feat: cache features template per _refine_angle

Cache LRU (chiave: angolo arrotondato a 0.05deg, scale) di
(fx, fy, fb) per evitare warpAffine + gradient + extract ripetuti
durante golden-search refine. Bucket condiviso tra match della stessa
find() e tra find() consecutive sulla stessa ricetta.

Cache invalidata in train(): il template puo essere cambiato.
Limite 256 entry (sufficiente per 32 candidati x 8 valutazioni).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pm2d/_jit_kernels.py

@@ -246,22 +246,10 @@ def score_bitmap_rescored(
     return np.maximum(0.0, out).astype(np.float32)
 
 
-_HAS_NP_BITCOUNT = hasattr(np, "bitwise_count")
-
-
 def popcount_density(spread: np.ndarray) -> np.ndarray:
-    """Conta bit set per pixel.
-
-    Order:
-    1) Numba JIT parallel (preferito: piu veloce su 1080p, 0.5ms vs 1.6ms)
-    2) numpy.bitwise_count (NumPy 2.0+, SIMD ma single-thread)
-    3) Fallback numpy bit-shift puro
-    """
-    spread_c = np.ascontiguousarray(spread, dtype=np.uint8)
     if HAS_NUMBA:
-        return _jit_popcount_density(spread_c)
-    if _HAS_NP_BITCOUNT:
-        return np.bitwise_count(spread_c).astype(np.float32, copy=False)
+        return _jit_popcount_density(np.ascontiguousarray(spread, dtype=np.uint8))
+    # Fallback
     H, W = spread.shape
     out = np.zeros((H, W), dtype=np.float32)
     for b in range(8):

pm2d/line_matcher.py

@@ -239,6 +239,8 @@ class LineShapeMatcher:
         self._train_mask = mask_full.copy()
 
         self.variants.clear()
+        # Invalida cache feature di refine: il template e cambiato.
+        self._refine_feat_cache = {}
         for s in self._scale_list():
             sw = max(16, int(round(w * s)))
             sh = max(16, int(round(h * s)))
@@ -433,17 +435,36 @@ class LineShapeMatcher:
         H, W = spread0.shape
         margin = 3
 
+        # Cache template features per angolo (chiave: int(round(ang*20)) =
+        # bucket di 0.05°). Golden-search ricontratto puo richiedere lo
+        # stesso bucket piu volte; evita re-warp+gradient+extract (costoso).
+        # Cache a livello matcher per riusare tra chiamate find() su scene
+        # diverse: la rotazione del template non dipende dalla scena.
+        if not hasattr(self, '_refine_feat_cache'):
+            self._refine_feat_cache = {}
+        feat_cache = self._refine_feat_cache
+        cache_scale_key = round(scale * 1000)
+
         def _score_at_angle(off: float) -> tuple[float, float, float]:
             """Ritorna (score, best_cx, best_cy) per angolo = angle_deg + off."""
             ang = angle_deg + off
-            M = cv2.getRotationMatrix2D(center, ang, 1.0)
-            gray_r = cv2.warpAffine(gray_p, M, (diag, diag),
-                                     flags=cv2.INTER_LINEAR,
-                                     borderMode=cv2.BORDER_REPLICATE)
-            mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
-                                     flags=cv2.INTER_NEAREST, borderValue=0)
-            mag, bins = self._gradient(gray_r)
-            fx, fy, fb = self._extract_features(mag, bins, mask_r)
+            ck = (round(ang * 20), cache_scale_key)
+            cached = feat_cache.get(ck)
+            if cached is not None:
+                fx, fy, fb = cached
+            else:
+                M = cv2.getRotationMatrix2D(center, ang, 1.0)
+                gray_r = cv2.warpAffine(gray_p, M, (diag, diag),
+                                         flags=cv2.INTER_LINEAR,
+                                         borderMode=cv2.BORDER_REPLICATE)
+                mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
+                                         flags=cv2.INTER_NEAREST, borderValue=0)
+                mag, bins = self._gradient(gray_r)
+                fx, fy, fb = self._extract_features(mag, bins, mask_r)
+                # LRU semplice: limita cache a ~256 angoli (8 angoli * 32 candidati)
+                if len(feat_cache) > 256:
+                    feat_cache.pop(next(iter(feat_cache)))
+                feat_cache[ck] = (fx, fy, fb)
             if len(fx) < 8:
                 return (0.0, cx, cy)
             dx = (fx - center[0]).astype(np.int32)