_stretch(out, factor=factor, rng=rng)
+    if rng.random() < 0.5:
+        angle = float(rng.uniform(-np.deg2rad(15), np.deg2rad(15)))
+        out = rotate_y(out, angle_rad=angle)
+    return out
diff --git a/data_only_viz/training/autolabel.py b/data_only_viz/training/autolabel.py
new file mode 100644
index 0000000..bbf63de
--- /dev/null
+++ b/data_only_viz/training/autolabel.py
@@ -0,0 +1,120 @@
+"""Rule-based labeler for j3d windows.
+
+Outputs one of {"debout", "assise", "danse", None}. None marks
+ambiguous windows that should be reviewed manually.
+
+Rules are tuned for SMPL-X joint indexing as used by Multi-HMR.
+"""
+from __future__ import annotations
+
+import argparse
+import logging
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+
+from data_only_viz.action_head import (
+    FeatureExtractor,
+    HIP_LEFT,
+    HIP_RIGHT,
+    WINDOW_LEN,
+)
+
+
+@dataclass(frozen=True)
+class AutoLabelConfig:
+    hip_y_seated_max: float = 0.55
+    knee_angle_seated_max: float = 2.0  # rad, ~115°
+    speed_static_max: float = 0.03  # m/s mean joint speed
+    speed_dance_min: float = 0.033  # m/s mean joint speed
+    accel_dance_min: float = 0.001
+
+
+DEFAULT_CFG = AutoLabelConfig()
+
+
+def autolabel_window(
+    frames: list[np.ndarray], cfg: AutoLabelConfig = DEFAULT_CFG
+) -> tuple[str | None, float]:
+    """Return (label, confidence). label is None when ambiguous."""
+    if len(frames) < WINDOW_LEN // 2:
+        return None, 0.0
+    cur = frames[-1]
+    hip_y = float((cur[HIP_LEFT, 1] + cur[HIP_RIGHT, 1]) * 0.5)
+    knee_angle = FeatureExtractor._mean_knee_angle(cur)
+    kin = FeatureExtractor.kinetics(frames)
+    speed = float(kin[0])
+    accel = float(kin[1])
+
+    if hip_y < cfg.hip_y_seated_max and knee_angle < cfg.knee_angle_seated_max:
+        conf = 0.5 + 0.5 * min(1.0, (cfg.hip_y_seated_max - hip_y) / 0.2)
+        return "assise", conf
+    if speed >= cfg.speed_dance_min or accel >= cfg.accel_dance_min:
+        conf = 0.5 + 0.5 * min(1.0, speed / 0.5)
+        return "danse", conf
+    if speed <= cfg.speed_static_max:
+        conf = 0.6
+        return "debout", conf
+    return None, 0.0
+
+
+def autolabel_dataset(
+    frames_jsonl: Path,
+    out_jsonl: Path,
+    window_len: int = WINDOW_LEN,
+    stride: int = 4,
+    keep_none: bool = True,
+) -> int:
+    """Glue: raw frames jsonl → sliding windows → auto-label → DatasetRow jsonl.
+
+    Returns the number of windows written.
+    """
+    from data_only_viz.training.dataset import (
+        DatasetRow,
+        load_frames_jsonl,
+        sliding_windows,
+        write_dataset_jsonl,
+    )
+
+    frames = load_frames_jsonl(frames_jsonl)
+    rows = []
+    for win in sliding_windows(frames, window_len=window_len, stride=stride):
+        frame_list = [win.j3d_stack[t] for t in range(win.j3d_stack.shape[0])]
+        label, conf = autolabel_window(frame_list)
+        if label is None and not keep_none:
+            continue
+        rows.append(
+            DatasetRow(
+                window_id=f"{win.session}_pid{win.pid_local}_t{int(win.first_ts*1000):08d}",
+                label=label if label is not None else "debout",
+                j3d_stack=win.j3d_stack,
+                session=win.session,
+                pid_local=win.pid_local,
+                auto_label_confidence=conf,
+                manually_validated=False,
+            )
+        )
+    out_jsonl.parent.mkdir(parents=True, exist_ok=True)
+    write_dataset_jsonl(rows, out_jsonl)
+    return len(rows)
+
+
+def _cli() -> None:
+    p = argparse.ArgumentParser()
+    p.add_argument(
+        "--frames",
+        required=True,
+        type=Path,
+        help="Raw frames jsonl from extract_j3d_offline.py",
+    )
+    p.add_argument("--out", required=True, type=Path, help="Auto-labeled windowed dataset jsonl")
+    p.add_argument("--stride", type=int, default=4)
+    args = p.parse_args()
+    logging.basicConfig(level=logging.INFO)
+    n = autolabel_dataset(args.frames, args.out, stride=args.stride)
+    print(f"wrote {n} windows to {args.out}")
+
+
+if __name__ == "__main__":
+    _cli()
diff --git a/data_only_viz/training/dataset.py b/data_only_viz/training/dataset.py
index 60cd2cd..27f2ecd 100644
--- a/data_only_viz/training/dataset.py
+++ b/data_only_viz/training/dataset.py
@@ -3,7 +3,7 @@ from __future__ import annotations
 
 import json
 import random
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Iterable, Iterator
 
