merge: precisione rotazione + perf propagate + robustezza server

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.gitignore

@@ -8,3 +8,5 @@ __pycache__/
 .DS_Store
 *.log
 models/
+# Ricette pre-trained (generate da utente, non versionare)
+recipes/*.npz

ROADMAP.md

@@ -2,6 +2,36 @@
 
 Lista ragionata di miglioramenti futuri. Priorità = impatto / effort, non urgenza temporale.
 
+## Fase 2 COMPLETATA (precisione rotazione + robustezza + perf)
+
+Root cause della rotazione imprecisa: lo score satura a 1.0 sulla spread
+bitmap dilatata (raggio 4-5) → il refine non vedeva gradiente né in angolo
+né in posizione, e `cv2.minMaxLoc` sul plateau saturo spostava il centro
+sull'angolo della finestra (errore sistematico 3·√2 ≈ 4.24 px).
+
+| Fix | Dettaglio |
+|---|---|
+| Refine su bitmap fine | `_refine_angle` ottimizza su spread raggio 1 (`spread_fine`, cached); score finale ricalcolato su spread coarse per mantenere semantica soglie |
+| Picco sub-pixel nel refine | centroide plateau / fit quadratico al posto di minMaxLoc (bias top-left) |
+| LM least-squares pos+angolo | `_subpixel_refine_lm` riscritto: snap edge ±2px lungo normale + LSQ 3x3 (dx, dy, dθ), ON di default |
+| Round feature offsets | troncamento `astype(int32)` → `np.round` (bias ~0.25 px) |
+| Centro rotazione coerente | `_prepare_padded_template`: rotazione attorno al centro reale del template nel padding (bias ≤0.5 px dipendente dall'angolo) |
+| `_angle_list` include estremo | range parziali ±tol ora testano anche +tol |
+| `_refine_pose_joint` rimosso | Nelder-Mead su funzione a gradini satura: terminava subito; param ora alias di refine_angle |
+| pyramid_propagate di default | kernel windowed (feature campionano l'intera scena: prima il crop troncava le feature → score 0); picchi = massimi locali (non top-K pixel); disattivato automaticamente per template elongati (>2:1) dove il picco top-level non localizza |
+| Piramide 3 livelli default | con clamp automatico sulla dimensione template (min 12 px al top) |
+| Cache scena: hash completo | prima hashava solo i primi 64KB → collisioni tra scene con stessa banda superiore → risultati della scena sbagliata |
+| Web server | lock matcher (race con threadpool FastAPI), LRU `_IMG_CACHE`, clamp ROI ovunque (400/422 invece di 500), `filtro_fp=off` disabilita davvero NCC, `_draw_matches` su crop locale |
+| GUI/legacy | centro overlay `(W-1)/2`→`W/2`, spread_radius default 5→4, EdgeShapeMatcher: angle list endpoint + cap candidati + save template_gray |
+
+Misure (GT sintetica 7 pose, scena 900x700, VPS 2 core):
+- Errore angolare mediano: **2.3° → 0.05°** (step 5°); a step 2° era 4.4° → **0.03°**
+- Errore posizione mediano: **4.24 px → 0.04 px**
+- find GT scene: 4.7s → 1.7s; scena reale 646x482: 1.14s → 0.81s
+- Benchmark suite 16 scenari: 96.5s → 84.2s, match count ≥ baseline
+  (eccezioni: dado_full -1 = match borderline su parte diversa;
+  lama_part_preciso 25→18 con baseline al cap max_matches)
+
 ## Fase 1 COMPLETATA (branch `speedFase1`)
 
 | ID | Voce | Status | Note |
@@ -84,9 +114,14 @@ Benchmark suite 16 scenari (4 immagini × full/part × fast/preciso):
 ## Target performance produzione
 
 Obiettivi da documento tecnico Vision Suite (Fase Beta):
-- [ ] **Precisione posizionale mediana**: <0.5 px → **raggiunto con subpixel (attualmente ~0.1-0.3 px atteso)**
-- [ ] **Precisione angolare mediana**: <1.0° → **raggiunto con refinement (~0.5°)**
-- [ ] **Latency mediana**: <50 ms su 1920×1080 → **attuale ~1.7s su 830×822 (serve GPU o ulteriore CPU)**
+- [x] **Precisione posizionale mediana**: <0.5 px → **0.04 px misurato su GT sintetica (Fase 2)**
+- [x] **Precisione angolare mediana**: <1.0° → **0.05° misurato su GT sintetica (Fase 2)**
+- [ ] **Latency mediana**: <50 ms su 1920×1080 → **~0.8s su 646×482 con 2 core; da misurare su hardware produzione**
 - [ ] **F1 score dataset sintetico**: >0.95 → **da misurare con dataset sintetico**
 
-Prossimo blocker per target: **latency**. Via più promettente: GPU (CuPy) o coarse-to-fine angolare.
+Prossimo blocker per target: **latency**. Nota: i kernel hot sono gia'
+Numba JIT (≈ velocita' C, prange parallelo): un port C++ dei kernel vale
+solo il margine SIMD esplicito (~2-4x con AVX2 su AND+popcount byte-wise).
+Prima di scriverlo conviene esaurire le vie algoritmiche rimaste:
+riduzione varianti al top-level (auto angle step per livello, stile
+Halcon), greediness di default, e GPU (CuPy/OpenCL) per scene 1080p.

benchmarks/test_suite.py

@@ -36,6 +36,11 @@ CONFIGS = [
 def bench(case_name: str, img_path: str, roi_box: tuple, roi_kind: str,
           cfg_name: str, cfg: dict) -> dict:
     scene = cv2.imread(str(TEST_DIR / img_path))
+    if scene is None:
+        # cv2.imread ritorna None silenzioso: senza check il crash arriva
+        # dopo, sullo slice, con un errore criptico.
+        raise FileNotFoundError(
+            f"Immagine di test non trovata o non leggibile: {TEST_DIR / img_path}")
     y0, y1, x0, x1 = roi_box
     roi = scene[y0:y1, x0:x1].copy()
     m = LineShapeMatcher(

pm2d/_jit_kernels.py

@@ -110,6 +110,63 @@ if HAS_NUMBA:
                     acc[y, x] *= inv
         return acc
 
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_strided(
+        spread: np.ndarray,
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint8,
+        bg: np.ndarray,
+        stride: nb.int32,
+    ) -> np.ndarray:
+        """Variante con sub-sampling: valuta solo pixel su griglia stride×stride.
+        Score restituito ha stessa shape (H, W); celle non valutate = 0.
+
+        4× speed-up con stride=2 (NMS recupera precisione in full-res).
+        Numba prange richiede step costante: itero su indici griglia e
+        moltiplico per stride dentro il body.
+        """
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((H, W), dtype=np.float32)
+        ny = (H + stride - 1) // stride
+        nx = (W + stride - 1) // stride
+        for yi in nb.prange(ny):
+            y = yi * stride
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint8(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                ddy = dy[i]
+                yy = y + ddy
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                x_lo = 0 if ddx >= 0 else -ddx
+                x_hi = W if ddx <= 0 else W - ddx
+                rem = x_lo % stride
+                if rem != 0:
+                    x_lo += stride - rem
+                x = x_lo
+                while x < x_hi:
+                    if spread[yy, x + ddx] & mask:
+                        acc[y, x] += 1.0
+                    x += stride
+        if N > 0:
+            inv = 1.0 / N
+            for yi in nb.prange(ny):
+                y = yi * stride
+                for xi in range(nx):
+                    x = xi * stride
+                    v = acc[y, x] * inv
+                    bgv = bg[y, x]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[y, x] = r if r > 0.0 else 0.0
+                    else:
+                        acc[y, x] = 0.0
+        return acc
+
     @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
     def _jit_score_bitmap_greedy(
         spread: np.ndarray,
@@ -142,7 +199,6 @@ if HAS_NUMBA:
                     b = bins[i]
                     mask = np.uint8(1) << b
                     if (bit_active & mask) == 0:
-                        # Nessun chance per questa feature
                         if hits + (N - i - 1) < min_req:
                             break
                         continue
@@ -215,6 +271,224 @@ if HAS_NUMBA:
                         acc[y, x] = 0.0
         return acc
 
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_window(
+        spread: np.ndarray,      # uint8 (H, W) - scena INTERA
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint8,
+        bg: np.ndarray,          # float32 (H, W) - scena intera
+        y0: nb.int64, x0: nb.int64,
+        wh: nb.int64, ww: nb.int64,
+    ) -> np.ndarray:
+        """Score rescored valutato SOLO nella finestra (y0, x0, wh, ww).
+
+        Le feature campionano lo spread dell'intera scena (bounds-checked
+        sui bordi scena): a differenza di chiamare il kernel su un crop,
+        le feature che escono dalla finestra NON contano come miss.
+        Usato dal path pyramid_propagate: costo ∝ area finestra.
+        """
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((wh, ww), dtype=np.float32)
+        for yi in nb.prange(wh):
+            y = y0 + yi
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint8(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                yy = y + dy[i]
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                xi_lo = 0
+                xi_hi = ww
+                lo = -(x0 + ddx)
+                if lo > xi_lo:
+                    xi_lo = lo
+                hi = W - (x0 + ddx)
+                if hi < xi_hi:
+                    xi_hi = hi
+                for xi in range(xi_lo, xi_hi):
+                    if spread[yy, x0 + xi + ddx] & mask:
+                        acc[yi, xi] += 1.0
+        if N > 0:
+            inv = 1.0 / N
+            for yi in nb.prange(wh):
+                for xi in range(ww):
+                    v = acc[yi, xi] * inv
+                    bgv = bg[y0 + yi, x0 + xi]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[yi, xi] = r if r > 0.0 else 0.0
+                    else:
+                        acc[yi, xi] = 0.0
+        return acc
+
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_window_u16(
+        spread: np.ndarray,      # uint16 (H, W)
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint16,
+        bg: np.ndarray,
+        y0: nb.int64, x0: nb.int64,
+        wh: nb.int64, ww: nb.int64,
+    ) -> np.ndarray:
+        """Versione uint16 (polarity 16-bin) del kernel windowed."""
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((wh, ww), dtype=np.float32)
+        for yi in nb.prange(wh):
+            y = y0 + yi
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint16(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                yy = y + dy[i]
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                xi_lo = 0
+                xi_hi = ww
+                lo = -(x0 + ddx)
+                if lo > xi_lo:
+                    xi_lo = lo
+                hi = W - (x0 + ddx)
+                if hi < xi_hi:
+                    xi_hi = hi
+                for xi in range(xi_lo, xi_hi):
+                    if spread[yy, x0 + xi + ddx] & mask:
+                        acc[yi, xi] += 1.0
+        if N > 0:
+            inv = 1.0 / N
+            for yi in nb.prange(wh):
+                for xi in range(ww):
+                    v = acc[yi, xi] * inv
+                    bgv = bg[y0 + yi, x0 + xi]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[yi, xi] = r if r > 0.0 else 0.0
+                    else:
+                        acc[yi, xi] = 0.0
+        return acc
+
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_top_max_per_variant(
+        spread: np.ndarray,            # uint8 (H, W)
+        dx_flat: np.ndarray,           # int32 (sum_N,)
+        dy_flat: np.ndarray,           # int32 (sum_N,)
+        bins_flat: np.ndarray,         # int8 (sum_N,)
+        offsets: np.ndarray,           # int32 (n_vars+1,) prefix sum
+        bit_active: np.uint8,
+        bg_per_variant: np.ndarray,    # float32 (n_vars, H, W) - 1 per scala
+        scale_idx: np.ndarray,         # int32 (n_vars,) idx in bg_per_variant
+    ) -> np.ndarray:
+        """Batch: per ogni variante calcola max score (rescored bg), ritorna
+        array float32 (n_vars,). Parallelismo prange ESTERNO sulle varianti
+        elimina overhead di n_vars chiamate JIT separate (avg ~20us per
+        chiamata su template piccoli) + pool thread Python.
+
+        Pensato per fase TOP del pruning quando n_vars >> n_threads.
+        """
+        n_vars = offsets.shape[0] - 1
+        H, W = spread.shape
+        out = np.zeros(n_vars, dtype=np.float32)
+        for vi in nb.prange(n_vars):
+            i0 = offsets[vi]; i1 = offsets[vi + 1]
+            N = i1 - i0
+            if N == 0:
+                out[vi] = -1.0
+                continue
+            si = scale_idx[vi]
+            inv = nb.float32(1.0 / N)
+            best = nb.float32(-1.0)
+            for y in range(H):
+                for x in range(W):
+                    s = nb.float32(0.0)
+                    for k in range(N):
+                        b = bins_flat[i0 + k]
+                        mask = np.uint8(1) << b
+                        if (bit_active & mask) == 0:
+                            continue
+                        ddy = dy_flat[i0 + k]
+                        yy = y + ddy
+                        if yy < 0 or yy >= H:
+                            continue
+                        ddx = dx_flat[i0 + k]
+                        xx = x + ddx
+                        if xx < 0 or xx >= W:
+                            continue
+                        if spread[yy, xx] & mask:
+                            s += nb.float32(1.0)
+                    s *= inv
+                    bgv = bg_per_variant[si, y, x]
+                    if bgv < 1.0:
+                        r = (s - bgv) / (1.0 - bgv + 1e-6)
+                        if r > best:
+                            best = r
+            out[vi] = best if best > 0.0 else 0.0
+        return out
+
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_u16(
+        spread: np.ndarray,      # uint16 (H, W) - 16 bit di polarity-aware
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint16,
+        bg: np.ndarray,
+    ) -> np.ndarray:
+        """Versione uint16 di _jit_score_bitmap_rescored per polarity 16-bin.
+
+        Identica logica ma mask = uint16(1) << b dove b in [0..15]
+        (orientamento mod 2π invece di mod π).
+        """
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((H, W), dtype=np.float32)
+        for y in nb.prange(H):
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint16(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                ddy = dy[i]
+                yy = y + ddy
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                x_lo = 0 if ddx >= 0 else -ddx
+                x_hi = W if ddx <= 0 else W - ddx
+                for x in range(x_lo, x_hi):
+                    if spread[yy, x + ddx] & mask:
+                        acc[y, x] += 1.0
+        if N > 0:
+            inv = 1.0 / N
+            for y in nb.prange(H):
+                for x in range(W):
+                    v = acc[y, x] * inv
+                    bgv = bg[y, x]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[y, x] = r if r > 0.0 else 0.0
+                    else:
+                        acc[y, x] = 0.0
+        return acc
+
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_popcount_density_u16(spread: np.ndarray) -> np.ndarray:
+        """Popcount per uint16 (16 bin polarity)."""
+        H, W = spread.shape
+        out = np.zeros((H, W), dtype=np.float32)
+        for y in nb.prange(H):
+            for x in range(W):
+                v = spread[y, x]
+                cnt = 0
+                for b in range(16):
+                    if v & (np.uint16(1) << b):
+                        cnt += 1
+                out[y, x] = float(cnt)
+        return out
+
     @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
     def _jit_popcount_density(spread: np.ndarray) -> np.ndarray:
         """Conta bit set per pixel: ritorna (H, W) float32 in [0..8]."""
@@ -241,11 +515,28 @@ if HAS_NUMBA:
         _jit_score_bitmap(spread, dx, dy, b, np.uint8(0xFF))
         bg = np.zeros((32, 32), dtype=np.float32)
         _jit_score_bitmap_rescored(spread, dx, dy, b, np.uint8(0xFF), bg)
+        _jit_score_bitmap_rescored_strided(
+            spread, dx, dy, b, np.uint8(0xFF), bg, np.int32(2),
+        )
         _jit_score_bitmap_greedy(
             spread, dx, dy, b, np.uint8(0xFF),
             np.float32(0.5), np.float32(0.8),
         )
+        offsets = np.array([0, 1], dtype=np.int32)
+        scale_idx = np.zeros(1, dtype=np.int32)
+        bg_pv = np.zeros((1, 32, 32), dtype=np.float32)
+        _jit_top_max_per_variant(
+            spread, dx, dy, b, offsets, np.uint8(0xFF), bg_pv, scale_idx,
+        )
+        _jit_score_bitmap_rescored_window(
+            spread, dx, dy, b, np.uint8(0xFF), bg, 4, 4, 8, 8,
+        )
         _jit_popcount_density(spread)
+        spread16 = np.zeros((32, 32), dtype=np.uint16)
+        _jit_score_bitmap_rescored_u16(
+            spread16, dx, dy, b, np.uint16(0xFFFF), bg,
+        )
+        _jit_popcount_density_u16(spread16)
 
 else:  # pragma: no cover
 
@@ -258,9 +549,30 @@ else:  # pragma: no cover
     def _jit_score_bitmap_rescored(spread, dx, dy, bins, bit_active, bg):
         raise RuntimeError("numba non disponibile")
 
