perf: boucle fermée pHash (2s→150ms) + batch CLIP (90 appels→1)

Boucle fermée : time.sleep(2.0) remplacé par _wait_for_screen_change()
qui poll le pHash toutes les 150ms. Sort dès que l'écran change.
4 occurrences remplacées.

Batch CLIP : filtre par distance AVANT le CLIP (90→~20 éléments),
puis embed_image_batch() en un seul appel GPU + np.dot vectorisé.

Estimé : 42s→~20s total workflow.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

visual_workflow_builder/backend/api_v3/execute.py

@@ -24,6 +24,48 @@ from . import api_v3_bp
 logger = logging.getLogger(__name__)
 
 
+def _wait_for_screen_change(pre_screen, max_wait=2.0, poll_interval=0.15):
+    """
+    Attend que l'écran change après un clic, au lieu d'un sleep fixe.
+    Compare le pHash de l'écran actuel avec le screenshot pré-clic.
+    Sort dès que la distance de Hamming >= 5 (changement détecté).
+    Fallback: sort après max_wait secondes si aucun changement.
+
+    Args:
+        pre_screen: PIL Image du screenshot avant le clic (peut être None)
+        max_wait: Temps max d'attente en secondes (défaut 2.0)
+        poll_interval: Intervalle de polling en secondes (défaut 0.15)
+    """
+    if pre_screen is None:
+        # Pas de screenshot pré-clic, fallback sur sleep classique
+        time.sleep(max_wait)
+        return
+
+    try:
+        from core.analytics.screen_change_detector import compute_phash
+        import mss as _mss
+        from PIL import Image as _PILImage
+
+        _pre_hash = compute_phash(pre_screen)
+        _start = time.time()
+
+        while time.time() - _start < max_wait:
+            time.sleep(poll_interval)
+            try:
+                with _mss.mss() as _sct:
+                    _grab = _sct.grab(_sct.monitors[0])
+                    _post_screen = _PILImage.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+                _post_hash = compute_phash(_post_screen)
+                _dist = _pre_hash - _post_hash
+                if _dist >= 5:  # Écran a changé
+                    break
+            except Exception:
+                break
+    except Exception:
+        # Si pHash/mss non disponible, fallback sur sleep
+        time.sleep(max_wait)
+
+
 from core.execution.input_handler import (
     safe_type_text as _shared_safe_type_text,
     check_screen_for_patterns as _shared_check_patterns,
@@ -722,6 +764,17 @@ def execute_action_with_coords(action_type: str, params: dict, coords: dict) ->
         x, y = coords['x'], coords['y']
         print(f"🖱️ [Self-Healing] Clic aux coordonnées choisies: ({x}, {y})")
 
+        # Capture pré-clic pour détection de changement
+        _pre_screen = None
+        try:
+            import mss as _mss
+            from PIL import Image as _PILImage
+            with _mss.mss() as _sct:
+                _grab = _sct.grab(_sct.monitors[0])
+                _pre_screen = _PILImage.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+        except Exception:
+            pass
+
         if action_type in ['double_click_anchor']:
             pyautogui.doubleClick(x, y)
         elif action_type in ['right_click_anchor']:
@@ -729,7 +782,7 @@ def execute_action_with_coords(action_type: str, params: dict, coords: dict) ->
         else:
             pyautogui.click(x, y)
 
-        time.sleep(2.0)  # Délai après le clic
+        _wait_for_screen_change(_pre_screen)
 
         return {
             'success': True,
@@ -759,6 +812,17 @@ def execute_action_with_static_coords(action_type: str, params: dict) -> dict:
 
         print(f"🖱️ [Self-Healing] Clic aux coordonnées statiques: ({x}, {y})")
 
+        # Capture pré-clic pour détection de changement
+        _pre_screen = None
+        try:
+            import mss as _mss
+            from PIL import Image as _PILImage
+            with _mss.mss() as _sct:
+                _grab = _sct.grab(_sct.monitors[0])
+                _pre_screen = _PILImage.frombytes('RGB', _grab.size, _grab.bgra, 'raw', 'BGRX')
+        except Exception:
+            pass
+
         if action_type in ['double_click_anchor']:
             pyautogui.doubleClick(x, y)
         elif action_type in ['right_click_anchor']:
@@ -766,7 +830,7 @@ def execute_action_with_static_coords(action_type: str, params: dict) -> dict:
         else:
             pyautogui.click(x, y)
 
-        time.sleep(2.0)
+        _wait_for_screen_change(_pre_screen)
 
         return {
             'success': True,
@@ -844,6 +908,7 @@ def execute_action(action_type: str, params: dict) -> dict:
                     _fc_target_desc = params.get('visual_anchor', {}).get('description', '')
 
                     x, y, confidence, method_used = None, None, 0, ''
+                    screen_img = None  # Screenshot pré-clic pour détection de changement
 
                     # === MÉTHODE 1 : Template matching direct (~1-10ms) ===
                     try:
@@ -922,7 +987,7 @@ def execute_action(action_type: str, params: dict) -> dict:
                         else:
                             pyautogui.click(x, y)
 
-                        time.sleep(2.0)
+                        _wait_for_screen_change(screen_img)
 
                         return {
                             'success': True,
@@ -969,7 +1034,7 @@ def execute_action(action_type: str, params: dict) -> dict:
                                 else:
                                     pyautogui.click(gx, gy)
 
-                                time.sleep(2.0)
+                                _wait_for_screen_change(screen_img)
 
                                 return {
                                     'success': True,