@@ -15,26 +15,35 @@ class RawFrame:
     ts: float
     session: str
     pid: int
-    j3d: np.ndarray  # (22, 3) float32
+    j3d: np.ndarray          # (32, 3) float32 (v3: body22 + 10 fingertips)
+    expression: np.ndarray | None = None  # (EXPR_DIM,) or None
+    mouth_open: float = 0.0
+    hands_kp: np.ndarray | None = None  # (42, 3) or None
 
 
 @dataclass
 class WindowRow:
-    j3d_stack: np.ndarray  # (window_len, 22, 3) float32
+    j3d_stack: np.ndarray          # (window_len, 32, 3) float32
     session: str
     pid_local: int
     first_ts: float
+    expr_stack: np.ndarray | None = None   # (window_len, 10) or None
+    mouth_open_stack: np.ndarray | None = None  # (window_len,) or None
+    hands_kp_stack: np.ndarray | None = None  # (window_len, 42, 3) or None
 
 
 @dataclass
 class DatasetRow:
     window_id: str
     label: str
-    j3d_stack: np.ndarray  # (window_len, 22, 3) float32
+    j3d_stack: np.ndarray          # (window_len, 32, 3) float32
     session: str
     pid_local: int
     auto_label_confidence: float
     manually_validated: bool
+    expr_stack: np.ndarray | None = None   # (window_len, 10) or None
+    mouth_open_stack: np.ndarray | None = None  # (window_len,) or None
+    hands_kp_stack: np.ndarray | None = None  # (window_len, 42, 3) or None
 
 
 def load_frames_jsonl(path: Path) -> list[RawFrame]:
@@ -45,11 +54,18 @@ def load_frames_jsonl(path: Path) -> list[RawFrame]:
             if not line:
                 continue
             d = json.loads(line)
+            expr_raw = d.get("expression")
+            expr = np.asarray(expr_raw, dtype=np.float32) if expr_raw is not None else None
+            hands_raw = d.get("hands_kp")
+            hands_kp = np.asarray(hands_raw, dtype=np.float32) if hands_raw is not None else None
             rows.append(RawFrame(
                 ts=float(d["ts"]),
                 session=str(d["session"]),
                 pid=int(d["pid"]),
                 j3d=np.asarray(d["j3d"], dtype=np.float32),
+                expression=expr,
+                mouth_open=float(d.get("mouth_open", 0.0)),
+                hands_kp=hands_kp,
             ))
     return rows
 
@@ -68,14 +84,43 @@ def sliding_windows(frames: list[RawFrame],
         for start in range(0, len(grp) - window_len + 1, stride):
             chunk = grp[start:start + window_len]
             stack = np.stack([c.j3d for c in chunk]).astype(np.float32)
+            # Expression stack: zeros if not present
+            if any(c.expression is not None for c in chunk):
+                expr_dim = max(
+                    (len(c.expression) for c in chunk if c.expression is not None),
+                    default=10,
+                )
+                expr_stack = np.zeros((window_len, expr_dim), dtype=np.float32)
+                for t, c in enumerate(chunk):
+                    if c.expression is not None:
+                        n = min(expr_dim, len(c.expression))
+                        expr_stack[t, :n] = c.expression[:n]
+            else:
+                expr_stack = None
+            mouth_stack = np.array(
+                [c.mouth_open for c in chunk], dtype=np.float32
+            )
+            # hands_kp stack: (window_len, 42, 3) if any frame has hands_kp
+            if any(c.hands_kp is not None for c in chunk):
+                hands_kp_stack = np.zeros((window_len, 42, 3), dtype=np.float32)
+                for t, c in enumerate(chunk):
+                    if c.hands_kp is not None:
+                        hk = np.asarray(c.hands_kp, dtype=np.float32)
+                        if hk.shape == (42, 3):
+                            hands_kp_stack[t] = hk
+            else:
+                hands_kp_stack = None
             yield WindowRow(j3d_stack=stack, session=sess,
-                            pid_local=pid, first_ts=chunk[0].ts)
+                            pid_local=pid, first_ts=chunk[0].ts,
+                            expr_stack=expr_stack,
+                            mouth_open_stack=mouth_stack,
+                            hands_kp_stack=hands_kp_stack)
 
 
 def write_dataset_jsonl(rows: Iterable[DatasetRow], path: Path) -> None:
     with path.open("w") as f:
         for r in rows:
-            f.write(json.dumps({
+            d: dict = {
                 "window_id": r.window_id,
                 "label": r.label,
                 "j3d": r.j3d_stack.astype(np.float32).tolist(),
@@ -83,7 +128,14 @@ def write_dataset_jsonl(rows: Iterable[DatasetRow], path: Path) -> None:
                 "pid_local": r.pid_local,
                 "auto_label_confidence": float(r.auto_label_confidence),
                 "manually_validated": bool(r.manually_validated),
-            }) + "\n")
+            }
+            if r.expr_stack is not None:
+                d["expr_stack"] = r.expr_stack.astype(np.float32).tolist()
+            if r.mouth_open_stack is not None:
+                d["mouth_open_stack"] = r.mouth_open_stack.astype(np.float32).tolist()
+            if r.hands_kp_stack is not None:
+                d["hands_kp_stack"] = r.hands_kp_stack.astype(np.float32).tolist()
+            f.write(json.dumps(d) + "\n")
 