+    def _jit_score_bitmap_rescored_strided(spread, dx, dy, bins, bit_active, bg, stride):
+        raise RuntimeError("numba non disponibile")
+
+    def _jit_score_bitmap_rescored_window(spread, dx, dy, bins, bit_active, bg, y0, x0, wh, ww):
+        raise RuntimeError("numba non disponibile")
+
+    def _jit_score_bitmap_rescored_window_u16(spread, dx, dy, bins, bit_active, bg, y0, x0, wh, ww):
+        raise RuntimeError("numba non disponibile")
+
     def _jit_score_bitmap_greedy(spread, dx, dy, bins, bit_active, min_score, greediness):
         raise RuntimeError("numba non disponibile")
 
+    def _jit_top_max_per_variant(
+        spread, dx_flat, dy_flat, bins_flat, offsets, bit_active,
+        bg_per_variant, scale_idx,
+    ):
+        raise RuntimeError("numba non disponibile")
+
+    def _jit_score_bitmap_rescored_u16(spread, dx, dy, bins, bit_active, bg):
+        raise RuntimeError("numba non disponibile")
+
+    def _jit_popcount_density_u16(spread):
+        raise RuntimeError("numba non disponibile")
+
     def _jit_popcount_density(spread):
         raise RuntimeError("numba non disponibile")
 
@@ -291,24 +603,71 @@ def score_bitmap(
 
 def score_bitmap_rescored(
     spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
-    bit_active: int, bg: np.ndarray,
+    bit_active: int, bg: np.ndarray, stride: int = 1,
 ) -> np.ndarray:
-    """Score bitmap + rescore fusi in un solo pass (JIT)."""
+    """Score bitmap + rescore fusi in un solo pass (JIT).
+
+    Dispatch per dtype: uint16 → kernel polarity 16-bin, uint8 → kernel
+    standard 8-bin (con eventuale stride > 1 per coarse top-level).
+    """
     if HAS_NUMBA and len(dx) > 0:
+        dx_c = np.ascontiguousarray(dx, dtype=np.int32)
+        dy_c = np.ascontiguousarray(dy, dtype=np.int32)
+        bins_c = np.ascontiguousarray(bins, dtype=np.int8)
+        bg_c = np.ascontiguousarray(bg, dtype=np.float32)
+        if spread.dtype == np.uint16:
+            spread_c = np.ascontiguousarray(spread, dtype=np.uint16)
+            return _jit_score_bitmap_rescored_u16(
+                spread_c, dx_c, dy_c, bins_c, np.uint16(bit_active), bg_c,
+            )
+        spread_c = np.ascontiguousarray(spread, dtype=np.uint8)
+        if stride > 1:
+            return _jit_score_bitmap_rescored_strided(
+                spread_c, dx_c, dy_c, bins_c, np.uint8(bit_active), bg_c,
+                np.int32(stride),
+            )
         return _jit_score_bitmap_rescored(
-            np.ascontiguousarray(spread, dtype=np.uint8),
-            np.ascontiguousarray(dx, dtype=np.int32),
-            np.ascontiguousarray(dy, dtype=np.int32),
-            np.ascontiguousarray(bins, dtype=np.int8),
-            np.uint8(bit_active),
-            np.ascontiguousarray(bg, dtype=np.float32),
+            spread_c, dx_c, dy_c, bins_c, np.uint8(bit_active), bg_c,
         )
-    # Fallback: chiamate separate
+    # Fallback: chiamate separate (stride ignorato in fallback)
     score = score_bitmap(spread, dx, dy, bins, bit_active)
     out = (score - bg) / (1.0 - bg + 1e-6)
     return np.maximum(0.0, out).astype(np.float32)
 
 
+def score_bitmap_rescored_window(
+    spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+    bit_active: int, bg: np.ndarray,
+    y0: int, x0: int, wh: int, ww: int,
+) -> np.ndarray:
+    """Score rescored solo nella finestra (y0, x0, wh, ww) della scena.
+
+    Le feature campionano l'INTERA scena: feature fuori finestra ma dentro
+    scena contano correttamente (chiamare il kernel su un crop le tratta
+    come miss e azzera lo score — il bug che rendeva inutilizzabile il
+    path pyramid_propagate). Fallback no-numba: kernel pieno + slice.
+    """
+    if HAS_NUMBA and len(dx) > 0:
+        dx_c = np.ascontiguousarray(dx, dtype=np.int32)
+        dy_c = np.ascontiguousarray(dy, dtype=np.int32)
+        bins_c = np.ascontiguousarray(bins, dtype=np.int8)
+        bg_c = np.ascontiguousarray(bg, dtype=np.float32)
+        if spread.dtype == np.uint16:
+            return _jit_score_bitmap_rescored_window_u16(
+                np.ascontiguousarray(spread, dtype=np.uint16),
+                dx_c, dy_c, bins_c, np.uint16(bit_active), bg_c,
+                int(y0), int(x0), int(wh), int(ww),
+            )
+        return _jit_score_bitmap_rescored_window(
+            np.ascontiguousarray(spread, dtype=np.uint8),
+            dx_c, dy_c, bins_c, np.uint8(bit_active), bg_c,
+            int(y0), int(x0), int(wh), int(ww),
+        )
+    # Fallback (lento, solo senza numba): score full-frame + slice finestra
+    full = score_bitmap_rescored(spread, dx, dy, bins, bit_active, bg)
+    return full[y0:y0 + wh, x0:x0 + ww]
+
+
 def score_bitmap_greedy(
     spread: np.ndarray, dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
     bit_active: int, min_score: float, greediness: float,
@@ -331,10 +690,78 @@ def score_bitmap_greedy(
     return score_bitmap(spread, dx, dy, bins, bit_active)
 
 
+def top_max_per_variant(
+    spread: np.ndarray,
+    dx_list: list, dy_list: list, bin_list: list,
+    bg_per_scale: dict,
+    variant_scales: list,
+    bit_active: int,
+) -> np.ndarray:
+    """Wrapper: prepara buffer flat e chiama kernel batch su tutte le varianti.
+
+    Parallelismo Numba prange-esterno sulle varianti (n_vars >> n_threads
+    tipicamente per top-pruning) → meglio del thread-pool Python che paga
+    overhead di n_vars chiamate JIT separate.
+    """
+    if not HAS_NUMBA or len(dx_list) == 0:
+        return np.array([], dtype=np.float32)
+    n_vars = len(dx_list)
+    sizes = [len(d) for d in dx_list]
+    offsets = np.zeros(n_vars + 1, dtype=np.int32)
+    offsets[1:] = np.cumsum(sizes)
+    total = int(offsets[-1])
+    dx_flat = np.empty(total, dtype=np.int32)
+    dy_flat = np.empty(total, dtype=np.int32)
+    bins_flat = np.empty(total, dtype=np.int8)
+    for vi, (dx, dy, bn) in enumerate(zip(dx_list, dy_list, bin_list)):
+        i0 = int(offsets[vi]); i1 = int(offsets[vi + 1])
+        dx_flat[i0:i1] = dx
+        dy_flat[i0:i1] = dy
+        bins_flat[i0:i1] = bn
+    # bg per variante: indicizzato per scala
+    scales_unique = sorted(bg_per_scale.keys())
+    scale_to_idx = {s: i for i, s in enumerate(scales_unique)}
+    H, W = spread.shape
+    bg_pv = np.empty((len(scales_unique), H, W), dtype=np.float32)
+    for s, idx in scale_to_idx.items():
+        bg_pv[idx] = bg_per_scale[s]
+    scale_idx = np.array(
+        [scale_to_idx[s] for s in variant_scales], dtype=np.int32,
+    )
+    return _jit_top_max_per_variant(
+        np.ascontiguousarray(spread, dtype=np.uint8),
+        dx_flat, dy_flat, bins_flat, offsets, np.uint8(bit_active),
+        bg_pv, scale_idx,
+    )
+
+
+_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
+    """
+    if spread.dtype == np.uint16:
+        spread_c = np.ascontiguousarray(spread, dtype=np.uint16)
+        if HAS_NUMBA:
+            return _jit_popcount_density_u16(spread_c)
+        if _HAS_NP_BITCOUNT:
+            return np.bitwise_count(spread_c).astype(np.float32, copy=False)
+        H, W = spread_c.shape
+        out = np.zeros((H, W), dtype=np.float32)
+        for b in range(16):
+            out += ((spread_c >> b) & 1).astype(np.float32)
+        return out
+    spread_c = np.ascontiguousarray(spread, dtype=np.uint8)
     if HAS_NUMBA:
-        return _jit_popcount_density(np.ascontiguousarray(spread, dtype=np.uint8))
-    # Fallback
+        return _jit_popcount_density(spread_c)
+    if _HAS_NP_BITCOUNT:
+        return np.bitwise_count(spread_c).astype(np.float32, copy=False)
     H, W = spread.shape
     out = np.zeros((H, W), dtype=np.float32)
     for b in range(8):

pm2d/auto_tune.py

@@ -61,6 +61,8 @@ def detect_rotational_symmetry(
 
     center = (w / 2.0, h / 2.0)
     ref = mag
+    # ref è costante nel loop sugli angoli: centra una volta sola
+    rm = ref - ref.mean()
 
     correlations: list[tuple[float, float]] = []
     for ang in np.arange(step_deg, 360.0, step_deg):
@@ -68,7 +70,6 @@ def detect_rotational_symmetry(
         rot = cv2.warpAffine(
             mag, M, (w, h), borderValue=0.0,
         )
-        rm = ref - ref.mean()
         rs = rot - rot.mean()
         denom = np.sqrt((rm * rm).sum() * (rs * rs).sum()) + 1e-9
         c = float((rm * rs).sum() / denom)
@@ -152,14 +153,124 @@ def _cache_key(template_bgr: np.ndarray, mask: np.ndarray | None) -> str:
     return h.hexdigest()
 
 
-def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
+def _self_validate(template_bgr: np.ndarray, params: dict,
+                   mask: np.ndarray | None = None) -> dict:
+    """Halcon-style self-validation: train il matcher coi parametri tentativi
+    e verifica che il template stesso sia trovato con recall ≥ 1.0.
+
+    Se recall < target o score basso, regola i parametri:
+    - alza weak_grad se troppi edge spuri (recall solido ma molti picchi falsi)
+    - abbassa strong_grad se troppe feature scartate (low feature count)
+    - riduce pyramid_levels se variants[0].levels[top] ha <8 feature
+
+    Halcon usa internamente questo loop in inspect_shape_model. Costo: 1
+    train + 1 find sul template (~50ms su template 100x100). Ne vale la
+    pena se evita match-time errors su scene reali.
+
+    Mutates `params` in place e ritorna lo stesso dict per chaining.
+    """
+    # Import lazy: evita ciclo (line_matcher importa nulla da auto_tune)
+    from pm2d.line_matcher import LineShapeMatcher
+
+    # Caso degenerato: troppe poche feature pre-validation → riduci soglia
+    if params.get("_n_strong_pixels", 0) < 30:
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.6)
+        params["strong_grad"] = max(30.0, params["strong_grad"] * 0.6)
+
+    # Train minimale: 1 sola pose orientazione 0 (range degenerato che
+    # produce comunque 1 variante via fallback in _angle_list).
+    m = LineShapeMatcher(
+        num_features=params["num_features"],
+        weak_grad=params["weak_grad"],
+        strong_grad=params["strong_grad"],
+        angle_range_deg=(0.0, 0.0),  # fallback _angle_list = [0.0]
+        angle_step_deg=10.0,
+        scale_range=(1.0, 1.0),
+        spread_radius=params["spread_radius"],
+        pyramid_levels=params["pyramid_levels"],
+    )
+    n_var = m.train(template_bgr, mask=mask)
+    if n_var == 0:
+        # Soglie troppo alte: nessuna variante generata → dimezza
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.5)
+        params["strong_grad"] = max(30.0, params["strong_grad"] * 0.5)
+        params["_validation"] = "fallback: soglie dimezzate (no variants)"
+        return params
+
+    # Verifica densita' feature al top-level (rischio collasso)
+    top_lvl = m.variants[0].levels[-1]
+    if top_lvl.n < 8 and params["pyramid_levels"] > 1:
+        params["pyramid_levels"] = max(1, params["pyramid_levels"] - 1)
+        params["_validation"] = (
+            f"pyramid_levels ridotto a {params['pyramid_levels']} "
+            f"(top aveva {top_lvl.n} feature)"
+        )
+        return params
+
+    # Self-find: cerca il template stesso nella propria immagine
+    h, w = template_bgr.shape[:2]
+    # Embed template in scena leggermente più grande per evitare bordo
+    pad = 20
+    canvas = np.full(
+        (h + 2 * pad, w + 2 * pad, 3 if template_bgr.ndim == 3 else 1),
+        128, dtype=np.uint8,
+    )
+    canvas[pad:pad + h, pad:pad + w] = template_bgr
+    matches = m.find(
+        canvas, min_score=0.3, max_matches=5,
+        verify_ncc=False,  # template stesso → NCC = 1 sempre, skip per velocita'
+        refine_angle=False, subpixel=False,
+        nms_iou_threshold=0.3,
+    )
+    if not matches:
+        # Nessun match sul proprio template: parametri troppo restrittivi
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.7)
+        params["strong_grad"] = max(30.0, params["strong_grad"] * 0.7)
+        params["num_features"] = max(48, int(params["num_features"] * 0.8))
+        params["_validation"] = "soglie/feature ridotte (no self-match)"
+        return params
+
+    # Misura score top match
+    top_score = float(matches[0].score)
+    params["_self_score"] = round(top_score, 3)
+    if top_score < 0.7:
+        # Score basso sul template stesso = parametri davvero subottimali
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.85)
+        params["_validation"] = (
+            f"weak_grad ridotto (self-score era {top_score:.2f})"
+        )
+    else:
+        params["_validation"] = f"OK (self-score {top_score:.2f})"
+    return params
+
+
+def auto_tune(
+    template_bgr: np.ndarray,
+    mask: np.ndarray | None = None,
+    angle_tolerance_deg: float | None = None,
+    angle_center_deg: float = 0.0,
+    self_validate: bool = True,
+) -> dict:
     """Analizza template e ritorna dict parametri suggeriti.
 
     Chiavi compatibili con edit_params PARAM_SCHEMA.
 
