diff --git a/data_only_viz/renderer.py b/data_only_viz/renderer.py
index 3ace31d..58d72fb 100644
--- a/data_only_viz/renderer.py
+++ b/data_only_viz/renderer.py
@@ -243,7 +243,9 @@ class MetalRenderer(NSObject):
 
     def _init_mesh_cpu_buffer(self) -> None:
         if getattr(self, "_mesh_cpu_buf", None) is None:
-            self._mesh_cpu_buf = np.zeros(MESH_MAX_VERTS * 5, dtype=np.float32)
+            self._mesh_cpu_buf = np.zeros(
+                MESH_MAX_VERTS * MESH_VERT_FLOATS, dtype=np.float32,
+            )
 
     # ---- Uniforms helpers ------------------------------------------
     def _update_uniforms(self) -> int:
@@ -401,7 +403,7 @@ class MetalRenderer(NSObject):
             ou triangle invalide (confiance basse)."""
             nonlocal n_verts
             tris = n_verts // 3
-            if tris >= MESH_MAX_TRIS:
+            if tris >= MESH_MAX_TRIS or n_verts + 3 > MESH_MAX_VERTS:
                 return False
             if i >= len(kp_list) or j >= len(kp_list) or k >= len(kp_list):
                 return True   # skip mais continue
@@ -464,7 +466,8 @@ class MetalRenderer(NSObject):
 
         if n_verts == 0:
             return 0
-        data = self._mesh_cpu_buf[: n_verts * 5].tobytes()
+        # Slice is exact — stale floats beyond n_verts*MESH_VERT_FLOATS never reach the GPU.
+        data = self._mesh_cpu_buf[: n_verts * MESH_VERT_FLOATS].tobytes()
         mv = self._mesh_buf.contents().as_buffer(len(data))
         mv[:] = data
         return n_verts // 3
diff --git a/data_only_viz/tests/test_renderer_allocations.py b/data_only_viz/tests/test_renderer_allocations.py
index 26bfb01..0c639d6 100644
--- a/data_only_viz/tests/test_renderer_allocations.py
+++ b/data_only_viz/tests/test_renderer_allocations.py
@@ -1,10 +1,15 @@
 # data_only_viz/tests/test_renderer_allocations.py
 """Renderer must reuse a preallocated CPU staging buffer for the skeleton."""
 
+import time
+import types
+import unittest.mock as mock
+
 import numpy as np
+import pytest
 
 from data_only_viz.renderer import MetalRenderer, SKEL_MAX_SEGS, MESH_MAX_VERTS
-from data_only_viz.state import State
+from data_only_viz.state import PoseKp, State
 
 
 def test_skeleton_cpu_buffer_is_preallocated_and_reused():
@@ -38,3 +43,49 @@ def test_mesh_cpu_buffer_is_preallocated_and_reused():
     assert buf.size == MESH_MAX_VERTS * 5
     r._init_mesh_cpu_buffer()
     assert r._mesh_cpu_buf is buf
+
+
+def test_mesh_buffer_overflow_guard():
+    """push_tri must stop writing before exceeding MESH_MAX_VERTS.
+
+    Strategy: build a State with enough face-person entries that the old
+    MESH_MAX_TRIS-only guard would allow n_verts > MESH_MAX_VERTS (old cap
+    was MESH_MAX_TRIS*3 = 24576), whereas the new guard caps at MESH_MAX_VERTS
+    (10475).  FACE_TRIANGLES has 36 tris → 108 verts per person.
+    200 persons × 108 = 21600 verts > 10475, so the vertex guard fires first.
+
+    _update_mesh calls self._mesh_buf.contents().as_buffer() — we stub that
+    out since no GPU is available in unit tests.
+    """
+    from data_only_viz.renderer import MESH_MAX_VERTS
+
+    r = MetalRenderer.__new__(MetalRenderer)
+    r._init_mesh_cpu_buffer()
+
+    # Stub _mesh_buf so the GPU upload path does not crash.
+    fake_mv = bytearray(MESH_MAX_VERTS * 5 * 4)  # max possible bytes
+
+    class FakeBuf:
+        def contents(self):
+            return self
+
+        def as_buffer(self, n):
+            return memoryview(fake_mv)[:n]
+
+    r._mesh_buf = FakeBuf()
+
+    # Build a State with 200 face persons, each with 478 keypoints at conf=1.0.
+    # With FACE_TRIANGLES (36 tris = 108 verts/person), 200 persons would write
+    # 21600 verts — well past MESH_MAX_VERTS — if the guard were absent.
+    s = State()
+    s.pose_last_t = time.monotonic()  # make pose_alive() return True
+    visible_kp = [PoseKp(x=0.5, y=0.5, z=0.0, c=1.0) for _ in range(478)]
+    s.persons_face = [list(visible_kp) for _ in range(200)]
+
+    n_tris = r._update_mesh(s)
+    n_verts = n_tris * 3
+
+    assert n_verts <= MESH_MAX_VERTS, (
+        f"Buffer overflow: _update_mesh wrote {n_verts} verts, "
+        f"but MESH_MAX_VERTS={MESH_MAX_VERTS}"
+    )