 
 def load_dataset_jsonl(path: Path) -> list[DatasetRow]:
@@ -94,6 +146,12 @@ def load_dataset_jsonl(path: Path) -> list[DatasetRow]:
             if not line:
                 continue
             d = json.loads(line)
+            expr_raw = d.get("expr_stack")
+            expr = np.asarray(expr_raw, dtype=np.float32) if expr_raw is not None else None
+            mouth_raw = d.get("mouth_open_stack")
+            mouth = np.asarray(mouth_raw, dtype=np.float32) if mouth_raw is not None else None
+            hands_raw = d.get("hands_kp_stack")
+            hands_kp = np.asarray(hands_raw, dtype=np.float32) if hands_raw is not None else None
             out.append(DatasetRow(
                 window_id=d["window_id"],
                 label=d["label"],
@@ -102,6 +160,9 @@ def load_dataset_jsonl(path: Path) -> list[DatasetRow]:
                 pid_local=int(d["pid_local"]),
                 auto_label_confidence=float(d["auto_label_confidence"]),
                 manually_validated=bool(d["manually_validated"]),
+                expr_stack=expr,
+                mouth_open_stack=mouth,
+                hands_kp_stack=hands_kp,
             ))
     return out
 
diff --git a/data_only_viz/training/eval.py b/data_only_viz/training/eval.py
new file mode 100644
index 0000000..c711187
--- /dev/null
+++ b/data_only_viz/training/eval.py
@@ -0,0 +1,87 @@
+"""Evaluate a trained action-head checkpoint."""
+from __future__ import annotations
+
+import argparse
+import json
+import time
+from pathlib import Path
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader
+
+from data_only_viz.action_head import ActionHeadModel, LABELS
+from data_only_viz.training.dataset import load_dataset_jsonl, split_by_session
+from data_only_viz.training.train_action_head import (
+    LABEL_TO_IDX,
+    WindowDataset,
+)
+
+
+def confusion_matrix(true: list[int], pred: list[int],
+                     num_classes: int = 3) -> np.ndarray:
+    cm = np.zeros((num_classes, num_classes), dtype=np.int64)
+    for t, p in zip(true, pred):
+        cm[t, p] += 1
+    return cm
+
+
+def evaluate(ckpt_path: Path, dataset_path: Path, device: str = "cpu",
+             seed: int = 0) -> dict:
+    rows = load_dataset_jsonl(dataset_path)
+    _train, _val, test_rows = split_by_session(rows, seed=seed)
+    if not test_rows:
+        test_rows = rows
+    ds = WindowDataset(test_rows, augment=False, seed=seed)
+    loader = DataLoader(ds, batch_size=64, shuffle=False)
+    model = ActionHeadModel().to(device).eval()
+    payload = torch.load(ckpt_path, map_location=device, weights_only=True)
+    model.load_state_dict(payload["model_state_dict"])
+    true: list[int] = []
+    pred: list[int] = []
+    with torch.no_grad():
+        for x, y in loader:
+            x = x.to(device); y = y.to(device)
+            B, T, _ = x.shape
+            h = model.init_hidden(batch=B, device=device)
+            logits = None
+            for t in range(T):
+                logits, h = model(x[:, t, :], h)
+            true.extend(y.cpu().tolist())
+            pred.extend(logits.argmax(-1).cpu().tolist())
+    cm = confusion_matrix(true, pred)
+    acc = float(np.trace(cm) / max(1, cm.sum()))
+    confusion_db = float((cm[0, 2] + cm[2, 0]) / max(1, cm.sum()))
+    feat_dim = ds[0][0].shape[-1]
+    bench_x = torch.zeros(1, feat_dim, device=device)
+    h = model.init_hidden(batch=1, device=device)
+    for _ in range(20):
+        _ = model(bench_x, h)
+    t0 = time.perf_counter()
+    N = 500
+    for _ in range(N):
+        _, h = model(bench_x, h)
+    lat_ms = (time.perf_counter() - t0) * 1000.0 / N
+    return {
+        "test_acc": acc,
+        "confusion_debout_danse": confusion_db,
+        "confusion_matrix": cm.tolist(),
+        "labels": list(LABELS),
+        "step_latency_ms": lat_ms,
+        "n_test": int(cm.sum()),
+    }
+
+
+def _cli() -> None:
+    p = argparse.ArgumentParser()
+    p.add_argument("--ckpt", required=True, type=Path)
+    p.add_argument("--dataset", required=True, type=Path)
+    p.add_argument("--device", default="cpu",
+                   choices=["cpu", "mps", "cuda"])
+    args = p.parse_args()
+    out = evaluate(args.ckpt, args.dataset, device=args.device)
+    print(json.dumps(out, indent=2))
+
+
+if __name__ == "__main__":
+    _cli()
diff --git a/data_only_viz/training/review.py b/data_only_viz/training/review.py
new file mode 100644
index 0000000..2cbb848
--- /dev/null
+++ b/data_only_viz/training/review.py
@@ -0,0 +1,116 @@
+"""Manual label review TUI.
+
+Reads an auto-labeled jsonl dataset, presents each window with:
+  - ASCII skeleton (front view) of last frame
+  - speed/accel/sym kinetics
+  - proposed label + confidence
+Keys:
+  1 = debout, 2 = assise, 3 = danse
+  ENTER = accept proposed label
+  S = skip (label = None, will not be saved)
+  Q = quit and write what we have so far
+
+Usage:
+    uv run python -m data_only_viz.training.review \\
+        --in ~/.cache/av-live-action/dataset/auto.jsonl \\
+        --out ~/.cache/av-live-action/dataset/reviewed.jsonl
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+import termios
+import tty
+from pathlib import Path
+
+import numpy as np
+
+from data_only_viz.action_head import LABELS
+from data_only_viz.training.autolabel import autolabel_window
+from data_only_viz.training.dataset import (
+    DatasetRow,
+    load_dataset_jsonl,
+    write_dataset_jsonl,
+)
+
+
+def _ascii_skeleton(j3d: np.ndarray, width: int = 40, height: int = 16) -> str:
+    pts = j3d[:, [0, 1]]  # x, y
+    mn = pts.min(axis=0)
+    mx = pts.max(axis=0)
+    rng = np.maximum(mx - mn, 1e-3)
+    norm = (pts - mn) / rng
+    grid = [[" "] * width for _ in range(height)]
+    for x, y in norm:
+        col = int(x * (width - 1))
+        row = int((1 - y) * (height - 1))
+        grid[row][col] = "*"
+    return "\n".join("".join(row) for row in grid)
+
+
+def _getch() -> str:
+    fd = sys.stdin.fileno()
+    old = termios.tcgetattr(fd)
+    try:
+        tty.setraw(fd)
+        return sys.stdin.read(1)
+    finally:
+        termios.tcsetattr(fd, termios.TCSADRAIN, old)
+
+
+def review(in_path: Path, out_path: Path,
+           sample_validated_fraction: float = 0.2,
+           seed: int = 0) -> None:
+    from data_only_viz.action_head import FeatureExtractor
+
+    rows = load_dataset_jsonl(in_path)
+    rng = np.random.default_rng(seed)
+    kept: list[DatasetRow] = []
+    for i, r in enumerate(rows):
+        proposed, conf = autolabel_window(list(r.j3d_stack))
+        is_none = proposed is None
+        sampled = rng.random() < sample_validated_fraction
+        if not is_none and not sampled and r.manually_validated:
+            kept.append(r)
+            continue
+        print("\033[2J\033[H")  # clear
+        print(f"[{i + 1}/{len(rows)}] {r.window_id}  proposed={proposed} conf={conf:.2f}")
+        print(_ascii_skeleton(r.j3d_stack[-1]))
+        kin = FeatureExtractor.kinetics(list(r.j3d_stack))
+        print(f"speed={kin[0]:.3f}  accel={kin[1]:.3f}  sym={kin[2]:+.3f}")
+        print("keys: 1=debout 2=assise 3=danse ENTER=accept S=skip Q=quit")
+        k = _getch().lower()
+        if k == "q":
+            break
+        if k == "s":
+            continue
+        if k == "\r":
+            chosen = proposed
+        elif k in ("1", "2", "3"):
+            chosen = LABELS[int(k) - 1]
+        else:
+            continue
+        if chosen is None:
+            continue
+        kept.append(DatasetRow(
+            window_id=r.window_id, label=chosen, j3d_stack=r.j3d_stack,
+            session=r.session, pid_local=r.pid_local,
+            auto_label_confidence=conf, manually_validated=True,
+        ))
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    write_dataset_jsonl(kept, out_path)
+    print(f"\nwrote {len(kept)} rows to {out_path}")
+
+
+def _cli() -> None:
+    p = argparse.ArgumentParser()
+    p.add_argument("--in", dest="in_path", required=True, type=Path)
+    p.add_argument("--out", dest="out_path", required=True, type=Path)
+    p.add_argument("--sample-fraction", type=float, default=0.2)
+    args = p.parse_args()
+    review(args.in_path, args.out_path,
+           sample_validated_fraction=args.sample_fraction)
+
+
+if __name__ == "__main__":