+    angle_tolerance_deg: se != None, restringe angle_range a
+    (center - tol, center + tol). Usare quando l'orientamento del
+    pezzo e' noto a priori (feeder con guida, posizionamento
+    meccanico): training molto piu rapido (24x meno varianti per
+    tol=15° vs 360° pieno).
+
+    self_validate: se True (default), dopo la stima dei parametri
+    esegue un dry-run del matching sul template stesso e regola
+    weak_grad/strong_grad/pyramid_levels se i parametri tentativi
+    non garantiscono auto-match (Halcon-style inspect_shape_model).
+
     Risultato cachato in-memory (LRU): ri-chiamare con stessa ROI è O(1).
     """
     ck = _cache_key(template_bgr, mask)
+    if angle_tolerance_deg is not None:
+        ck = f"{ck}|tol={angle_tolerance_deg}|c={angle_center_deg}"
     cached = _TUNE_CACHE.get(ck)
     if cached is not None:
         _TUNE_CACHE.move_to_end(ck)
@@ -208,8 +319,13 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
     # spread_radius proporzionale a risoluzione + pyramid (tolleranza ~1% dim)
     spread_radius = int(np.clip(max(3, min_side * 0.02), 3, 8))
 
-    # angle range ridotto se simmetria rotazionale
-    angle_max = 360.0 / sym["order"] if sym["order"] > 1 else 360.0
+    # angle range: priorita' a tolerance hint utente, poi simmetria rotazionale.
+    if angle_tolerance_deg is not None:
+        angle_min = float(angle_center_deg - angle_tolerance_deg)
+        angle_max = float(angle_center_deg + angle_tolerance_deg)
+    else:
+        angle_min = 0.0
+        angle_max = 360.0 / sym["order"] if sym["order"] > 1 else 360.0
 
     # min_score: se entropia orient alta → template distintivo → soglia alta ok
     # se entropia bassa → template ambiguo → soglia più permissiva
@@ -220,12 +336,15 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
     else:
         min_score = 0.45
 
-    # angle step: 5° default; se simmetria, mantengo step ma range ridotto
-    angle_step = 5.0
+    # angle step adattivo (Halcon-style): atan(2/max_side) deg, clampato.
+    # Template grande → step fine (rotazione minima visibile su perimetro).
+    # Template piccolo → step grosso (over-sampling = sprecato).
+    max_side = max(h, w)
+    angle_step = float(np.clip(np.degrees(np.arctan2(2.0, max_side)), 1.0, 8.0))
 
     result = {
         "backend": "line",
-        "angle_min": 0.0,
+        "angle_min": angle_min,
         "angle_max": angle_max,
         "angle_step": angle_step,
         "scale_min": 1.0,
@@ -244,7 +363,15 @@ def auto_tune(template_bgr: np.ndarray, mask: np.ndarray | None = None) -> dict:
         "_symmetry_order": sym["order"],
         "_symmetry_conf": round(sym["confidence"], 2),
         "_orient_entropy": round(stats["orient_entropy"], 2),
+        "_n_strong_pixels": stats["n_strong"],
     }
+    # Halcon-style self-validation: dry-run training+find sul template per
+    # auto-correggere parametri tentativi che non garantirebbero match.
+    if self_validate:
+        result = _self_validate(template_bgr, result, mask=mask)
+        # Round numerici dopo eventuali aggiustamenti
+        result["weak_grad"] = round(result["weak_grad"], 1)
+        result["strong_grad"] = round(result["strong_grad"], 1)
     # Store in LRU cache
     _TUNE_CACHE[ck] = dict(result)
     _TUNE_CACHE.move_to_end(ck)

pm2d/bench.py

@@ -0,0 +1,179 @@
+"""Benchmark suite per LineShapeMatcher.
+
+Usage:
+    python -m pm2d.bench [--quick]
+
+Misura tempi find() su 3 template-tipo × 3 scene-tipo × N config:
+- Template: rettangolo 80×80, L-shape 120×120, cerchio 150×150
+- Scene: pulita 800×600, cluttered 1080×1920, multi-pezzo 1080×1920
+- Config: baseline, polarity, gpu, pyramid_propagate, greediness=0.7
+
+Per ogni config stampa: ms/find, ms per fase (profile), n. match.
+Output tabellare per detectare regressioni in CI.
+"""
+
+from __future__ import annotations
+
+import argparse
+import time
+
+import cv2
+import numpy as np
+
+from pm2d.line_matcher import LineShapeMatcher, opencl_available
+
+
+# ---------- Sintetizzatori template/scena ----------
+
+def _tpl_rect() -> np.ndarray:
+    t = np.zeros((80, 80, 3), np.uint8)
+    cv2.rectangle(t, (15, 15), (65, 65), (255, 255, 255), 3)
+    return t
+
+
+def _tpl_lshape() -> np.ndarray:
+    t = np.zeros((120, 120, 3), np.uint8)
+    cv2.rectangle(t, (20, 20), (50, 100), (255, 255, 255), -1)
+    cv2.rectangle(t, (20, 70), (100, 100), (255, 255, 255), -1)
+    return t
+
+
+def _tpl_circle() -> np.ndarray:
+    t = np.zeros((150, 150, 3), np.uint8)
+    cv2.circle(t, (75, 75), 60, (255, 255, 255), 4)
+    return t
+
+
+def _scene_clean(W: int, H: int, n_pieces: int = 1) -> np.ndarray:
+    np.random.seed(0)
+    s = np.zeros((H, W, 3), np.uint8)
+    for _ in range(n_pieces):
+        cx = np.random.randint(80, W - 80)
+        cy = np.random.randint(80, H - 80)
+        cv2.rectangle(s, (cx - 25, cy - 25), (cx + 25, cy + 25), (255, 255, 255), 3)
+    return s
+
+
+def _scene_cluttered(W: int, H: int) -> np.ndarray:
+    np.random.seed(0)
+    s = np.random.randint(50, 200, (H, W, 3), np.uint8)
+    cv2.rectangle(s, (300, 200), (350, 250), (255, 255, 255), 3)
+    cv2.rectangle(s, (1500, 800), (1550, 850), (255, 255, 255), 3)
+    return s
+
+
+# ---------- Single benchmark ----------
+
+def _bench_config(template, scene, config_name: str,
+                  init_kw: dict, find_kw: dict,
+                  n_iter: int = 5) -> dict:
+    m = LineShapeMatcher(**init_kw)
+    t0 = time.perf_counter()
+    n_var = m.train(template)
+    t_train = time.perf_counter() - t0
+
+    # Warmup (Numba JIT)
+    m.find(scene, **find_kw)
+    m.find(scene, **find_kw)
+
+    # Run
+    times_ms = []
+    for _ in range(n_iter):
+        t0 = time.perf_counter()
+        matches = m.find(scene, **find_kw)
+        times_ms.append((time.perf_counter() - t0) * 1000.0)
+
+    # Profile (1 iter)
+    m.find(scene, profile=True, **find_kw)
+    prof = m.get_last_profile() or {}
+
+    return {
+        "config": config_name,
+        "n_variants": n_var,
+        "t_train_s": round(t_train, 3),
+        "ms_avg": round(float(np.mean(times_ms)), 1),
+        "ms_min": round(float(np.min(times_ms)), 1),
+        "ms_max": round(float(np.max(times_ms)), 1),
+        "n_matches": len(matches),
+        "profile_ms": {k: round(v, 1) for k, v in prof.items()},
+    }
+
+
+# ---------- Suite ----------
+
+CONFIGS = [
+    ("baseline",
+     {"angle_step_deg": 10, "pyramid_levels": 2},
+     {"min_score": 0.4, "verify_threshold": 0.2}),
+    ("polarity",
+     {"angle_step_deg": 10, "pyramid_levels": 2, "use_polarity": True},
+     {"min_score": 0.4, "verify_threshold": 0.2}),
+    ("propagate",
+     {"angle_step_deg": 10, "pyramid_levels": 3},
+     {"min_score": 0.4, "verify_threshold": 0.2,
+      "pyramid_propagate": True, "propagate_topk": 4}),
+    ("greedy_07",
+     {"angle_step_deg": 10, "pyramid_levels": 2},
+     {"min_score": 0.4, "verify_threshold": 0.2, "greediness": 0.7}),
+    ("stride2",
+     {"angle_step_deg": 10, "pyramid_levels": 2},
+     {"min_score": 0.4, "verify_threshold": 0.2, "coarse_stride": 2}),
+]
+
+if opencl_available():
+    CONFIGS.append(
+        ("gpu_umat",
+         {"angle_step_deg": 10, "pyramid_levels": 2, "use_gpu": True},
+         {"min_score": 0.4, "verify_threshold": 0.2})
+    )
+
+
+SCENARIOS = [
+    ("rect_80 vs scene_800x600", _tpl_rect, lambda: _scene_clean(800, 600, 1)),
+    ("lshape_120 vs scene_1080x1920_clutter",
+     _tpl_lshape, lambda: _scene_cluttered(1920, 1080)),
+    ("circle_150 vs scene_clean_3pieces",
+     _tpl_circle, lambda: _scene_clean(1920, 1080, 3)),
+]
+
+
+def run(quick: bool = False) -> int:
+    n_iter = 2 if quick else 5
+    print(f"=== PM2D Benchmark Suite ({len(SCENARIOS)} scenarios x "
+          f"{len(CONFIGS)} configs, n_iter={n_iter}) ===\n")
+    rows = []
+    for sc_name, tpl_fn, scn_fn in SCENARIOS:
+        template = tpl_fn()
+        scene = scn_fn()
+        print(f"--- Scenario: {sc_name} (tpl={template.shape}, "
+              f"scn={scene.shape}) ---")
+        for cfg_name, init_kw, find_kw in CONFIGS:
+            r = _bench_config(template, scene, cfg_name, init_kw, find_kw,
+                              n_iter=n_iter)
+            r["scenario"] = sc_name
+            rows.append(r)
+            prof_str = " ".join(
+                f"{k}={v:.1f}" for k, v in r["profile_ms"].items()
+            )
+            print(f"  {cfg_name:14s}  {r['ms_avg']:6.1f}ms  "
+                  f"(min {r['ms_min']:.1f} max {r['ms_max']:.1f})  "
+                  f"vars={r['n_variants']:3d}  "
+                  f"matches={r['n_matches']:2d}")
+            if prof_str:
+                print(f"     profile: {prof_str}")
+        print()
+    print("=== Done ===")
+    return 0
+
+
+def main(argv: list[str] | None = None) -> int:
+    p = argparse.ArgumentParser(description="PM2D benchmark suite")
+    p.add_argument("--quick", action="store_true",
+                   help="2 iterazioni per config invece di 5 (smoke test)")
+    args = p.parse_args(argv)
+    return run(quick=args.quick)
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(main())

pm2d/eval.py

@@ -0,0 +1,217 @@
+"""CLI validation harness per LineShapeMatcher.
+
+Usage:
+    python -m pm2d.eval dataset.json [opzioni]
+
+Formato dataset (JSON):
+    {
+      "template": "path/to/template.png",
+      "mask": "path/to/mask.png",  # opzionale
+      "params": {                   # opzionali, override su matcher init
+        "use_polarity": true,
+        "angle_step_deg": 5,
+        ...
+      },
+      "find_params": {              # opzionali, passati a find()
+        "min_score": 0.6,
+        "use_soft_score": true,
+        ...
+      },
+      "scenes": [
+        {
+          "image": "path/to/scene1.png",
+          "ground_truth": [
+            {"cx": 320.0, "cy": 240.0, "angle_deg": 12.0,
+             "scale": 1.0, "tolerance_px": 5.0,
+             "tolerance_deg": 3.0}
+          ]
+        }
+      ]
+    }
+
+Output: report precision/recall/IoU/timing per ogni scena + aggregati.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import math
+import sys
+import time
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+from pm2d.line_matcher import LineShapeMatcher, _poly_iou, _oriented_bbox_polygon
+
+
+def _load_image(path: str | Path) -> np.ndarray:
+    img = cv2.imread(str(path), cv2.IMREAD_UNCHANGED)
+    if img is None:
+        raise FileNotFoundError(f"Immagine non trovata: {path}")
+    if img.ndim == 2:
+        img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
+    return img
+
+
+def _gt_to_poly(gt: dict, tw: int, th: int) -> np.ndarray:
+    """Costruisce bbox poligonale per un ground truth."""
+    s = float(gt.get("scale", 1.0))
+    return _oriented_bbox_polygon(
+        float(gt["cx"]), float(gt["cy"]),
+        tw * s, th * s, float(gt["angle_deg"]),
+    )
+
+
+def _match_to_gt(match, gt: dict, tw: int, th: int,
+                 iou_thr: float = 0.3) -> bool:
+    """True se il match corrisponde al ground truth.
+
+    Criterio: distanza centro <= tolerance_px AND |angle_deg - gt| <= tolerance_deg
+    OR IoU bbox >= iou_thr (fallback per pose con tolerance ampie).
+    """
+    tol_px = float(gt.get("tolerance_px", 5.0))
+    tol_deg = float(gt.get("tolerance_deg", 3.0))
+    dx = match.cx - float(gt["cx"])
+    dy = match.cy - float(gt["cy"])
+    dist = math.hypot(dx, dy)
+    da = abs((match.angle_deg - float(gt["angle_deg"]) + 180) % 360 - 180)
+    if dist <= tol_px and da <= tol_deg:
+        return True
+    # Fallback IoU
+    poly_gt = _gt_to_poly(gt, tw, th)
+    poly_m = match.bbox_poly
+    if _poly_iou(poly_m, poly_gt) >= iou_thr:
+        return True
+    return False
+
+
+def evaluate_scene(matcher: LineShapeMatcher, scene_bgr: np.ndarray,
+                   gt_list: list[dict], find_params: dict,
+                   tw: int, th: int) -> dict:
+    """Esegue match e calcola TP/FP/FN per una scena."""
+    t0 = time.time()
+    matches = matcher.find(scene_bgr, **find_params)
+    elapsed = time.time() - t0
+
+    gt_matched = [False] * len(gt_list)
+    match_is_tp = [False] * len(matches)
+    iou_per_match = [0.0] * len(matches)
+    for i, m in enumerate(matches):
+        for j, gt in enumerate(gt_list):
+            if gt_matched[j]:
+                continue
+            if _match_to_gt(m, gt, tw, th):
+                gt_matched[j] = True
+                match_is_tp[i] = True
+                # Calcolo IoU per metrica
+                poly_gt = _gt_to_poly(gt, tw, th)
+                iou_per_match[i] = _poly_iou(m.bbox_poly, poly_gt)
+                break
+    tp = sum(match_is_tp)
+    fp = len(matches) - tp
+    fn = len(gt_list) - sum(gt_matched)
+    return {
+        "n_matches": len(matches),
+        "n_gt": len(gt_list),
+        "tp": tp, "fp": fp, "fn": fn,
+        "find_time_s": elapsed,
+        "iou_mean": float(np.mean([i for i, t in zip(iou_per_match, match_is_tp) if t])
+                          if tp > 0 else 0.0),
+        "diag": (matcher.get_last_diag()
+                 if hasattr(matcher, "get_last_diag") else None),
+    }
+
+
+def run(dataset_path: str, scene_filter: str | None = None,
+        verbose: bool = False) -> dict:
+    """Esegue eval su dataset, ritorna report aggregato."""
+    dataset_path = Path(dataset_path)
+    base = dataset_path.parent
+    with open(dataset_path) as f:
+        ds = json.load(f)
+
+    template = _load_image(base / ds["template"])
+    mask = None
+    if ds.get("mask"):
+        mask_img = cv2.imread(str(base / ds["mask"]), cv2.IMREAD_GRAYSCALE)
+        if mask_img is not None:
+            mask = (mask_img > 128).astype(np.uint8) * 255
+    init_params = ds.get("params", {})
+    find_params = ds.get("find_params", {})
+
+    matcher = LineShapeMatcher(**init_params)
+    n_var = matcher.train(template, mask=mask)
+    tw, th = matcher.template_size
+    print(f"Template: {ds['template']} ({tw}x{th}), {n_var} varianti")
+    print(f"Param matcher: {init_params}")
+    print(f"Param find:    {find_params}")
+    print()
+
+    scenes = ds["scenes"]
+    if scene_filter:
+        scenes = [s for s in scenes if scene_filter in s["image"]]
+
+    rows = []
+    tot_tp = tot_fp = tot_fn = 0
+    tot_time = 0.0
+    for sc in scenes:
+        scene = _load_image(base / sc["image"])
+        gt = sc.get("ground_truth", [])
+        result = evaluate_scene(matcher, scene, gt, find_params, tw, th)
+        rows.append({"scene": sc["image"], **result})
+        tot_tp += result["tp"]; tot_fp += result["fp"]; tot_fn += result["fn"]
+        tot_time += result["find_time_s"]
+        prec = result["tp"] / max(1, result["tp"] + result["fp"])
+        rec = result["tp"] / max(1, result["tp"] + result["fn"])
+        line = (f"  {sc['image']:30s}  "
+                f"TP={result['tp']} FP={result['fp']} FN={result['fn']}  "
+                f"P={prec:.2f} R={rec:.2f}  "
+                f"IoU={result['iou_mean']:.2f}  "
+                f"t={result['find_time_s']*1000:.0f}ms")
+        print(line)
+        if verbose and result["diag"] and hasattr(matcher, "_format_diag"):
+            print(f"     diag: {matcher._format_diag(result['diag'])}")
+
+    # Aggregati
+    precision = tot_tp / max(1, tot_tp + tot_fp)
+    recall = tot_tp / max(1, tot_tp + tot_fn)
+    f1 = 2 * precision * recall / max(1e-9, precision + recall)
+    print()
+    print(f"AGGREGATO: precision={precision:.3f} recall={recall:.3f} "
+          f"F1={f1:.3f}  TP={tot_tp} FP={tot_fp} FN={tot_fn}")
+    print(f"TIME: total={tot_time:.2f}s avg={tot_time / max(1, len(scenes)) * 1000:.0f}ms/scene")
+
+    return {
+        "precision": precision, "recall": recall, "f1": f1,
+        "tp": tot_tp, "fp": tot_fp, "fn": tot_fn,
+        "total_time_s": tot_time, "n_scenes": len(scenes),
+        "per_scene": rows,
+    }
+
+
+def main(argv: list[str] | None = None) -> int:
+    p = argparse.ArgumentParser(
+        description="pm2d-eval: validation harness per LineShapeMatcher"
+    )
+    p.add_argument("dataset", help="JSON dataset (template + scenes + GT)")
+    p.add_argument("--scene-filter", default=None,
+                   help="Filtro substring sui nomi scena (debug)")
+    p.add_argument("--verbose", "-v", action="store_true",
+                   help="Stampa diag dict per ogni scena")
+    p.add_argument("--out", default=None,
+                   help="Salva report JSON su file")
+    args = p.parse_args(argv)
+    report = run(args.dataset, scene_filter=args.scene_filter,
+                 verbose=args.verbose)
+    if args.out:
+        with open(args.out, "w") as f:
+            json.dump(report, f, indent=2)
+        print(f"Report salvato: {args.out}")
+    return 0 if report["f1"] > 0.5 else 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())

