diff --git a/public/js/ar/solve-cv.js b/public/js/ar/solve-cv.js
index 1df2931..1eda0b8 100644
--- a/public/js/ar/solve-cv.js
+++ b/public/js/ar/solve-cv.js
@@ -4,7 +4,7 @@
  * becomes a 3D world point (from scene.json anchors, same anchorWorldMatrix
  * convention as v2), and ONE solvePnP recovers the camera pose directly:
  *   - stability scales with marker spread instead of degrading into fusion tuning
- *   - the wall crest makes the point set non-coplanar => planar ambiguity gone
+ *   - a non-coplanar crest (one standing vertical) removes planar ambiguity
  *   - single visible marker still works (IPPE init + iterative refine w/ prior)
  *
  * The solve itself runs in cv-worker.js. OpenCV.js embeds ~8MB of WASM as a
@@ -95,6 +95,7 @@ export class CvBoardSolver {
     this.lastPose = null;           // newest smoothed pose from the worker
     this.lastReproj = 0;
     this.lastCount = 0;
+    this.lastReject = '';           // why the worker last refused a pose (HUD)
     this._v = new THREE.Vector3();
     this._m4 = new THREE.Matrix4();
     this._q = new THREE.Quaternion();
@@ -126,7 +127,10 @@ export class CvBoardSolver {
 
   _onPose(m) {
     this.inFlight = false;
-    if (!m.ok) return;
+    // a rejected pose leaves the previous one in place: better a slightly stale
+    // camera than a mirrored one that throws every ghost across the table
+    if (!m.ok) { this.lastReject = m.why || 'fail'; return; }
+    this.lastReject = '';
     this.lastReproj = m.reproj || 0;
     this.lastCount = m.n / 4;
     this._m4.set(