+    _cli()
diff --git a/data_only_viz/training/train_action_head.py b/data_only_viz/training/train_action_head.py
new file mode 100644
index 0000000..fd4ea53
--- /dev/null
+++ b/data_only_viz/training/train_action_head.py
@@ -0,0 +1,205 @@
+"""Train ActionHead on the windowed dataset.
+
+Usage:
+    uv run python -m data_only_viz.training.train_action_head \
+        --dataset ~/.cache/av-live-action/dataset/dataset.jsonl \
+        --ckpt-out ~/.cache/av-live-action/checkpoints/action_head.pt \
+        --device mps --epochs 50 --batch-size 128
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import logging
+from collections import Counter
+from pathlib import Path
+
+import numpy as np
+import torch
+from torch import nn
+from torch.utils.data import DataLoader, Dataset
+
+from data_only_viz.action_head import (
+    ActionHeadModel,
+    EXPR_DIM,
+    FeatureExtractor,
+    HANDS_KP_FLAT,
+    HIP_LEFT,
+    HIP_RIGHT,
+    LABELS,
+)
+from data_only_viz.training.augment import random_augment
+from data_only_viz.training.dataset import (
+    DatasetRow,
+    load_dataset_jsonl,
+    split_by_session,
+)
+
+LOG = logging.getLogger("train_action_head")
+LABEL_TO_IDX = {l: i for i, l in enumerate(LABELS)}
+
+
+class WindowDataset(Dataset[tuple[torch.Tensor, int]]):
+    def __init__(self, rows: list[DatasetRow],
+                 augment: bool = False, seed: int = 0) -> None:
+        self._rows = rows
+        self._augment = augment
+        self._rng = np.random.default_rng(seed)
+
+    def __len__(self) -> int:
+        return len(self._rows)
+
+    def __getitem__(self, idx: int) -> tuple[torch.Tensor, int]:
+        row = self._rows[idx]
+        stack = row.j3d_stack
+        if self._augment:
+            stack = random_augment(stack, self._rng)
+        T = stack.shape[0]
+        # expression and mouth_open stacks (zeros if absent / legacy)
+        if row.expr_stack is not None:
+            expr_s = row.expr_stack.astype(np.float32)
+        else:
+            expr_s = np.zeros((T, EXPR_DIM), dtype=np.float32)
+        if row.mouth_open_stack is not None:
+            mouth_s = row.mouth_open_stack.astype(np.float32)
+        else:
+            mouth_s = np.zeros(T, dtype=np.float32)
+        feats = []
+        prev = stack[0]
+        prev_vel = np.zeros_like(prev)
+        for t in range(T):
+            cur = stack[t]
+            vel = cur - prev
+            accel = vel - prev_vel
+            hip_y = float((cur[HIP_LEFT, 1] + cur[HIP_RIGHT, 1]) * 0.5)
+            knee_angle = FeatureExtractor._mean_knee_angle(cur)
+            sym = FeatureExtractor._symmetry_score(vel)
+            expr_t = expr_s[t] if t < len(expr_s) else np.zeros(EXPR_DIM, dtype=np.float32)
+            expr_vec = np.zeros(EXPR_DIM, dtype=np.float32)
+            n = min(EXPR_DIM, len(expr_t))
+            expr_vec[:n] = expr_t[:n]
+            mouth_t = float(mouth_s[t]) if t < len(mouth_s) else 0.0
+            # hands_kp at frame t (42, 3); zeros if row has none
+            if row.hands_kp_stack is not None:
+                hands_t = row.hands_kp_stack[t]
+                hands_flat = hands_t.reshape(-1).astype(np.float32, copy=False)
+            else:
+                hands_flat = np.zeros(HANDS_KP_FLAT, dtype=np.float32)
+            feat = np.concatenate([
+                cur.reshape(-1), vel.reshape(-1), accel.reshape(-1),
+                hands_flat,
+                expr_vec,
+                np.array([hip_y, knee_angle, sym, mouth_t], dtype=np.float32),
+            ]).astype(np.float32, copy=False)
+            feats.append(feat)
+            prev_vel = vel
+            prev = cur
+        x = torch.from_numpy(np.stack(feats))
+        y = LABEL_TO_IDX[row.label]
+        return x, y
+
+
+def _class_weights(rows: list[DatasetRow]) -> torch.Tensor:
+    counts = Counter(r.label for r in rows)
+    total = sum(counts.values())
+    weights = torch.tensor([
+        total / (len(LABELS) * counts.get(l, 1)) for l in LABELS
+    ], dtype=torch.float32)
+    return weights
+
+
+def _run_epoch(model: nn.Module, loader: DataLoader, loss_fn: nn.Module,
+               optim: torch.optim.Optimizer | None,
+               device: str) -> tuple[float, float]:
+    train_mode = optim is not None
+    model.train(train_mode)
+    total_loss = 0.0
+    correct = 0
+    seen = 0
+    for x, y in loader:
+        x = x.to(device)
+        y = y.to(device)
+        B, T, _ = x.shape
+        h = model.init_hidden(batch=B, device=device)
+        logits_last: torch.Tensor | None = None
+        for t in range(T):
+            logits, h = model(x[:, t, :], h)
+            logits_last = logits
+        assert logits_last is not None
+        loss = loss_fn(logits_last, y)
+        if train_mode:
+            optim.zero_grad()
+            loss.backward()
+            optim.step()
+        total_loss += float(loss) * B
+        correct += int((logits_last.argmax(-1) == y).sum())
+        seen += B
+    return total_loss / max(1, seen), correct / max(1, seen)
+
+
+def train(*,
+          dataset_path: Path,
+          ckpt_out: Path,
+          epochs: int = 50,
+          batch_size: int = 128,
+          lr: float = 1e-3,
+          device: str = "cpu",
+          seed: int = 0,
+          log_every: int = 1,
+          ) -> dict[str, list[float]]:
+    torch.manual_seed(seed)
+    rows = load_dataset_jsonl(dataset_path)
+    train_rows, val_rows, _test_rows = split_by_session(rows, seed=seed)
+    LOG.info("train=%d val=%d", len(train_rows), len(val_rows))
+    train_ds = WindowDataset(train_rows, augment=True, seed=seed)
+    val_ds = WindowDataset(val_rows, augment=False, seed=seed)
+    train_loader = DataLoader(train_ds, batch_size=batch_size, shuffle=True)
+    val_loader = DataLoader(val_ds, batch_size=batch_size, shuffle=False)
+    model = ActionHeadModel().to(device)
+    weights = _class_weights(train_rows).to(device)
+    loss_fn = nn.CrossEntropyLoss(weight=weights)
+    optim = torch.optim.AdamW(model.parameters(), lr=lr)
+    history: dict[str, list[float]] = {
+        "train_loss": [], "train_acc": [], "val_loss": [], "val_acc": [],
+    }
+    best_val_acc = -1.0
+    ckpt_out.parent.mkdir(parents=True, exist_ok=True)
+    for ep in range(epochs):
+        tl, ta = _run_epoch(model, train_loader, loss_fn, optim, device)
+        with torch.no_grad():
+            vl, va = _run_epoch(model, val_loader, loss_fn, None, device)
+        history["train_loss"].append(tl)
+        history["train_acc"].append(ta)
+        history["val_loss"].append(vl)
+        history["val_acc"].append(va)
+        if ep % log_every == 0 or ep == epochs - 1:
+            LOG.info("ep=%d train_loss=%.4f train_acc=%.3f val_loss=%.4f val_acc=%.3f",
+                     ep, tl, ta, vl, va)
+        if va > best_val_acc:
+            best_val_acc = va
+            torch.save({"model_state_dict": model.state_dict(),
+                        "version": 1, "val_acc": va}, ckpt_out)
+    return history
+
+
+def _cli() -> None:
+    p = argparse.ArgumentParser()
+    p.add_argument("--dataset", required=True, type=Path)
+    p.add_argument("--ckpt-out", required=True, type=Path)
+    p.add_argument("--epochs", type=int, default=50)
+    p.add_argument("--batch-size", type=int, default=128)
+    p.add_argument("--lr", type=float, default=1e-3)
+    p.add_argument("--device", default="cpu",
+                   choices=["cpu", "mps", "cuda"])
+    p.add_argument("--seed", type=int, default=0)
+    args = p.parse_args()
+    logging.basicConfig(level=logging.INFO,
+                        format="%(asctime)s [%(name)s] %(message)s")
+    hist = train(dataset_path=args.dataset, ckpt_out=args.ckpt_out,
+                 epochs=args.epochs, batch_size=args.batch_size,
+                 lr=args.lr, device=args.device, seed=args.seed)
+    print(json.dumps({"final": {k: v[-1] for k, v in hist.items()}}))
+
+
+if __name__ == "__main__":
+    _cli()
diff --git a/docs/superpowers/plans/2026-05-13-action-head.md b/docs/superpowers/plans/2026-05-13-action-head.md
index 389ce06..c06fa34 100644
--- a/docs/superpowers/plans/2026-05-13-action-head.md
+++ b/docs/superpowers/plans/2026-05-13-action-head.md
@@ -1,12 +1,18 @@
 # action-head Implementation Plan
 