pm2d/gui.py

@@ -196,8 +196,10 @@ def _warp_template_edges_to_scene(
     edge = cv2.Canny(template_gray, canny_low, canny_high)
     # Matrice affine: scala + rotazione attorno al centro template, poi traslazione
     Ht, Wt = h, w
-    cx_t = (Wt - 1) / 2.0
-    cy_t = (Ht - 1) / 2.0
+    # Centro coerente con la convenzione train (center = w / 2.0, no -1):
+    # (Wt-1)/2 introduceva uno shift di 0.5px per template di lato pari.
+    cx_t = Wt / 2.0
+    cy_t = Ht / 2.0
     M = cv2.getRotationMatrix2D((cx_t, cy_t), angle_deg, scale)
     # Traslazione per portare centro template a (cx, cy) della scena
     M[0, 2] += cx - cx_t
@@ -492,7 +494,9 @@ def run(
     num_features: int = 96,
     weak_grad: float = 30.0,
     strong_grad: float = 60.0,
-    spread_radius: int = 5,
+    # 4 allineato col default del matcher: raggio 5 peggiora la precisione
+    # di rotazione (spread troppo largo appiattisce il picco angolare).
+    spread_radius: int = 4,
     pyramid_levels: int = 3,
     min_score: float = 0.55,
     max_matches: int = 25,

pm2d/matcher.py

@@ -91,8 +91,16 @@ class EdgeShapeMatcher:
         a0, a1 = self.angle_range_deg
         if self.angle_step_deg <= 0 or a0 >= a1:
             return [float(a0)]
-        n = int(np.floor((a1 - a0) / self.angle_step_deg))
-        return [float(a0 + i * self.angle_step_deg) for i in range(n)]
+        # n+1 valori per includere l'estremo superiore del range: con il
+        # solo floor un range [0, 90] step 5 si fermava a 85° (off-by-one).
+        n = int(np.floor((a1 - a0) / self.angle_step_deg)) + 1
+        angles = [float(a0 + i * self.angle_step_deg) for i in range(n)]
+        if a1 - a0 >= 360.0:
+            # Range che copre il giro completo: a0+360° è la stessa pose di
+            # a0, escludi il duplicato (variante inutile in train/find).
+            eps = 1e-6
+            angles = [a for a in angles if a < a0 + 360.0 - eps]
+        return angles
 
     def train(self, template_bgr: np.ndarray) -> int:
         """Genera varianti per tutte le combinazioni (angolo, scala)."""
@@ -222,6 +230,14 @@ class EdgeShapeMatcher:
             for y, x in zip(ys, xs):
                 candidates.append((float(res[y, x]), int(x), int(y), ti))
 
+        # Cap candidati top-level: senza limite np.where con soglia bassa
+        # può generare migliaia di candidati, ognuno con un matchTemplate
+        # full-res nel refinement. Tieni solo i migliori per score.
+        max_candidates = max(1, max_matches * 10)
+        if len(candidates) > max_candidates:
+            candidates.sort(key=lambda c: -c[0])
+            candidates = candidates[:max_candidates]
+
         # Refinement a risoluzione piena: per ogni candidato top, finestra locale
         refined: list[tuple[float, int, int, int]] = []
         margin = sf + 4
@@ -294,6 +310,10 @@ class EdgeShapeMatcher:
         )
         arrays = {f"edge_{i}": t.edge for i, t in enumerate(self.templates)}
         arrays.update({f"mask_{i}": t.mask for i, t in enumerate(self.templates)})
+        # Persisti anche il grayscale originale: senza, l'overlay edge
+        # spariva dopo load() (template_gray restava None).
+        if self.template_gray is not None:
+            arrays["template_gray"] = self.template_gray
         np.savez_compressed(path, params=params, meta=meta, **arrays)
 
     @classmethod
@@ -312,6 +332,10 @@ class EdgeShapeMatcher:
             top_score_factor=float(p[12]) if len(p) > 12 else 0.6,
         )
         m.template_size = (int(p[8]), int(p[9]))
+        # Retrocompatibilità: modelli salvati prima non hanno template_gray
+        # (resta None: overlay edge non disponibile ma find() funziona).
+        if "template_gray" in z.files:
+            m.template_gray = z["template_gray"]
         meta = z["meta"]
         for i in range(len(meta)):
             m.templates.append(

pm2d/web/server.py

@@ -12,6 +12,7 @@ from __future__ import annotations
 import hashlib
 import os
 import tempfile
+import threading
 import time
 import uuid
 from collections import OrderedDict
@@ -48,6 +49,10 @@ IMAGES_DIR = Path(_images_dir_raw)
 if not IMAGES_DIR.is_absolute():
     IMAGES_DIR = PROJECT_ROOT / IMAGES_DIR
 
+# Cartella ricette pre-trained (V feature: save/load matcher)
+RECIPES_DIR = PROJECT_ROOT / "recipes"
+RECIPES_DIR.mkdir(exist_ok=True)
+
 from pm2d.line_matcher import LineShapeMatcher, Match
 from pm2d.auto_tune import auto_tune
 
@@ -60,20 +65,30 @@ STATIC_DIR.mkdir(exist_ok=True)
 CACHE_DIR = Path(tempfile.gettempdir()) / "pm2d_cache"
 CACHE_DIR.mkdir(exist_ok=True)
 
-# Cache in-memory (soft, ricaricata da disco se mancante)
-_IMG_CACHE: dict[str, np.ndarray] = {}
+# Cache in-memory (soft, ricaricata da disco se mancante).
+# LRU con capacità limitata: senza eviction le immagini si accumulavano
+# senza limite (leak di memoria su server long-running).
+_IMG_CACHE: OrderedDict[str, np.ndarray] = OrderedDict()
+_IMG_CACHE_SIZE = 64
 
 # Cache matcher addestrati: (roi_hash, params_hash) -> LineShapeMatcher
 # LRU con capacità limitata
 _MATCHER_CACHE: OrderedDict = OrderedDict()
 _MATCHER_CACHE_SIZE = 8
 
+# Lock globale matcher: gli endpoint girano nel threadpool FastAPI ma i
+# matcher condivisi (_MATCHER_CACHE, _RECIPE_MATCHERS) mutano stato interno
+# durante train()/find(). Serializzare il matching è la soluzione semplice
+# e corretta (un lock per-ricetta sarebbe over-engineering).
+_MATCHER_LOCK = threading.Lock()
+
 
 def _matcher_cache_key(roi: np.ndarray, tech: dict) -> str:
     h = hashlib.md5()
     h.update(roi.tobytes())
     # Solo parametri che influenzano il training
     relevant = ("num_features", "weak_grad", "strong_grad",
+                "min_feature_spacing",
                 "angle_min", "angle_max", "angle_step",
                 "scale_min", "scale_max", "scale_step",
                 "spread_radius", "pyramid_levels")
@@ -97,23 +112,32 @@ def _cache_put_matcher(key: str, matcher) -> None:
         _MATCHER_CACHE.popitem(last=False)
 
 
+def _img_cache_put(key: str, value: np.ndarray) -> None:
+    """Inserisce in _IMG_CACHE con eviction LRU (cap _IMG_CACHE_SIZE)."""
+    _IMG_CACHE[key] = value
+    _IMG_CACHE.move_to_end(key)
+    while len(_IMG_CACHE) > _IMG_CACHE_SIZE:
+        _IMG_CACHE.popitem(last=False)
+
+
 def _store_image(img: np.ndarray) -> str:
     iid = uuid.uuid4().hex[:12]
     cv2.imwrite(str(CACHE_DIR / f"{iid}.png"), img)
-    _IMG_CACHE[iid] = img
+    _img_cache_put(iid, img)
     return iid
 
 
 def _load_image(iid: str) -> np.ndarray | None:
     cached = _IMG_CACHE.get(iid)
     if cached is not None:
+        _IMG_CACHE.move_to_end(iid)  # LRU touch
         return cached
     p = CACHE_DIR / f"{iid}.png"
     if not p.exists():
         return None
     img = cv2.imread(str(p))
     if img is not None:
-        _IMG_CACHE[iid] = img
+        _img_cache_put(iid, img)
     return img
 
 app = FastAPI(title="PM2D Webapp", version="1.0.0")
@@ -126,46 +150,144 @@ def _encode_png(img: np.ndarray) -> bytes:
     return buf.tobytes()
 
