fix: switch to Lasso regression for better axis stability

- Replace Ridge with Lasso (L1) regression to concentrate weights on
  fewer dimensions, improving stability measurement
- Default alpha changed to 0.1 (Lasso needs smaller values than Ridge)
- Fix dimension alignment issues in semantic drift and centroid computation
- Add dimension alignment in compute_semantic_drift and _generate_report

Results with Lasso alpha=0.1:
- 9/10 axes now stable (>0.7): [1, 2, 3, 4, 5, 7, 8, 9, 10]
- Axis 6 reordered (0.25-0.5 range)
- Axis 8 shows inflection points in 2016→2017→2018
- Overtone shift detected on all stable axes (1.3-1.9 range)

scripts/motion_drift.py

@@ -157,7 +157,7 @@ def compute_axis_stability(
     Returns dict with stability_matrix, stable_axes, reordered_axes, unstable_axes,
     and weight_vectors for downstream interpretation.
     """
-    from sklearn.linear_model import Ridge
+    from sklearn.linear_model import Lasso
     from sklearn.preprocessing import StandardScaler
 
     # Load data per window
@@ -169,16 +169,12 @@ def compute_axis_stability(
         if not motion_scores or not fused:
             continue
 
-        # Build feature matrix and targets
-        # Use motions that have both SVD scores and fused embeddings
         common = [m for m in motion_scores if m in fused]
         if len(common) < 50:
             continue
 
-        # Feature matrix: fused embeddings (align dimensions)
         dim = min(len(fused[m]) for m in common)
         X = np.array([fused[m][:dim] for m in common])
-        # Target matrix: SVD scores (n_common × n_components)
         Y = np.array([motion_scores[m][:n_components] for m in common])
 
         window_data[w] = (X, Y)
@@ -188,20 +184,21 @@ def compute_axis_stability(
             con, windows, n_components, stability_threshold
         )
 
-    # Fit Ridge regression per axis per window
+    # Fit Lasso regression per axis per window
+    # Lasso (L1) produces sparse weight vectors, concentrating on the
+    # most important embedding dimensions for each axis
     weight_vectors: Dict[str, Dict[int, np.ndarray]] = {}
     window_list = sorted(window_data.keys())
 
     for w in window_list:
         X, Y = window_data[w]
-        # Normalize features
         scaler = StandardScaler()
         X_scaled = scaler.fit_transform(X)
 
         weights = {}
         for comp_idx in range(n_components):
             y = Y[:, comp_idx]
-            model = Ridge(alpha=regression_alpha)
+            model = Lasso(alpha=regression_alpha, max_iter=5000)
             model.fit(X_scaled, y)
             weights[comp_idx + 1] = model.coef_
 
@@ -754,8 +751,10 @@ def compute_semantic_drift(
             if not valid_motions:
                 continue
 
-            # Compute centroid
-            vectors = np.array([fused[m] for m in valid_motions])
+            # Compute centroid (align dimensions)
+            vectors = [fused[m] for m in valid_motions]
+            dim = min(len(v) for v in vectors)
+            vectors = np.array([v[:dim] for v in vectors])
             centroid = np.mean(vectors, axis=0)
             centroids.append(centroid)
             window_centroids[w] = {
@@ -770,6 +769,10 @@ def compute_semantic_drift(
         drift_values = []
         for i in range(len(centroids) - 1):
             a, b = centroids[i], centroids[i + 1]
+            # Align dimensions
+            dim = min(len(a), len(b))
+            a = a[:dim]
+            b = b[:dim]
             norm_a = np.linalg.norm(a)
             norm_b = np.linalg.norm(b)
             if norm_a == 0 or norm_b == 0:
@@ -1292,8 +1295,8 @@ def main(argv: Optional[List[str]] = None) -> int:
     p.add_argument(
         "--regression-alpha",
         type=float,
-        default=1.0,
-        help="Ridge regression regularization strength (default: 1.0)",
+        default=0.1,
+        help="Lasso regression regularization strength (default: 0.1)",
     )
     args = p.parse_args(argv)