 > **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+>
+> **STATUS 2026-05-13 22:50** — Implementation **complete** (16/17 tasks, Task 16 is a manual gate). 39 tests green. Key deviations from this document, captured in the "Post-impl deviations" section below:
+> - Task 14 pivoted from "modify `multi_hmr_worker_coreml.py` + CLI flag" to **standalone publisher thread `data_only_viz/action_head_pub.py`** + 3-line wire-in in `multi.py` (avoids collision with the user's parallel iteration on `multi_hmr_worker.py`). The MultiHMR backend is selected via env var `MULTIHMR_BACKEND=pytorch|coreml`, not a CLI flag.
+> - Task 11 pivoted from "refactor `MultiHMRWorker` with `create_for_offline()`" to **standalone script using `MultiHMRCoreMLBackend.infer()` directly** — no worker refactor.
+> - j3d is approximated from SMPL-X v3d via a fixed 22-vertex anchor set (`SMPLX_JOINT_ANCHOR_VERTS`), with a MediaPipe 33→22 fallback. The same anchor set is shared between live serve (`action_head_pub.py`) and offline extract (`scripts/extract_j3d_offline.py`) to avoid train/serve skew.
+> - Studio train wrapper added as Task 8.5 (`data_only_viz/scripts/train_on_studio.sh`), validated end-to-end smoke 160 windows × 3 epochs MPS in ~4 s.
 