 
+def _clamp_roi(x: int, y: int, w: int, h: int,
+               img_w: int, img_h: int) -> tuple[int, int, int, int]:
+    """Clampa la ROI dentro i limiti immagine.
+
+    Una ROI fuori immagine causava slice vuote → crash 500 negli endpoint
+    che non clampavano. Solleva 400 se la ROI risultante è degenere
+    (lato < 16 px: sotto questa soglia il train non estrae abbastanza
+    edge feature e produce 0 varianti → find() esplode con 500).
+    """
+    x = max(0, min(int(x), img_w - 1))
+    y = max(0, min(int(y), img_h - 1))
+    w = min(int(w), img_w - x)
+    h = min(int(h), img_h - y)
+    if w < 16 or h < 16:
+        raise HTTPException(
+            400, f"ROI fuori immagine o degenere: [{x}, {y}, {w}, {h}] "
+                 f"su immagine {img_w}x{img_h} (lato minimo 16 px)")
+    return x, y, w, h
+
+
+def _check_trained(m: "LineShapeMatcher", n_variants: int) -> None:
+    """Solleva 422 se il train non ha prodotto varianti.
+
+    Succede con ROI senza contrasto (sfondo uniforme) o troppo piccola:
+    senza questo check il find() successivo esplode con RuntimeError → 500.
+    """
+    if n_variants <= 0 or not m.variants:
+        raise HTTPException(
+            422, "La ROI non contiene abbastanza edge feature per il "
+                 "training (zona troppo uniforme o piccola): scegliere "
+                 "una regione con contorni netti")
+
+
 def _draw_matches(scene: np.ndarray, matches: list[Match],
-                   template_gray: np.ndarray | None) -> np.ndarray:
+                   template_gray: np.ndarray | None,
+                   matcher: "LineShapeMatcher | None" = None) -> np.ndarray:
+    """Disegna SOLO UCS (richiesta utente) per ogni match trovato.
+
+    UCS = sistema di coordinate (X rosso, Y verde) posizionato sul
+    baricentro feature del modello, ruotato secondo l'angolo del match.
+    Niente edge, niente cerchietti feature, niente bbox: i match sulla
+    scena reale devono essere puliti, gli edge filtrati si vedono solo
+    nell'anteprima modello.
+    """
     out = scene.copy()
-    H, W = scene.shape[:2]
-    palette = [
-        (0, 255, 0), (0, 200, 255), (255, 100, 100), (255, 200, 0),
-        (200, 0, 255), (100, 255, 200), (255, 0, 0), (0, 255, 255),
-    ]
+    # Lunghezza assi UCS: stessa formula dell'anteprima modello
+    # (0.15 * max lato template) scalata per m.scale → coerenza dimensionale.
+    if matcher is not None and matcher.template_size != (0, 0):
+        L_base = int(0.15 * max(matcher.template_size))
+    else:
+        L_base = 30
+    H_scene, W_scene = scene.shape[:2]
+
     for i, m in enumerate(matches):
-        color = palette[i % len(palette)]
-        if template_gray is not None:
+        # UCS posizionato esattamente sul CENTRO POSE del match (m.cx, m.cy):
+        # equivale al centro template traslato alla scena, ruotato con
+        # m.angle_deg. Coerente con UCS dell'anteprima modello che ora
+        # e' anche sul centro ROI (vedi preview_edges).
+        ax = np.deg2rad(m.angle_deg)
+        ca, sa = np.cos(ax), np.sin(ax)
+        cx, cy = int(round(m.cx)), int(round(m.cy))
+        # Overlay edge del modello orientato (richiesta utente):
+        # warpa template alla pose, applica hysteresis identica al matcher,
+        # disegna pixel edge come overlay verde tenue. Maschera col
+        # _train_mask warpato + erode per rimuovere edge sui BORDI del
+        # rettangolo template (transizione bordo nero → scena = falso edge
+        # che appariva come "ROI" attorno a ogni match).
+        if template_gray is not None and matcher is not None:
             t = template_gray
             th, tw = t.shape
-            edge = cv2.Canny(t, 50, 150)
-            cx_t = (tw - 1) / 2.0; cy_t = (th - 1) / 2.0
-            M = cv2.getRotationMatrix2D((cx_t, cy_t), m.angle_deg, m.scale)
-            M[0, 2] += m.cx - cx_t
-            M[1, 2] += m.cy - cy_t
-            warped = cv2.warpAffine(edge, M, (W, H),
-                                     flags=cv2.INTER_NEAREST, borderValue=0)
-            mask = warped > 0
-            if mask.any():
-                overlay = np.zeros_like(out)
-                overlay[mask] = color
-                out[mask] = (0.3 * out[mask] + 0.7 * overlay[mask]).astype(np.uint8)
-        poly = m.bbox_poly.astype(np.int32).reshape(-1, 1, 2)
-        cv2.polylines(out, [poly], True, color, 2, cv2.LINE_AA)
-        p0 = tuple(m.bbox_poly[0].astype(int))
-        p1 = tuple(m.bbox_poly[1].astype(int))
-        cv2.line(out, p0, p1, color, 4, cv2.LINE_AA)
-        cx, cy = int(round(m.cx)), int(round(m.cy))
-        cv2.drawMarker(out, (cx, cy), color, cv2.MARKER_CROSS, 22, 2, cv2.LINE_AA)
-        L = int(np.linalg.norm(m.bbox_poly[1] - m.bbox_poly[0])) // 2
-        a = np.deg2rad(m.angle_deg)
-        cv2.arrowedLine(out, (cx, cy),
-                        (int(cx + L * np.cos(a)), int(cy - L * np.sin(a))),
-                        color, 2, cv2.LINE_AA, tipLength=0.2)
-        label = f"#{i+1} {m.angle_deg:.0f}d s={m.scale:.2f} {m.score:.2f}"
-        cv2.putText(out, label, (cx + 8, cy - 8),
-                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, 2, cv2.LINE_AA)
+            # Centro template coerente col training: in train si usa
+            # `center = (diag / 2.0, diag / 2.0)` (no -1). Usare (tw-1)/2
+            # introduceva uno shift di 0.5px per template di lato pari.
+            cx_t = tw / 2.0; cy_t = th / 2.0
+            # Lavora su un CROP locale della scena di lato = diagonale del
+            # template ruotato+scalato (+margine), come _verify_ncc: warp
+            # + Sobel sull'INTERA scena per ogni match erano O(W·H) cadauno
+            # (costosissimo su scene grandi con molti match).
+            diag = int(np.ceil(np.hypot(tw, th) * m.scale)) + 8
+            x0 = int(round(m.cx)) - diag // 2
+            y0 = int(round(m.cy)) - diag // 2
+            gx0 = max(0, x0); gy0 = max(0, y0)
+            gx1 = min(W_scene, x0 + diag); gy1 = min(H_scene, y0 + diag)
+            cw, ch_ = gx1 - gx0, gy1 - gy0
+            if cw >= 3 and ch_ >= 3:
+                M = cv2.getRotationMatrix2D((cx_t, cy_t), m.angle_deg, m.scale)
+                # Porta il centro template a (m.cx - gx0, m.cy - gy0) del crop
+                M[0, 2] += (m.cx - gx0) - cx_t
+                M[1, 2] += (m.cy - gy0) - cy_t
+                warped_gray = cv2.warpAffine(
+                    t, M, (cw, ch_),
+                    flags=cv2.INTER_LINEAR, borderValue=0)
+                # Maschera: train_mask se disponibile, altrimenti rettangolo pieno
+                mask_src = (matcher._train_mask if matcher._train_mask is not None
+                            else np.full((th, tw), 255, dtype=np.uint8))
+                warped_mask = cv2.warpAffine(
+                    mask_src, M, (cw, ch_),
+                    flags=cv2.INTER_NEAREST, borderValue=0)
+                # Erode minimo (3x3) per togliere SOLO artefatti border-padding
+                # (~1px di bordo nero da warpAffine borderValue=0). Erode piu'
+                # grande spostava visualmente l'edge verso l'interno e creava
+                # apparente "traslazione fissa" rispetto al bordo del pezzo.
+                kernel_er = np.ones((3, 3), np.uint8)
+                warped_mask = cv2.erode(warped_mask, kernel_er)
+                mag, _ = matcher._gradient(warped_gray)
+                if matcher.weak_grad < matcher.strong_grad:
+                    edge_mask = matcher._hysteresis_mask(mag)
+                else:
+                    edge_mask = mag >= matcher.strong_grad
+                edge_mask = edge_mask & (warped_mask > 0)
+                if edge_mask.any():
+                    # Edge ritraslati nel sistema scena: blend solo sul crop
+                    # (addWeighted lascia invariati i pixel con overlay nullo,
+                    # quindi l'output visivo è identico al full-frame).
+                    sub = out[gy0:gy1, gx0:gx1]
+                    edge_overlay = np.zeros_like(sub)
+                    # Ciano (cambiato da verde): non collide col verde dell'asse
+                    # Y dell'UCS che altrimenti scompariva nell'overlay edge.
+                    edge_overlay[edge_mask] = (255, 200, 0)  # ciano (BGR)
+                    out[gy0:gy1, gx0:gx1] = cv2.addWeighted(
+                        sub, 1.0, edge_overlay, 0.6, 0)
+        L = max(20, int(L_base * m.scale))
+        # X axis = rotazione di (1, 0) con cv2 matrix → (cos, -sin)
+        x_end = (int(cx + L * ca), int(cy - L * sa))
+        # Y axis = rotazione di (0, 1) con cv2 matrix → (sin, cos)
+        # A m.angle_deg=0 deve puntare GIU' (image y-down convenzione modello)
+        y_end = (int(cx + L * sa), int(cy + L * ca))
+        cv2.arrowedLine(out, (cx, cy), x_end,
+                        (0, 0, 255), 2, cv2.LINE_AA, tipLength=0.2)
+        cv2.putText(out, "X", (x_end[0] + 4, x_end[1] + 5),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
+        cv2.arrowedLine(out, (cx, cy), y_end,
+                        (0, 255, 0), 2, cv2.LINE_AA, tipLength=0.2)
+        cv2.putText(out, "Y", (y_end[0] + 4, y_end[1] + 12),
+                    cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 1, cv2.LINE_AA)
+        # Origine UCS: cerchio bianco con bordo nero
+        cv2.circle(out, (cx, cy), 4, (0, 0, 0), -1, cv2.LINE_AA)
+        cv2.circle(out, (cx, cy), 3, (255, 255, 255), -1, cv2.LINE_AA)
     return out
 
 
@@ -193,7 +315,9 @@ class MatchParams(BaseModel):
     num_features: int = 96
     weak_grad: float = 30.0
     strong_grad: float = 60.0
-    spread_radius: int = 5
+    # 4 allineato col default del matcher: raggio 5 peggiora la precisione
+    # di rotazione (spread troppo largo appiattisce il picco angolare).
+    spread_radius: int = 4
     pyramid_levels: int = 3
     verify_threshold: float = 0.4
 
@@ -213,6 +337,7 @@ class MatchResp(BaseModel):
     find_time: float
     num_variants: int
     annotated_id: str
+    diag: dict | None = None  # CC: diagnostica pipeline (drop reasons)
 
 
 class TuneParams(BaseModel):
@@ -249,9 +374,9 @@ PRECISION_ANGLE_STEP = {
 # Un operatore sceglie il livello di rigore, non un numero astratto.
 FILTRO_FP_MAP = {
     "off":      0.0,   # disabilitato: mantieni tutti i match shape-based
-    "leggero":  0.20,  # tollera variazioni intensità/illuminazione forti
-    "medio":    0.35,  # default bilanciato (consigliato)
-    "forte":    0.50,  # scarta match con intensità molto diversa dal template
+    "leggero":  0.30,  # tollera variazioni intensità/illuminazione forti
+    "medio":    0.50,  # default bilanciato (consigliato)
+    "forte":    0.70,  # scarta match con intensità molto diversa dal template
 }
 
 
@@ -267,6 +392,29 @@ class SimpleMatchParams(BaseModel):
     penalita_scala: float = 0.0       # 0 = score shape invariante, >0 = penalizza scala != 1
     min_score: float = 0.65
     max_matches: int = 25
+    # --- Override edge da pannello "Anteprima edge" (None = auto_tune) ---
+    # Quando settati, sovrascrivono i valori derivati da auto_tune e
+    # vengono usati identici sia nel training del matcher sia nel find.
+    # Salvati nella ricetta cosi' la stessa pulizia rumore e' replicata
+    # quando la ricetta viene caricata.
+    edge_weak_grad: float | None = None
+    edge_strong_grad: float | None = None
+    edge_num_features: int | None = None
+    edge_min_feature_spacing: int | None = None
+    # --- Halcon-mode flags (default off = backward compat) ---
+    # Init-time (richiede ri-train se cambiato)
+    use_polarity: bool = False        # F: 16 bin orientation mod 2pi
+    use_gpu: bool = False             # R: OpenCL UMat (silent fallback)
+    # Find-time (no retrain)
+    min_recall: float = 0.0           # M: filtra match con poche feature combaciate
+    use_soft_score: bool = False      # Y: cosine sim continua dei gradients
+    subpixel_lm: bool = False         # Z: precisione 0.05 px
+    nms_iou_threshold: float = 0.3    # A: IoU bbox poligonale
+    coarse_stride: int = 1            # sub-sampling top-level (>=1)
+    pyramid_propagate: bool = False   # propagazione candidati top->full
+    greediness: float = 0.0           # early-exit kernel (0..1)
+    refine_pose_joint: bool = False   # Nelder-Mead 3D (cx, cy, angle)
+    search_roi: list[int] | None = None  # [x, y, w, h] limita area
 
 
 def _simple_to_technical(
@@ -301,10 +449,24 @@ def _simple_to_technical(
     smin, smax, sstep = SCALE_PRESETS.get(p.scala, (1.0, 1.0, 0.1))
     ang_step = PRECISION_ANGLE_STEP.get(p.precisione, 5.0)
 
+    # Override edge dal pannello "Anteprima edge" se utente li ha settati.
+    # Questi sostituiscono i valori auto_tune nel training del matcher,
+    # garantendo che la selezione edge identica a quella del preview
+    # venga usata sia in training sia in find.
+    weak_g = (p.edge_weak_grad if p.edge_weak_grad is not None
+              else tune["weak_grad"])
+    strong_g = (p.edge_strong_grad if p.edge_strong_grad is not None
+                else tune["strong_grad"])
+    n_feat = (p.edge_num_features if p.edge_num_features is not None
+              else nf)
+    min_sp = (p.edge_min_feature_spacing if p.edge_min_feature_spacing is not None
+              else 3)
+
     return {
-        "num_features": nf,
-        "weak_grad": tune["weak_grad"],
-        "strong_grad": tune["strong_grad"],
+        "num_features": n_feat,
+        "weak_grad": weak_g,
+        "strong_grad": strong_g,
+        "min_feature_spacing": min_sp,
         "spread_radius": spread,
         "pyramid_levels": pyr,
         "angle_min": 0.0,
@@ -316,7 +478,13 @@ def _simple_to_technical(
         "min_score": p.min_score,
         "max_matches": p.max_matches,
         "nms_radius": 0,
-        "verify_threshold": FILTRO_FP_MAP.get(p.filtro_fp, 0.35),
+        # Fallback = livello "medio" della mappa (no valore hardcoded
+        # che divergerebbe se la mappa cambia).
+        "verify_threshold": FILTRO_FP_MAP.get(p.filtro_fp, FILTRO_FP_MAP["medio"]),
+        # "off" deve disabilitare DAVVERO il verify NCC: passare solo
+        # verify_threshold=0.0 lascerebbe attivo il calcolo NCC (che può
+        # comunque scartare match con score negativo / patch uniformi).
+        "verify_ncc": p.filtro_fp != "off",
         "scale_penalty": p.penalita_scala,
     }
 
@@ -450,9 +618,7 @@ def match(p: MatchParams):
     if model is None or scene is None:
         raise HTTPException(404, "Immagini non trovate")
     x, y, w, h = p.roi
-    x = max(0, x); y = max(0, y)
-    w = max(1, min(w, model.shape[1] - x))
-    h = max(1, min(h, model.shape[0] - y))
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
 
     tech_for_cache = {
@@ -466,33 +632,38 @@ def match(p: MatchParams):
         "pyramid_levels": p.pyramid_levels,
     }
     key = _matcher_cache_key(roi_img, tech_for_cache)
-    m = _cache_get_matcher(key)
-    if m is None:
-        m = LineShapeMatcher(
-            num_features=p.num_features,
-            weak_grad=p.weak_grad, strong_grad=p.strong_grad,
-            angle_range_deg=(p.angle_min, p.angle_max),
-            angle_step_deg=p.angle_step,
-            scale_range=(p.scale_min, p.scale_max),
-            scale_step=p.scale_step,
-            spread_radius=p.spread_radius,
-            pyramid_levels=p.pyramid_levels,
+    # Lock globale: matcher condivisi tra thread del pool FastAPI
+    with _MATCHER_LOCK:
+        m = _cache_get_matcher(key)
+        if m is None:
+            m = LineShapeMatcher(
+                num_features=p.num_features,
+                weak_grad=p.weak_grad, strong_grad=p.strong_grad,
+                angle_range_deg=(p.angle_min, p.angle_max),
+                angle_step_deg=p.angle_step,
+                scale_range=(p.scale_min, p.scale_max),
+                scale_step=p.scale_step,
+                spread_radius=p.spread_radius,
+                pyramid_levels=p.pyramid_levels,
+            )
+            t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
+            _check_trained(m, n)
+            _cache_put_matcher(key, m)
+        else:
+            n = len(m.variants); t_train = 0.0
+        nms = p.nms_radius if p.nms_radius > 0 else None
+        t0 = time.time()
+        matches = m.find(
+            scene, min_score=p.min_score, max_matches=p.max_matches,
+            nms_radius=nms, verify_threshold=p.verify_threshold,
+            # Soglia 0 = filtro FP disattivato: skippa proprio il calcolo NCC
+            verify_ncc=p.verify_threshold > 0.0,
         )
-        t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
-        _cache_put_matcher(key, m)
-    else:
-        n = len(m.variants); t_train = 0.0
-    nms = p.nms_radius if p.nms_radius > 0 else None
-    t0 = time.time()
-    matches = m.find(
-        scene, min_score=p.min_score, max_matches=p.max_matches,
-        nms_radius=nms, verify_threshold=p.verify_threshold,
-    )
-    t_find = time.time() - t0
+        t_find = time.time() - t0
 
-    # Render annotated image
-    tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
-    annotated = _draw_matches(scene, matches, tg)
+        # Render annotated image
+        tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
+        annotated = _draw_matches(scene, matches, tg, matcher=m)
     ann_id = _store_image(annotated)
 
     return MatchResp(
@@ -503,6 +674,7 @@ def match(p: MatchParams):
         ) for m_ in matches],
         train_time=t_train, find_time=t_find,
         num_variants=n, annotated_id=ann_id,
+        diag=m.get_last_diag() if hasattr(m, "get_last_diag") else None,
     )
 
