From 8d6d42ed3cf15958f000bf0918cdbbdfa83847b7 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?L=27=C3=A9lectron=20rare?=
 <108685187+electron-rare@users.noreply.github.com>
Date: Thu, 2 Jul 2026 18:47:57 +0200
Subject: [PATCH] feat(pose): wrist-anchored xy mods + wrist markers

The X/Y voice mods were anchored to the 21-landmark centroid, so
opening or pinching the fingers shifted the mod values (gesture/mod
cross-talk). cx/cy now track the WRIST (kp 0). The renderer draws a
cross marker at each wrist on the video whenever features exist, so
the control point stays visible even for far hands the overlay
filters out (user request live 2026-07-02).
---
 data_only_viz/hand_features.py |  8 ++++++--
 data_only_viz/renderer.py      | 10 ++++++++++
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/data_only_viz/hand_features.py b/data_only_viz/hand_features.py
index bb7650d..0b76516 100644
--- a/data_only_viz/hand_features.py
+++ b/data_only_viz/hand_features.py
@@ -62,8 +62,12 @@ class HandFeatureExtractor:
     def _features(self, lm: list) -> dict:
         xs = [_finite(_coord(p, "x", 0), 0.5) for p in lm[:21]]
         ys = [_finite(_coord(p, "y", 1), 0.5) for p in lm[:21]]
-        cx = _clamp(sum(xs) / len(xs), 0.0, 1.0)
-        cy = _clamp(sum(ys) / len(ys), 0.0, 1.0)
+        # cx/cy anchor = the WRIST (kp 0), not the landmark centroid: the
+        # centroid shifts when fingers open/pinch, cross-talking gestures
+        # into the X/Y voice mods. The wrist decouples them (user request
+        # live 2026-07-02).
+        cx = _clamp(xs[WRIST], 0.0, 1.0)
+        cy = _clamp(ys[WRIST], 0.0, 1.0)
         size = math.hypot(xs[MIDDLE_MCP] - xs[WRIST], ys[MIDDLE_MCP] - ys[WRIST])
         size = size if size > 1e-4 else 1e-4
         span = math.hypot(xs[THUMB_TIP] - xs[PINKY_TIP],
diff --git a/data_only_viz/renderer.py b/data_only_viz/renderer.py
index 2b11c35..0f7373a 100644
--- a/data_only_viz/renderer.py
+++ b/data_only_viz/renderer.py
@@ -575,6 +575,16 @@ class MetalRenderer(NSObject):
                     content_pid=pid_s,
                 ):
                     push_panel(ax, ay, bx, by, gconf, gpid)
+                # Wrist marker ON the video: the X/Y voice mods are anchored
+                # to the wrist, so its control point stays visible whenever
+                # features exist (even for far hands the overlay filters
+                # out). push_seg applies the video mirror like the overlay.
+                if _g_cx is not None and _g_cy is not None:
+                    _mx = 0.015 / asp
+                    push_seg(_g_cx - _mx, _g_cy, _g_cx + _mx, _g_cy,
+                             1.0, pid_s)
+                    push_seg(_g_cx, _g_cy - 0.015, _g_cx, _g_cy + 0.015,
+                             1.0, pid_s)
                 if h_kp is not None:
                     for seg in panel_segments(
                         h_kp, side_name, lhand_bones_p, asp,