 **Goal:** Implement a real-time per-person action classifier (debout/assise/danse) on top of Multi-HMR `j3d`, with OSC output enriched by softmax probabilities and kinetics scalars (speed/accel/symmetry).
 
 **Architecture:** GRU-1-layer + MLP head streaming inference, fed by a 16-frame ring buffer per person. Trained windowed on Studio M3 Ultra (PyTorch MPS), inferred streaming on M5. Hybrid auto-labeler (rules on j3d) + manual review for dataset. Inference ≤ 2 ms/person M5 in eager PyTorch — no CoreML conversion needed.
 
-**Tech Stack:** Python 3.11 + uv, PyTorch (MPS for train, CPU for M5 inference), numpy, python-osc, pytest. Reuses existing `data_only_viz` infrastructure (`multi_hmr_worker_coreml.py`, `pose_bridge.py`, `tracker.py`, `state.py`).
+**Tech Stack:** Python 3.11 + uv, PyTorch (MPS for train, CPU for M5 inference), numpy, python-osc, pytest. Reuses existing `data_only_viz` infrastructure (`multi_hmr_worker.py`, `multihmr_coreml.py`, `pose_bridge.py`, `tracker.py`, `state.py`).
 
 **Reference spec:** `docs/superpowers/specs/2026-05-13-action-head-design.md`
 
@@ -1839,6 +1845,8 @@ git commit -m "feat(data-only-viz): action capture script"
 
 ## Task 11 — Extract j3d offline
 
+> **SUPERSEDED 2026-05-13.** The implemented script does NOT refactor `multi_hmr_worker.py`. Instead it uses the standalone `MultiHMRCoreMLBackend.infer()` from `data_only_viz/multihmr_coreml.py` directly. Output jsonl rows contain a (22, 3) `j3d` extracted via `SMPLX_JOINT_ANCHOR_VERTS` (shared with `action_head_pub.py` to avoid train/serve skew), not the raw v3d. See actual file at `data_only_viz/scripts/extract_j3d_offline.py`. The body below documents the original intent — keep as historical context.
+
 **Files:**
 - Create: `data_only_viz/scripts/extract_j3d_offline.py`
 