 
@@ -518,41 +690,61 @@ def match_simple(p: SimpleMatchParams):
     if model is None or scene is None:
         raise HTTPException(404, "Immagini non trovate")
     x, y, w, h = p.roi
-    x = max(0, x); y = max(0, y)
-    w = max(1, min(w, model.shape[1] - x))
-    h = max(1, min(h, model.shape[0] - y))
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
 
     tech = _simple_to_technical(p, roi_img)
 
     key = _matcher_cache_key(roi_img, tech)
-    m = _cache_get_matcher(key)
-    if m is None:
-        m = LineShapeMatcher(
-            num_features=tech["num_features"],
-            weak_grad=tech["weak_grad"], strong_grad=tech["strong_grad"],
-            angle_range_deg=(tech["angle_min"], tech["angle_max"]),
-            angle_step_deg=tech["angle_step"],
-            scale_range=(tech["scale_min"], tech["scale_max"]),
-            scale_step=tech["scale_step"],
-            spread_radius=tech["spread_radius"],
-            pyramid_levels=tech["pyramid_levels"],
+    # Halcon-mode init params: incidono sul training, includere in cache key
+    halcon_init_key = f"|pol={p.use_polarity}|gpu={p.use_gpu}"
+    key = key + halcon_init_key
+    # Lock globale: matcher condivisi tra thread del pool FastAPI
+    with _MATCHER_LOCK:
+        m = _cache_get_matcher(key)
+        if m is None:
+            m = LineShapeMatcher(
+                num_features=tech["num_features"],
+                weak_grad=tech["weak_grad"], strong_grad=tech["strong_grad"],
+                angle_range_deg=(tech["angle_min"], tech["angle_max"]),
+                angle_step_deg=tech["angle_step"],
+                scale_range=(tech["scale_min"], tech["scale_max"]),
+                scale_step=tech["scale_step"],
+                spread_radius=tech["spread_radius"],
+                min_feature_spacing=tech.get("min_feature_spacing", 3),
+                pyramid_levels=tech["pyramid_levels"],
+                use_polarity=p.use_polarity,
+                use_gpu=p.use_gpu,
+            )
+            t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
+            _check_trained(m, n)
+            _cache_put_matcher(key, m)
+        else:
+            n = len(m.variants); t_train = 0.0
+        nms = tech["nms_radius"] if tech["nms_radius"] > 0 else None
+        search_roi_t = tuple(p.search_roi) if p.search_roi else None
+        t0 = time.time()
+        matches = m.find(
+            scene, min_score=tech["min_score"], max_matches=tech["max_matches"],
+            nms_radius=nms, verify_threshold=tech["verify_threshold"],
+            # filtro_fp="off" → verify NCC davvero disabilitato
+            verify_ncc=tech.get("verify_ncc", True),
+            scale_penalty=tech.get("scale_penalty", 0.0),
+            # Halcon-mode flags
+            min_recall=p.min_recall,
+            use_soft_score=p.use_soft_score,
+            subpixel_lm=p.subpixel_lm,
+            nms_iou_threshold=p.nms_iou_threshold,
+            coarse_stride=p.coarse_stride,
+            pyramid_propagate=p.pyramid_propagate,
+            greediness=p.greediness,
+            refine_pose_joint=p.refine_pose_joint,
+            search_roi=search_roi_t,
         )
-        t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
-        _cache_put_matcher(key, m)
-    else:
-        n = len(m.variants); t_train = 0.0
-    nms = tech["nms_radius"] if tech["nms_radius"] > 0 else None
-    t0 = time.time()
-    matches = m.find(
-        scene, min_score=tech["min_score"], max_matches=tech["max_matches"],
-        nms_radius=nms, verify_threshold=tech["verify_threshold"],
-        scale_penalty=tech.get("scale_penalty", 0.0),
-    )
-    t_find = time.time() - t0
+        t_find = time.time() - t0
 
-    tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
-    annotated = _draw_matches(scene, matches, tg)
+        tg = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY)
+        annotated = _draw_matches(scene, matches, tg, matcher=m)
     ann_id = _store_image(annotated)
 
     return MatchResp(
@@ -562,6 +754,7 @@ def match_simple(p: SimpleMatchParams):
         ) for mt in matches],
         train_time=t_train, find_time=t_find,
         num_variants=n, annotated_id=ann_id,
+        diag=m.get_last_diag() if hasattr(m, "get_last_diag") else None,
     )
 
 
@@ -571,9 +764,299 @@ def tune(p: TuneParams):
     if model is None:
         raise HTTPException(404, "Immagine non trovata")
     x, y, w, h = p.roi
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
     roi_img = model[y:y + h, x:x + w]
     t = auto_tune(roi_img)
-    return {k: v for k, v in t.items() if not k.startswith("_")}
+    # Esponi parametri tecnici + meta diagnostica (_self_score, _validation,
+    # _symmetry_order, _orient_entropy) per feedback UI.
+    return t
+
+
+# --- V: Save/Load ricette pre-trained ---
+
+class SaveRecipeParams(BaseModel):
+    model_id: str
+    scene_id: str | None = None
+    roi: list[int]
+    # Riusa stessi param simple per training equivalente
+    tipo: str = "intero"
+    simmetria: str = "nessuna"
+    scala: str = "fissa"
+    precisione: str = "normale"
+    use_polarity: bool = False
+    use_gpu: bool = False
+    # Override edge dal pannello "Anteprima edge" (None = auto_tune)
+    edge_weak_grad: float | None = None
+    edge_strong_grad: float | None = None
+    edge_num_features: int | None = None
+    edge_min_feature_spacing: int | None = None
+    name: str  # nome file ricetta (no path)
+
+
+class EdgePreviewParams(BaseModel):
+    model_id: str
+    roi: list[int]
+    weak_grad: float = 30.0
+    strong_grad: float = 60.0
+    num_features: int = 96
+    min_feature_spacing: int = 3
+    use_polarity: bool = False
+
+
+@app.post("/preview_edges")
+def preview_edges(p: EdgePreviewParams):
+    """Estrae edge feature dalla ROI con i parametri dati e ritorna
+    immagine annotata con i pixel selezionati come overlay.
+
+    Permette tuning interattivo delle soglie weak/strong_grad e
+    num_features per "togliere le sporcizie" (rumore di sfondo,
+    edge spuri) prima di trainare il matcher vero.
+    """
+    model = _load_image(p.model_id)
+    if model is None:
+        raise HTTPException(404, "Modello non trovato")
+    x, y, w, h = p.roi
+    H_m, W_m = model.shape[:2]
+    x = max(0, min(int(x), W_m - 1)); y = max(0, min(int(y), H_m - 1))
+    w = max(1, min(int(w), W_m - x)); h = max(1, min(int(h), H_m - y))
+    roi_img = model[y:y + h, x:x + w]
+    # Matcher temporaneo solo per estrazione feature (no train completo)
+    m = LineShapeMatcher(
+        weak_grad=p.weak_grad,
+        strong_grad=p.strong_grad,
+        num_features=p.num_features,
+        min_feature_spacing=p.min_feature_spacing,
+        use_polarity=p.use_polarity,
+    )
+    gray = cv2.cvtColor(roi_img, cv2.COLOR_BGR2GRAY) if roi_img.ndim == 3 else roi_img
+    mag, bins = m._gradient(gray)
+    fx, fy, fb = m._extract_features(mag, bins, None)
+    # Mostra anche i pixel "weak/strong" come heatmap di sfondo
+    out = roi_img.copy() if roi_img.ndim == 3 else cv2.cvtColor(roi_img, cv2.COLOR_GRAY2BGR)
+    # Overlay magnitude leggera
+    mag_norm = np.clip(mag / max(1.0, mag.max()) * 255, 0, 255).astype(np.uint8)
+    mag_color = cv2.applyColorMap(mag_norm, cv2.COLORMAP_BONE)
+    out = cv2.addWeighted(out, 0.6, mag_color, 0.4, 0)
+    # Pixel "strong" con hysteresis: contorno verde scuro tenue
+    if m.weak_grad < m.strong_grad:
+        edge_mask = m._hysteresis_mask(mag).astype(np.uint8) * 255
+    else:
+        edge_mask = (mag >= m.strong_grad).astype(np.uint8) * 255
+    edge_overlay = np.zeros_like(out)
+    edge_overlay[edge_mask > 0] = (0, 80, 0)  # verde scuro
+    out = cv2.addWeighted(out, 1.0, edge_overlay, 0.5, 0)
+    # Feature scelte: cerchietti colorati per bin
+    bin_colors = [
+        (255, 0, 0), (255, 128, 0), (255, 255, 0), (0, 255, 0),
+        (0, 255, 255), (0, 128, 255), (0, 0, 255), (255, 0, 255),
+        (255, 100, 100), (255, 180, 100), (255, 230, 100), (180, 255, 100),
+        (100, 255, 200), (100, 180, 255), (180, 100, 255), (255, 100, 200),
+    ]
+    for i in range(len(fx)):
+        b = int(fb[i])
+        col = bin_colors[b % len(bin_colors)]
+        cv2.circle(out, (int(fx[i]), int(fy[i])), 2, col, -1, cv2.LINE_AA)
+    # UCS sul CENTRO ROI (coerente con _draw_matches che usa centro pose).
+    # In questo modo l'UCS visualizzato nel modello = UCS del match (modulo
+    # rotazione/traslazione data dalla pose del pezzo trovato).
+    rh, rw = roi_img.shape[:2]
+    bx, by = (rw - 1) // 2, (rh - 1) // 2
+    axis_len = max(20, int(0.15 * max(rw, rh)))
+    cv2.arrowedLine(out, (bx, by), (bx + axis_len, by),
+                    (0, 0, 255), 2, cv2.LINE_AA, tipLength=0.2)
+    cv2.putText(out, "X", (bx + axis_len + 4, by + 5),
+                cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1, cv2.LINE_AA)
+    cv2.arrowedLine(out, (bx, by), (bx, by + axis_len),
+                    (0, 255, 0), 2, cv2.LINE_AA, tipLength=0.2)
+    cv2.putText(out, "Y", (bx + 4, by + axis_len + 12),
+                cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 1, cv2.LINE_AA)
+    cv2.circle(out, (bx, by), 4, (0, 0, 0), -1, cv2.LINE_AA)
+    cv2.circle(out, (bx, by), 3, (255, 255, 255), -1, cv2.LINE_AA)
+    bary_cx, bary_cy = float(bx), float(by)
+    img_id = _store_image(out)
+    n_edge_strong = int((mag >= m.strong_grad).sum())
+    n_edge_total = int(edge_mask.sum() / 255)
+    return {
+        "preview_id": img_id,
+        "n_features": len(fx),
+        "n_edge_strong": n_edge_strong,
+        "n_edge_after_hysteresis": n_edge_total,
+        "mag_max": float(mag.max()),
+        "mag_p50": float(np.percentile(mag, 50)),
+        "mag_p85": float(np.percentile(mag, 85)),
+        "ucs_baricentro": (
+            {"cx": round(bary_cx, 2), "cy": round(bary_cy, 2)}
+            if bary_cx is not None else None
+        ),
+    }
+
+
+@app.post("/recipes")
+def save_recipe(p: SaveRecipeParams):
+    """Allena matcher e salva su disco come ricetta riutilizzabile."""
+    model = _load_image(p.model_id)
+    if model is None:
+        raise HTTPException(404, "Modello non trovato")
+    x, y, w, h = p.roi
+    x, y, w, h = _clamp_roi(x, y, w, h, model.shape[1], model.shape[0])
+    roi_img = model[y:y + h, x:x + w]
+    sp = SimpleMatchParams(
+        model_id=p.model_id, scene_id=p.scene_id or p.model_id, roi=p.roi,
+        tipo=p.tipo, simmetria=p.simmetria, scala=p.scala,
+        precisione=p.precisione,
+        use_polarity=p.use_polarity, use_gpu=p.use_gpu,
+        edge_weak_grad=p.edge_weak_grad,
+        edge_strong_grad=p.edge_strong_grad,
+        edge_num_features=p.edge_num_features,
+        edge_min_feature_spacing=p.edge_min_feature_spacing,
+    )
+    tech = _simple_to_technical(sp, roi_img)
+    m = LineShapeMatcher(
+        num_features=tech["num_features"],
+        weak_grad=tech["weak_grad"], strong_grad=tech["strong_grad"],
+        angle_range_deg=(tech["angle_min"], tech["angle_max"]),
+        angle_step_deg=tech["angle_step"],
+        scale_range=(tech["scale_min"], tech["scale_max"]),
+        scale_step=tech["scale_step"],
+        spread_radius=tech["spread_radius"],
+        pyramid_levels=tech["pyramid_levels"],
+        use_polarity=p.use_polarity,
+        use_gpu=p.use_gpu,
+    )
+    # Lock globale: serializza il training pesante col matching in corso
+    with _MATCHER_LOCK:
+        n_var = m.train(roi_img)
+    _check_trained(m, n_var)
+    safe_name = "".join(c for c in p.name if c.isalnum() or c in "._-")
+    if not safe_name:
+        raise HTTPException(400, "Nome ricetta non valido")
+    if not safe_name.endswith(".npz"):
+        safe_name += ".npz"
+    target = RECIPES_DIR / safe_name
+    m.save_model(str(target))
+    return {"name": safe_name, "size": target.stat().st_size,
+            "n_variants": len(m.variants)}
+
+
+@app.get("/recipes")
+def list_recipes():
+    files = []
+    if RECIPES_DIR.is_dir():
+        for f in sorted(RECIPES_DIR.glob("*.npz")):
+            files.append({"name": f.name, "size": f.stat().st_size})
+    return {"files": files, "dir": str(RECIPES_DIR)}
+
+
+# Cache di matcher caricati da .npz (V feature). Key: nome ricetta.
+_RECIPE_MATCHERS: OrderedDict = OrderedDict()
+_RECIPE_MATCHERS_SIZE = 4
+
+
+def _recipe_matchers_put(name: str, matcher: LineShapeMatcher) -> None:
+    """Inserisce in _RECIPE_MATCHERS con eviction LRU (cap _RECIPE_MATCHERS_SIZE)."""
+    _RECIPE_MATCHERS[name] = matcher
+    _RECIPE_MATCHERS.move_to_end(name)
+    while len(_RECIPE_MATCHERS) > _RECIPE_MATCHERS_SIZE:
+        _RECIPE_MATCHERS.popitem(last=False)
+
+
+@app.post("/recipes/{name}/load")
+def load_recipe(name: str):
+    """Carica ricetta .npz e popola cache matcher in memoria.
+
+    Una volta caricata, /match_recipe la usa direttamente senza
+    re-train. Halcon-equivalent read_shape_model + handle.
+    """
+    safe_name = "".join(c for c in name if c.isalnum() or c in "._-")
+    if not safe_name.endswith(".npz"):
+        safe_name += ".npz"
+    path = RECIPES_DIR / safe_name
+    if not path.is_file():
+        raise HTTPException(404, f"Ricetta non trovata: {safe_name}")
+    m = LineShapeMatcher.load_model(str(path))
+    with _MATCHER_LOCK:
+        _recipe_matchers_put(safe_name, m)
+    return {
+        "name": safe_name,
+        "n_variants": len(m.variants),
+        "template_size": list(m.template_size),
+        "use_polarity": m.use_polarity,
+    }
+
+
+class RecipeMatchParams(BaseModel):
+    recipe: str
+    scene_id: str
+    # Solo find-time params (training gia' fatto offline)
+    min_score: float = 0.65
+    max_matches: int = 25
+    min_recall: float = 0.0
+    use_soft_score: bool = False
+    subpixel_lm: bool = False
+    nms_iou_threshold: float = 0.3
+    coarse_stride: int = 1
+    pyramid_propagate: bool = False
+    greediness: float = 0.0
+    refine_pose_joint: bool = False
+    search_roi: list[int] | None = None
+    # Allineato a MatchParams.verify_threshold (0.4): valori divergenti
+    # davano risultati diversi tra /match e /match_recipe a parità di scena.
+    verify_threshold: float = 0.4
+    scale_penalty: float = 0.0
+
+
+@app.post("/match_recipe", response_model=MatchResp)
+def match_recipe(p: RecipeMatchParams):
+    """Match con ricetta pre-trained: zero training, solo find."""
+    safe_name = p.recipe if p.recipe.endswith(".npz") else f"{p.recipe}.npz"
+    scene = _load_image(p.scene_id)
+    if scene is None:
+        raise HTTPException(404, "Scena non trovata")
+    search_roi_t = tuple(p.search_roi) if p.search_roi else None
+    # Lock globale: matcher condivisi tra thread del pool FastAPI
+    with _MATCHER_LOCK:
+        m = _RECIPE_MATCHERS.get(safe_name)
+        if m is not None:
+            _RECIPE_MATCHERS.move_to_end(safe_name)  # LRU touch
+        else:
+            # Auto-load on demand: stessa eviction LRU di load_recipe
+            # (senza cap la cache cresceva senza limite)
+            path = RECIPES_DIR / safe_name
+            if not path.is_file():
+                raise HTTPException(404, f"Ricetta non trovata: {safe_name}")
+            m = LineShapeMatcher.load_model(str(path))
+            _recipe_matchers_put(safe_name, m)
+        t0 = time.time()
+        matches = m.find(
+            scene,
+            min_score=p.min_score, max_matches=p.max_matches,
+            verify_threshold=p.verify_threshold,
+            # Soglia 0 = filtro FP disattivato: skippa proprio il calcolo NCC
+            verify_ncc=p.verify_threshold > 0.0,
+            scale_penalty=p.scale_penalty,
+            min_recall=p.min_recall,
+            use_soft_score=p.use_soft_score,
+            subpixel_lm=p.subpixel_lm,
+            nms_iou_threshold=p.nms_iou_threshold,
+            coarse_stride=p.coarse_stride,
+            pyramid_propagate=p.pyramid_propagate,
+            greediness=p.greediness,
+            refine_pose_joint=p.refine_pose_joint,
+            search_roi=search_roi_t,
+        )
+        t_find = time.time() - t0
+        tg = m.template_gray if m.template_gray is not None else np.zeros((1, 1), np.uint8)
+        annotated = _draw_matches(scene, matches, tg, matcher=m)
+    ann_id = _store_image(annotated)
+    return MatchResp(
+        matches=[MatchResult(
+            cx=mt.cx, cy=mt.cy, angle_deg=mt.angle_deg, scale=mt.scale,
+            score=mt.score, bbox_poly=mt.bbox_poly.tolist(),
+        ) for mt in matches],
+        train_time=0.0, find_time=t_find,
+        num_variants=len(m.variants), annotated_id=ann_id,
+        diag=m.get_last_diag() if hasattr(m, "get_last_diag") else None,
+    )
 
 
 # Mount static

pm2d/web/static/app.js

@@ -19,6 +19,7 @@ const PALETTE = [
 const state = {
   model: null, scene: null, roi: null, drag: null,
   matches: [], annotatedImg: null,
+  active_recipe: null,  // V: ricetta caricata (string nome) o null
 };
 
 // ---------- Forms ----------
@@ -52,6 +53,63 @@ function readUserParams() {
       document.getElementById("p-penalita-scala").value),
     min_score: parseFloat(document.getElementById("p-min-score").value),
     max_matches: parseInt(document.getElementById("p-max-matches").value, 10),
+    ...readEdgeOverrides(),
+    ...readHalconFlags(),
+  };
+}
+
+function readEdgeOverrides() {
+  // Override edge dal pannello "Anteprima edge". Settati = utente li ha
+  // toccati (anche se uguali al default attuale). Vengono propagati a
+  // _simple_to_technical e usati identici sia in training sia in find.
+  // Inoltre salvati nella ricetta cosi' si replicano al load.
+  const _v = (id, parser) => {
+    const el = document.getElementById(id);
+    if (!el) return null;
+    const v = parser(el.value);
+    return Number.isFinite(v) ? v : null;
+  };
+  // Sempre passa i valori correnti degli slider: e' la richiesta utente
+  // che i param di pulizia rumore vengano usati anche nel find/ricetta.
+  const polCb = document.getElementById("hc-use-polarity");
+  return {
+    edge_weak_grad: _v("ep-weak", parseFloat),
+    edge_strong_grad: _v("ep-strong", parseFloat),
+    edge_num_features: _v("ep-nf", parseInt),
+    edge_min_feature_spacing: _v("ep-sp", parseInt),
+    use_polarity: polCb?.checked || document.getElementById("ep-pol")?.checked,
+  };
+}
+
+function readHalconFlags() {
+  // Halcon-mode toggle: tutti i flag default-off, esposti via "Modalità Halcon"
+  const $cb = (id) => document.getElementById(id)?.checked ?? false;
+  const $num = (id, def) => {
+    const v = parseFloat(document.getElementById(id)?.value);
+    return Number.isFinite(v) ? v : def;
+  };
+  const $int = (id, def) => {
+    const v = parseInt(document.getElementById(id)?.value, 10);
+    return Number.isFinite(v) ? v : def;
+  };
+  const roiStr = document.getElementById("hc-search-roi")?.value.trim() ?? "";
+  let search_roi = null;
+  if (roiStr) {
+    const p = roiStr.split(/[ ,;]+/).map((x) => parseInt(x, 10));
+    if (p.length === 4 && p.every((v) => Number.isFinite(v))) search_roi = p;
+  }
+  return {
+    use_polarity: $cb("hc-use-polarity"),
+    use_gpu: $cb("hc-use-gpu"),
+    use_soft_score: $cb("hc-soft-score"),
+    subpixel_lm: $cb("hc-subpixel-lm"),
+    refine_pose_joint: $cb("hc-refine-joint"),
+    pyramid_propagate: $cb("hc-pyr-propagate"),
+    min_recall: $num("hc-min-recall", 0),
+    nms_iou_threshold: $num("hc-nms-iou", 0.3),
+    greediness: $num("hc-greediness", 0),
+    coarse_stride: $int("hc-coarse-stride", 1),
+    search_roi: search_roi,
   };
 }
 
@@ -274,7 +332,43 @@ function setupROI() {
 }
 
 // ---------- Match action ----------
+async function doMatchRecipe() {
+  if (!state.scene) { setStatus("Carica scena"); return; }
+  setStatus(`Match ricetta ${state.active_recipe}...`);
+  const hc = readHalconFlags();
+  const body = {
+    recipe: state.active_recipe,
+    scene_id: state.scene.id,
+    min_score: parseFloat(document.getElementById("p-min-score").value),
+    max_matches: parseInt(document.getElementById("p-max-matches").value, 10),
+    verify_threshold: 0.50,
+    ...hc,
+  };
+  const r = await fetch("/match_recipe", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify(body),
+  });
+  if (!r.ok) { setStatus(`Errore: ${await r.text()}`); return; }
+  const data = await r.json();
+  state.matches = data.matches;
+  state.annotatedImg = await loadImage(
+    `/image/${data.annotated_id}/raw?t=${Date.now()}`);
+  renderScene();
+  renderLegend();
+  document.getElementById("t-train").textContent = "—";
+  document.getElementById("t-find").textContent = `${data.find_time.toFixed(2)}s`;
+  document.getElementById("t-var").textContent = data.num_variants;
+  document.getElementById("t-match").textContent = data.matches.length;
+  renderDiag(data.diag, data.matches.length);
+  setStatus(`${data.matches.length} match trovati (ricetta ${state.active_recipe})`);
+}
+
 async function doMatch() {
+  // Path V: ricetta caricata → bypass training, solo find su scena
+  if (state.active_recipe) {
+    return doMatchRecipe();
+  }
   if (!state.model) { setStatus("Carica modello"); return; }
   if (!state.scene) { setStatus("Carica scena"); return; }
   if (!state.roi) { setStatus("Seleziona ROI sul modello"); return; }
@@ -294,12 +388,17 @@ async function doMatch() {
     const SCALE_MAP = {fissa:[1,1,0.1], mini:[0.9,1.1,0.05],
                        medio:[0.75,1.25,0.05], max:[0.5,1.5,0.05]};
     const PREC_MAP = {veloce:10, normale:5, preciso:2};
-    const FP_MAP = {off:0, leggero:0.20, medio:0.35, forte:0.50};
+    // Allineato a FILTRO_FP_MAP server-side (server.py)
+    const FP_MAP = {off:0, leggero:0.30, medio:0.50, forte:0.70};
     const [smin, smax, sstep] = SCALE_MAP[user.scala];
+    // NB: SYM_MAP[invariante]=0 e' valido (zero rotazioni). Uso ?? per
+    // distinguere "chiave mancante" da "valore zero": altrimenti 0 || 360
+    // collassa invariante a 360 = bug "simmetria non ha effetto".
+    const angMax = SYM_MAP[user.simmetria] ?? 360;
     body = {
       model_id: state.model.id, scene_id: state.scene.id, roi: state.roi,
-      angle_min: 0, angle_max: SYM_MAP[user.simmetria] || 360,
-      angle_step: PREC_MAP[user.precisione] || 5,
+      angle_min: 0, angle_max: angMax,
+      angle_step: PREC_MAP[user.precisione] ?? 5,
       scale_min: smin, scale_max: smax, scale_step: sstep,
       min_score: user.min_score, max_matches: user.max_matches,
       num_features: adv.num_features ?? 96,
@@ -307,7 +406,7 @@ async function doMatch() {
       strong_grad: adv.strong_grad ?? 60,
       spread_radius: adv.spread_radius ?? 5,
       pyramid_levels: adv.pyramid_levels ?? 3,
-      verify_threshold: adv.verify_threshold ?? (FP_MAP[user.filtro_fp] ?? 0.35),
+      verify_threshold: adv.verify_threshold ?? (FP_MAP[user.filtro_fp] ?? 0.50),
       nms_radius: adv.nms_radius ?? 0,
     };
   } else {
@@ -335,6 +434,7 @@ async function doMatch() {
   document.getElementById("t-find").textContent = `${data.find_time.toFixed(2)}s`;
   document.getElementById("t-var").textContent = data.num_variants;
   document.getElementById("t-match").textContent = data.matches.length;
+  renderDiag(data.diag, data.matches.length);
   setStatus(`${data.matches.length} match trovati${hasAdv ? " (avanzato)" : ""}`);
 }
 
@@ -362,6 +462,305 @@ function setStatus(s) {
 }
 
 // ---------- Init ----------
+// ---------- Edge preview (clean rumore) ----------
+let _epDebounce = null;
+let _epLastImg = null;
+
+async function fetchEdgePreview() {
+  if (!state.model || !state.roi) {
+    document.getElementById("edge-preview-info").textContent =
+      "Disegna prima la ROI sul modello";
+    return;
+  }
+  const body = {
+    model_id: state.model.id,
+    roi: state.roi,
+    weak_grad: parseFloat(document.getElementById("ep-weak").value),
+    strong_grad: parseFloat(document.getElementById("ep-strong").value),
+    num_features: parseInt(document.getElementById("ep-nf").value, 10),
+    min_feature_spacing: parseInt(document.getElementById("ep-sp").value, 10),
+    use_polarity: document.getElementById("ep-pol").checked,
+  };
+  try {
+    const r = await fetch("/preview_edges", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    });
+    if (!r.ok) throw new Error(await r.text());
+    const j = await r.json();
+    _epLastImg = await loadImage(`/image/${j.preview_id}/raw?t=${Date.now()}`);
+    drawEdgePreview();
+    const ucs = j.ucs_baricentro
+      ? ` | UCS=(${j.ucs_baricentro.cx},${j.ucs_baricentro.cy})`
+      : "";
+    document.getElementById("edge-preview-info").innerHTML =
+      `<b>${j.n_features}</b> feature scelte (di ${j.n_edge_after_hysteresis} edge totali)<br>` +
+      `mag: max=${j.mag_max.toFixed(0)} p50=${j.mag_p50.toFixed(0)} ` +
+      `p85=${j.mag_p85.toFixed(0)}${ucs}`;
+  } catch (e) {
+    document.getElementById("edge-preview-info").textContent =
+      `Errore preview: ${e.message}`;
+  }
+}
+
+function drawEdgePreview() {
+  const cnv = document.getElementById("c-edge-preview");
+  if (!_epLastImg) return;
+  const ctx = cnv.getContext("2d");
+  // Fit-contain
+  const r = Math.min(cnv.width / _epLastImg.width,
+                     cnv.height / _epLastImg.height);
+  const w = _epLastImg.width * r;
+  const h = _epLastImg.height * r;
+  const ox = (cnv.width - w) / 2;
+  const oy = (cnv.height - h) / 2;
+  ctx.fillStyle = "#000"; ctx.fillRect(0, 0, cnv.width, cnv.height);
+  ctx.imageSmoothingEnabled = false;
+  ctx.drawImage(_epLastImg, ox, oy, w, h);
+}
+
+function scheduleEdgePreview() {
+  if (_epDebounce) clearTimeout(_epDebounce);
+  _epDebounce = setTimeout(fetchEdgePreview, 200);
+}
+
+function bindEdgePreviewControls() {
+  const slid = (id, valEl) => {
+    const el = document.getElementById(id);
+    const v = document.getElementById(valEl);
+    el.addEventListener("input", () => {
+      v.textContent = el.value;
+      scheduleEdgePreview();
+    });
+  };
+  slid("ep-weak", "ep-weak-v");
+  slid("ep-strong", "ep-strong-v");
+  slid("ep-nf", "ep-nf-v");
+  slid("ep-sp", "ep-sp-v");
+  document.getElementById("ep-pol").addEventListener("change",
+    scheduleEdgePreview);
+  // Auto-refresh quando il pannello viene aperto
+  document.getElementById("edge-preview-panel").addEventListener("toggle",
+    (e) => { if (e.target.open) fetchEdgePreview(); });
+  document.getElementById("btn-edge-apply").addEventListener("click", () => {
+    // Copia i valori correnti nei campi avanzati
+    const map = {
+      "ep-weak": "adv-weak_grad",
+      "ep-strong": "adv-strong_grad",
+      "ep-nf": "adv-num_features",
+      "ep-sp": "adv-min_feature_spacing",
+    };
+    for (const [src, dst] of Object.entries(map)) {
+      const dstEl = document.getElementById(dst);
+      if (dstEl) dstEl.value = document.getElementById(src).value;
+    }
+    // use_polarity: alla checkbox della modalita Halcon
+    const polCb = document.getElementById("hc-use-polarity");
+    if (polCb) polCb.checked = document.getElementById("ep-pol").checked;
+    // Apri pannello Avanzate per feedback
+    const advDetails = document.querySelectorAll("#col-params details");
+    advDetails.forEach((d) => { d.open = true; });
+    alert("Parametri edge applicati. Esegui MATCH per usare i valori scelti.");
+  });
+}
+
+// ---------- CC: Diagnostica match ----------
+function renderDiag(diag, n_matches) {
+  const el = document.getElementById("diag-content");
+  if (!diag) {
+    el.innerHTML = '<em style="color:#888">Diagnostica non disponibile</em>';
+    return;
+  }
+  const dropTotal = (diag.drop_ncc_low || 0) + (diag.drop_min_score_post_avg || 0)
+    + (diag.drop_recall_low || 0) + (diag.drop_bbox_out_of_scene || 0)
+    + (diag.drop_nms_iou || 0);
+  // Hint contestuali se 0 match
+  let hint = "";
+  if (n_matches === 0) {
+    if (diag.n_after_pre_nms === 0) {
+      hint = `<div style="color:#f88; margin-top:6px">⚠ Nessun candidato sopra soglia.
+        Prova: ↓ <b>min_score</b> o ↓ <b>top_thresh</b> (currently ${diag.top_thresh_used.toFixed(2)})</div>`;
+    } else if (diag.drop_ncc_low > 0 && dropTotal === diag.drop_ncc_low) {
+      hint = `<div style="color:#f88; margin-top:6px">⚠ ${diag.drop_ncc_low} candidati droppati da NCC.
+        Prova: ↓ <b>verify_threshold</b> (filtro_fp più leggero)</div>`;
+    } else if (diag.drop_min_score_post_avg > 0) {
+      hint = `<div style="color:#f88; margin-top:6px">⚠ ${diag.drop_min_score_post_avg} match sotto min_score post-NCC.
+        Prova: ↓ <b>min_score</b></div>`;
+    } else if (diag.drop_recall_low > 0) {
+      hint = `<div style="color:#f88; margin-top:6px">⚠ ${diag.drop_recall_low} match con recall < ${diag.min_recall_used}.
+        Prova: ↓ <b>min_recall</b></div>`;
+    } else if (diag.drop_bbox_out_of_scene > 0) {
+      hint = `<div style="color:#f88; margin-top:6px">⚠ ${diag.drop_bbox_out_of_scene} match con bbox fuori scena.
+        Centro derivato male: aumenta <b>min_score</b> o restringi <b>search_roi</b></div>`;
+    }
+  }
+  const flags = [];
+  if (diag.use_polarity) flags.push("polarity");
+  if (diag.use_soft_score) flags.push("soft");
+  if (diag.subpixel_lm) flags.push("subpix-LM");
+  el.innerHTML = `
+    <div><b>Pipeline pruning:</b></div>
+    <div>varianti: ${diag.n_variants_total} → top_eval=${diag.n_variants_top_evaluated}
+       → top_pass=${diag.n_variants_top_passed} → full_eval=${diag.n_variants_full_evaluated}</div>
+    <div><b>Candidati:</b> raw=${diag.n_raw_candidates}
+       → pre_nms=${diag.n_after_pre_nms} → final=${diag.n_final}</div>
+    <div><b>Drop reasons:</b> NCC=${diag.drop_ncc_low}, score=${diag.drop_min_score_post_avg},
+       recall=${diag.drop_recall_low}, bbox=${diag.drop_bbox_out_of_scene}, NMS=${diag.drop_nms_iou}</div>
+    <div><b>Soglie:</b> top=${diag.top_thresh_used.toFixed(2)},
+       min_score=${diag.min_score_used.toFixed(2)},
+       NCC=${diag.verify_threshold_used.toFixed(2)},
+       recall=${diag.min_recall_used.toFixed(2)}</div>
+    ${flags.length ? `<div><b>Flag attivi:</b> ${flags.join(", ")}</div>` : ""}
+    ${hint}
+  `;
+  // Auto-apri pannello se 0 match (segnala problema)
+  if (n_matches === 0) {
+    document.getElementById("diag-panel").open = true;
+  }
+}
+
+// ---------- Auto-tune (Halcon-style) ----------
+async function doAutoTune() {
+  if (!state.model || !state.roi) {
+    alert("Seleziona modello e disegna ROI prima di Auto-tune.");
+    return;
+  }
+  const status = document.getElementById("status");
+  status.textContent = "Analisi ROI in corso...";
+  try {
+    const r = await fetch("/auto_tune", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({
+        model_id: state.model.id,
+        roi: state.roi,
+      }),
+    });
+    if (!r.ok) throw new Error(await r.text());
+    const t = await r.json();
+    // Applica ai campi avanzati (override automatico)
+    for (const [key] of ADV_PARAMS) {
+      const el = document.getElementById(`adv-${key}`);
+      if (el && t[key] !== undefined) el.value = String(t[key]);
+    }
+    // Espandi la sezione Avanzate per mostrare i valori applicati
+    const advDetails = document.querySelector("#col-params details:last-of-type");
+    if (advDetails) advDetails.open = true;
+    // Feedback diagnostico
+    const lines = [
+      `weak/strong: ${t.weak_grad} / ${t.strong_grad}`,
+      `feature: ${t.num_features}, piramide: ${t.pyramid_levels}`,
+      `angle: [${t.angle_min}..${t.angle_max}]@${t.angle_step}°`,
+    ];
+    if (t._symmetry_order > 1) {
+      lines.push(`simmetria rotaz. ${t._symmetry_order}x (conf ${t._symmetry_conf})`);
+    }
+    if (t._self_score !== undefined) {
+      lines.push(`self-validation: ${t._validation}`);
+    }
+    status.textContent = `Auto-tune OK — ${lines[0]}`;
+    alert("Auto-tune completato:\n\n" + lines.join("\n"));
+  } catch (e) {
+    status.textContent = `Auto-tune errore: ${e.message}`;
+    alert(`Errore auto-tune: ${e.message}`);
+  }
+}
+
+// ---------- V: Recipe load/list/unload ----------
+async function refreshRecipeList() {
+  try {
+    const r = await fetch("/recipes");
+    if (!r.ok) return;
+    const j = await r.json();
+    const sel = document.getElementById("hc-recipe-list");
+    const cur = sel.value;
+    sel.innerHTML = '<option value="">— ricette disponibili —</option>';
+    for (const f of j.files) {
+      const o = document.createElement("option");
+      o.value = f.name;
+      o.textContent = `${f.name}  (${(f.size / 1024).toFixed(1)} KB)`;
+      sel.appendChild(o);
+    }
+    if (cur) sel.value = cur;
+  } catch (e) { /* silent */ }
+}
+
+async function loadRecipe() {
+  const sel = document.getElementById("hc-recipe-list");
+  const name = sel.value;
+  if (!name) {
+    alert("Seleziona una ricetta dalla lista.");
+    return;
+  }
+  try {
+    const r = await fetch(`/recipes/${encodeURIComponent(name)}/load`, {
+      method: "POST",
+    });
+    if (!r.ok) throw new Error(await r.text());
+    const j = await r.json();
+    state.active_recipe = j.name;
+    document.getElementById("recipe-status").textContent =
+      `Caricata: ${j.name} — ${j.n_variants} varianti, ` +
+      `${j.template_size[0]}x${j.template_size[1]} px` +
+      (j.use_polarity ? " (polarity)" : "");
+    document.getElementById("recipe-status").style.color = "#0c0";
+    document.getElementById("btn-unload-recipe").disabled = false;
+  } catch (e) {
+    alert(`Errore caricamento: ${e.message}`);
+  }
+}
+
+function unloadRecipe() {
+  state.active_recipe = null;
+  document.getElementById("recipe-status").textContent = "Nessuna ricetta caricata";
+  document.getElementById("recipe-status").style.color = "#888";
+  document.getElementById("btn-unload-recipe").disabled = true;
+}
+
+// ---------- V: Save recipe ----------
+async function saveRecipe() {
+  if (!state.model || !state.roi) {
+    alert("Seleziona modello e disegna ROI prima di salvare la ricetta.");
+    return;
+  }
+  const name = document.getElementById("hc-recipe-name").value.trim();
+  if (!name) {
+    alert("Inserisci un nome per la ricetta.");
+    return;
+  }
+  const user = readUserParams();
+  const body = {
+    model_id: state.model.id,
+    scene_id: state.scene?.id || state.model.id,
+    roi: state.roi,
+    tipo: user.tipo,
+    simmetria: user.simmetria,
+    scala: user.scala,
+    precisione: user.precisione,
+    use_polarity: user.use_polarity,
+    use_gpu: user.use_gpu,
+    edge_weak_grad: user.edge_weak_grad,
+    edge_strong_grad: user.edge_strong_grad,
+    edge_num_features: user.edge_num_features,
+    edge_min_feature_spacing: user.edge_min_feature_spacing,
+    name: name,
+  };
+  try {
+    const r = await fetch("/recipes", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    });
+    if (!r.ok) throw new Error(await r.text());
+    const j = await r.json();
+    alert(`Ricetta salvata: ${j.name}\n${j.n_variants} varianti, ${j.size} bytes`);
+    refreshRecipeList();
+  } catch (e) {
+    alert(`Errore salvataggio: ${e.message}`);
+  }
+}
+
 window.addEventListener("DOMContentLoaded", async () => {
   buildAdvancedForm();
   setupROI();
@@ -389,6 +788,15 @@ window.addEventListener("DOMContentLoaded", async () => {
     e.target.value = "";  // consente re-upload stesso file
   });
   document.getElementById("btn-match").addEventListener("click", doMatch);
+  document.getElementById("btn-autotune").addEventListener("click", doAutoTune);
+  document.getElementById("btn-save-recipe").addEventListener("click",
+    saveRecipe);
+  document.getElementById("btn-load-recipe").addEventListener("click",
+    loadRecipe);
+  document.getElementById("btn-unload-recipe").addEventListener("click",
+    unloadRecipe);
+  refreshRecipeList();
+  bindEdgePreviewControls();
   const slider = document.getElementById("p-min-score");
   slider.addEventListener("input", (e) => {
     document.getElementById("v-score").textContent =