@@ -2196,6 +2204,8 @@ git commit -m "feat(data-only-viz): pose_bridge /pose/action + /pose/kin"
 
 ## Task 14 — Wire ActionHead into multi_hmr_worker_coreml
 
+> **SUPERSEDED 2026-05-13.** No `multi_hmr_worker_coreml.py` file exists in the current repo — the user pivoted to `multihmr_coreml.py` (standalone backend) selected via env `MULTIHMR_BACKEND=pytorch|coreml` inside `multi_hmr_worker.py`. To avoid colliding with that file under active iteration, ActionHead wiring was implemented as a standalone publisher thread in `data_only_viz/action_head_pub.py` plus a 3-line wire-in inside `data_only_viz/multi.py` (`__init__` instantiates and `.start()`s the publisher). The publisher polls `state.persons_smplx` (preferred) and `state.persons_body3d` (MediaPipe fallback) at 30 Hz, deduplicates by timestamp, extracts j3d22 via shared `SMPLX_JOINT_ANCHOR_VERTS` / `MEDIAPIPE_TO_22` index maps, runs `ActionHead.step()` per pid, and emits OSC via the existing `PoseSoundBridge`. No CLI flag was added. See `data_only_viz/action_head_pub.py` and `data_only_viz/multi.py:22,97-98`. The body below documents the original intent — keep as historical context.
+
 **Files:**
 - Modify: `data_only_viz/multi_hmr_worker_coreml.py`
 - Modify: `data_only_viz/main.py`
diff --git a/docs/superpowers/specs/2026-05-13-action-head-design.md b/docs/superpowers/specs/2026-05-13-action-head-design.md
index 7c262e2..e62a35e 100644
--- a/docs/superpowers/specs/2026-05-13-action-head-design.md
+++ b/docs/superpowers/specs/2026-05-13-action-head-design.md
@@ -1,11 +1,17 @@
 # action-head — Classifier d'action temps réel au-dessus de Multi-HMR
 
 > **Date** : 2026-05-13
-> **Status** : design approuvé, prêt pour implementation plan
+> **Status** : design approuvé — **implémenté 2026-05-13 22:50**, 16/17 tasks, 39 tests verts. Task 16 (E2E gate) reste manuel (requiert capture + train réel).
 > **Authors** : L'Electron Rare + Claude
 > **Companion plans** :
 > - `2026-05-13-multihmr-coreml-hybrid-backbone.md`
 > - `2026-05-13-studio-train-deploy-m5.md`
+>
+> **Déviations notables vs design original** (cf. plan `2026-05-13-action-head.md` pour le détail) :
+> - **Wiring worker** : standalone publisher thread `data_only_viz/action_head_pub.py` + 3 lignes dans `multi.py`, au lieu de modifier directement `multi_hmr_worker.py` (qui était en cours d'évolution par l'utilisateur en parallèle). Backend Multi-HMR sélectionné par env `MULTIHMR_BACKEND=pytorch|coreml`, pas par flag CLI.
+> - **Source j3d** : approximée via 22 vertex anchors (`SMPLX_JOINT_ANCHOR_VERTS`) sur le mesh SMPL-X 10475-vert, partagés entre serve live (`action_head_pub.py`) et extraction offline (`scripts/extract_j3d_offline.py`) pour éviter le train/serve skew. Fallback MediaPipe 33→22 (`MEDIAPIPE_TO_22`) quand `persons_smplx` est vide. **Limitation** : ces 22 indices sont approximatifs ; pour des j3d SMPL-X corrects, brancher `J_regressor @ v3d` quand le module SMPL-X est dispo.
+> - **Extract offline** : pas de refactor de `MultiHMRWorker`, on utilise `MultiHMRCoreMLBackend.infer()` directement (commit user `9e7a9f8`).
+> - **Studio launch** : wrapper bash `data_only_viz/scripts/train_on_studio.sh` (Task 8.5) qui rsync + ssh + uv sync + train MPS + ckpt back. Validé end-to-end sur dataset smoke 160 windows × 3 epochs en ~4 s wallclock.
 