pm2d/web/static/index.html

@@ -26,6 +26,10 @@
       <div class="picker-list"></div>
     </div>
     <button class="btn btn-go" id="btn-match">▶ MATCH</button>
+    <button class="btn" id="btn-autotune"
+            title="Analizza ROI e derivata parametri ottimali (Halcon-style)">
+      ⚙ Auto-tune
+    </button>
     <label class="btn" title="Carica nuovo file nella cartella immagini">
       ⬆ Carica file
       <input type="file" id="file-upload" accept="image/*" hidden>
@@ -41,6 +45,40 @@
       <canvas id="c-model" width="380" height="420"></canvas>
     </div>
     <div id="roi-info">ROI: (nessuna)</div>
+    <details id="edge-preview-panel" style="margin-top:10px">
+      <summary>🔬 Anteprima edge / pulizia rumore</summary>
+      <div style="font-size:11px; color:#aaa; margin:4px 0">
+        Regola le soglie per togliere edge spuri (sporcizie). UCS rosso/verde
+        sul baricentro feature.
+      </div>
+      <div class="ep-grid">
+        <label class="ep-row">weak_grad <span id="ep-weak-v">30</span>
+          <input type="range" id="ep-weak" min="5" max="200" value="30" step="1">
+        </label>
+        <label class="ep-row">strong_grad <span id="ep-strong-v">60</span>
+          <input type="range" id="ep-strong" min="10" max="400" value="60" step="1">
+        </label>
+        <label class="ep-row">num_features <span id="ep-nf-v">96</span>
+          <input type="range" id="ep-nf" min="16" max="300" value="96" step="1">
+        </label>
+        <label class="ep-row">spacing <span id="ep-sp-v">3</span>
+          <input type="range" id="ep-sp" min="1" max="15" value="3" step="1">
+        </label>
+        <label class="ep-row" style="flex-direction:row; gap:6px">
+          <input type="checkbox" id="ep-pol"> polarity
+        </label>
+        <button class="btn" id="btn-edge-apply" type="button"
+                style="grid-column:1/-1">
+          ✓ Applica ai parametri Avanzate
+        </button>
+      </div>
+      <div class="canvas-wrap" style="margin-top:6px">
+        <canvas id="c-edge-preview" width="380" height="380"></canvas>
+      </div>
+      <div id="edge-preview-info" style="font-size:11px; color:#888; margin-top:4px">
+        Disegna ROI e apri questo pannello per generare anteprima
+      </div>
+    </details>
   </section>
 
   <section class="col" id="col-scene">
@@ -64,8 +102,8 @@
     <div class="field">
       <label>Simmetria</label>
       <select id="p-simmetria">
+        <option value="nessuna" selected>Nessuna (0..360°)</option>
         <option value="invariante">Invariante (cerchi — no rotazione)</option>
-        <option value="nessuna">Nessuna (0..360°)</option>
         <option value="bilaterale">Bilaterale (speculare 180°)</option>
         <option value="rot_3">Rotazionale 3× (120°)</option>
         <option value="rot_4">Rotazionale 4× (90°)</option>
@@ -129,6 +167,77 @@
       <input type="number" id="p-max-matches" value="25" min="1" max="200">
     </div>
 
+    <details>
+      <summary>Modalità Halcon</summary>
+      <div class="halcon-grid">
+        <label class="hc-row" title="16-bin orientation polarity-aware (mod 2π)">
+          <input type="checkbox" id="hc-use-polarity">
+          <span>Polarity 16-bin (F)</span>
+        </label>
+        <label class="hc-row" title="Score continuo cos(θ_t-θ_s) invece di bin">
+          <input type="checkbox" id="hc-soft-score">
+          <span>Soft-margin score (Y)</span>
+        </label>
+        <label class="hc-row" title="Sub-pixel refinement gradient field LM">
+          <input type="checkbox" id="hc-subpixel-lm">
+          <span>Sub-pixel LM 0.05 px (Z)</span>
+        </label>
+        <label class="hc-row" title="Refine congiunto Nelder-Mead (cx,cy,θ)">
+          <input type="checkbox" id="hc-refine-joint">
+          <span>Refine pose joint</span>
+        </label>
+        <label class="hc-row" title="Pyramid candidates propagation">
+          <input type="checkbox" id="hc-pyr-propagate">
+          <span>Pyramid propagate</span>
+        </label>
+        <label class="hc-row" title="OpenCL GPU offload (silent fallback CPU)">
+          <input type="checkbox" id="hc-use-gpu">
+          <span>GPU OpenCL (R)</span>
+        </label>
+
+        <div class="hc-row hc-num">
+          <label>Min recall (M)</label>
+          <input type="number" id="hc-min-recall" value="0.0" min="0" max="1" step="0.05">
+        </div>
+        <div class="hc-row hc-num">
+          <label>NMS IoU thr (A)</label>
+          <input type="number" id="hc-nms-iou" value="0.3" min="0" max="1" step="0.05">
+        </div>
+        <div class="hc-row hc-num">
+          <label>Greediness</label>
+          <input type="number" id="hc-greediness" value="0.0" min="0" max="1" step="0.1">
+        </div>
+        <div class="hc-row hc-num">
+          <label>Coarse stride</label>
+          <input type="number" id="hc-coarse-stride" value="1" min="1" max="4" step="1">
+        </div>
+        <div class="hc-row hc-num" style="grid-column:1/-1">
+          <label title="Limita area di ricerca scena: x,y,w,h (vuoto = tutta scena)">
+            Search ROI (x,y,w,h)
+          </label>
+          <input type="text" id="hc-search-roi" placeholder="es. 100,50,800,400">
+        </div>
+
+        <div class="hc-row" style="grid-column:1/-1; border-top:1px solid #444; padding-top:8px">
+          <label>Ricetta pre-trained (V)</label>
+          <div style="display:flex; gap:6px; margin-top:4px">
+            <input type="text" id="hc-recipe-name" placeholder="nome_ricetta" style="flex:1">
+            <button class="btn" id="btn-save-recipe" type="button">💾 Salva</button>
+          </div>
+          <div style="display:flex; gap:6px; margin-top:6px; align-items:center">
+            <select id="hc-recipe-list" style="flex:1">
+              <option value="">— ricette disponibili —</option>
+            </select>
+            <button class="btn" id="btn-load-recipe" type="button">📂 Carica</button>
+            <button class="btn" id="btn-unload-recipe" type="button" disabled>✖ Stacca</button>
+          </div>
+          <div id="recipe-status" style="margin-top:4px; font-size:11px; color:#888">
+            Nessuna ricetta caricata
+          </div>
+        </div>
+      </div>
+    </details>
+
     <details>
       <summary>Avanzate</summary>
       <div id="adv-form"></div>
@@ -139,6 +248,16 @@
     <div class="kv"><span>find:</span><span id="t-find">-</span></div>
     <div class="kv"><span>varianti:</span><span id="t-var">-</span></div>
     <div class="kv"><span>match:</span><span id="t-match">-</span></div>
+
+    <details id="diag-panel" style="margin-top:10px">
+      <summary>🔍 Diagnostica (CC)</summary>
+      <div id="diag-content" style="font-family:monospace; font-size:11px;
+                                     background:#1a1a1a; padding:8px;
+                                     border-radius:3px; margin-top:6px;
+                                     line-height:1.5">
+        <em style="color:#888">Esegui un MATCH per vedere la diagnostica</em>
+      </div>
+    </details>
   </section>
 </main>
 

pm2d/web/static/style.css

@@ -156,3 +156,35 @@ footer h2 {
 }
 
 #col-model, #col-scene { min-width: 0; }
+
+/* Halcon-mode panel */
+.halcon-grid {
+  display: grid;
+  grid-template-columns: 1fr 1fr;
+  gap: 6px 12px;
+  margin-top: 6px;
+  font-size: 12px;
+}
+.hc-row {
+  display: flex; align-items: center; gap: 6px;
+}
+.hc-row.hc-num {
+  flex-direction: column; align-items: flex-start;
+}
+.hc-row.hc-num label { font-size: 11px; color: #aaa; }
+.hc-row.hc-num input { width: 100%; }
+
+/* Edge preview panel */
+.ep-grid {
+  display: grid;
+  grid-template-columns: 1fr 1fr;
+  gap: 6px 12px;
+  margin-top: 6px;
+  font-size: 12px;
+}
+.ep-row {
+  display: flex; flex-direction: column; gap: 2px;
+  font-size: 11px; color: #aaa;
+}
+.ep-row input[type="range"] { width: 100%; }
+.ep-row span { color: #fff; font-weight: bold; font-family: monospace; }

pyproject.toml

@@ -12,6 +12,10 @@ dependencies = [
     "uvicorn[standard]>=0.34",
 ]
 
+[project.scripts]
+pm2d-eval = "pm2d.eval:main"
+pm2d-bench = "pm2d.bench:main"
+
 [dependency-groups]
 dev = [
     "httpx>=0.28.1",