 ## TL;DR
 
@@ -94,7 +100,9 @@ class ActionHead:
     def forget(self, pid: int) -> None: ...
 ```
 
-Aucune modification de l'API publique de `multi_hmr_worker_coreml.py` autre que :
+**Note d'implémentation 2026-05-13** : la section ci-dessous décrit l'intention originale. L'implémentation réelle est dans `data_only_viz/action_head_pub.py` (publisher thread) — pas de modification de `multi_hmr_worker.py`. Voir l'en-tête du document pour les déviations.
+
+Aucune modification de l'API publique de `multi_hmr_worker.py` n'est requise au-delà de :
 - Construction d'une `ActionHead` au startup.
 - Appel `.step()` après chaque détection.
 - Appel `.forget()` synchronisé avec `tracker.purge()`.
diff --git a/sound_algo/data_only/scene_pose_action.scd b/sound_algo/data_only/scene_pose_action.scd
new file mode 100644
index 0000000..096c119
--- /dev/null
+++ b/sound_algo/data_only/scene_pose_action.scd
@@ -0,0 +1,131 @@
+// =====================================================================
+//  scene_pose_action.scd -- pilote les FX live a partir de l'action_head.
+//
+//  Necessite : data_feeds.scd deja charge (handlers OSCdef \poseAction,
+//  \poseKin, \poseEnter, \poseLeave). Les dicts ~poseState et ~poseKin
+//  doivent etre populates en temps reel par data_only_viz.action_head_pub.
+//
+//  Mapping :
+//    speed (m/s mean joint)       -> drive amount    0..1
+//    accel (m/s2)                 -> filter cutoff   200 Hz..6 kHz
+//    symmetry (-1..1)             -> stereo width    0..1
+//    label argmax (0=debout, 1=assise, 2=danse) :
+//      - debout  -> reverb mid, kick on, melody pad
+//      - assise  -> pad ambient, kick off, lo-fi
+//      - danse   -> drive max, kick punchy, acid lead
+//
+//  Usage live (a evaluer bloc par bloc dans le SC IDE) :
+//      [0] SETUP : declare le helper ~mapPoseToFx + lance la Routine.
+//      [1] START : ~scenePoseAction.start
+//      [2] STOP  : ~scenePoseAction.stop
+//      [3] TWEAK : ajuster les seuils dans ~paConfig
+// =====================================================================
+
+// [0] SETUP -----------------------------------------------------------
+(
+~paConfig = (
+	speedMin:      0.0,    speedMax:    0.8,
+	accelMin:      0.0,    accelMax:    3.0,
+	cutoffMin:     200,    cutoffMax:   6000,
+	rate:          10,     // refresh Hz
+);
+
+~mapPoseToFx = {
+	var avgSpeed = 0, avgAccel = 0, avgSym = 0, avgProbs = [0, 0, 0];
+	var n = max(1, ~poseKin.size);
+	var labelIdx;
+
+	~poseKin.values.do { |kin|
+		avgSpeed = avgSpeed + (kin[\speed] ? 0);
+		avgAccel = avgAccel + (kin[\accel] ? 0);
+		avgSym   = avgSym   + (kin[\symmetry] ? 0);
+	};
+	avgSpeed = avgSpeed / n;
+	avgAccel = avgAccel / n;
+	avgSym   = avgSym   / n;
+
+	~poseState.values.do { |ps|
+		var p = ps[\probs] ? [0, 0, 0];
+		avgProbs = avgProbs.collect { |v, i| v + (p[i] ? 0) };
+	};
+	avgProbs = avgProbs.collect { _ / n };
+	labelIdx = avgProbs.maxIndex ? 0;
+
+	// Drive + filter + width
+	if (~fxDrive.notNil) {
+		~fxDrive.(avgSpeed.linlin(~paConfig[\speedMin], ~paConfig[\speedMax], 0, 1));
+	};
+	if (~fxCut.notNil) {
+		~fxCut.(avgAccel.linexp(~paConfig[\accelMin], ~paConfig[\accelMax],
+								~paConfig[\cutoffMin], ~paConfig[\cutoffMax]));
+	};
+	if (~fxSt.notNil) { ~fxSt.(avgSym.linlin(-1, 1, 0, 1)) };
+
+	// Label-driven track switch (smooth crossfade between presets)
+	switch (labelIdx,
+		0, {
+			if (~fxComp.notNil) { ~fxComp.(0.7) };
+			if (~fxRev.notNil)  { ~fxRev.(0.4) };
+		},
+		1, {
+			if (~fxComp.notNil) { ~fxComp.(0.3) };
+			if (~fxRev.notNil)  { ~fxRev.(0.6) };
+		},
+		2, {
+			if (~fxComp.notNil) { ~fxComp.(0.9) };
+			if (~fxRev.notNil)  { ~fxRev.(0.2) };
+			if (~fxDrive.notNil){ ~fxDrive.(avgSpeed.linlin(0, 0.6, 0.3, 1.0)) };
+		}
+	);
+
+	[labelIdx, avgSpeed.round(0.01), avgAccel.round(0.01),
+	 avgSym.round(0.01)];
+};
+
+~scenePoseAction = ~scenePoseAction ?? {
+	(
+		running: false,
+		routine: nil,
+		start: {
+			if (~scenePoseAction[\running]) {
+				"[scene_pose_action] already running".postln;
+			} {
+				~scenePoseAction[\running] = true;
+				~scenePoseAction[\routine] = Routine({
+					inf.do {
+						var snapshot = ~mapPoseToFx.();
+						if (~scenePoseAction[\verbose] ? false) {
+							("[pose] label=" ++ snapshot[0]
+							 ++ " s=" ++ snapshot[1]
+							 ++ " a=" ++ snapshot[2]
+							 ++ " sym=" ++ snapshot[3]).postln;
+						};
+						(1 / ~paConfig[\rate]).wait;
+					};
+				}).play;
+				"[scene_pose_action] STARTED".postln;
+			};
+		},
+		stop: {
+			~scenePoseAction[\routine] !? { |r| r.stop };
+			~scenePoseAction[\routine] = nil;
+			~scenePoseAction[\running] = false;
+			"[scene_pose_action] STOPPED".postln;
+		},
+		verbose: false,
+	)
+};
+
+"[OK] SETUP scene_pose_action".postln;
+)
+
+// [1] START -----------------------------------------------------------
+// ~scenePoseAction[\verbose] = true;   // optionnel : log chaque tick
+// ~scenePoseAction[\start].();
+
+// [2] STOP ------------------------------------------------------------
+// ~scenePoseAction[\stop].();
+
+// [3] TWEAK ------------------------------------------------------------
+// ~paConfig[\speedMax] = 1.2;
+// ~paConfig[\rate]     = 20;