fix: by_text dans build_replay + anchor matching pour icônes sans texte

build_replay (stream_processor.py) :
- Remplir by_text depuis vision_info.text ou som_element.label
- VLM identification pour les éléments sans texte (icônes)
- Nettoyage du bavardage VLM (retrait préfixes courants)

resolve_target (api_stream.py) :
- Nouveau som_anchor_match : template matching du crop anchor vs régions YOLO
- Pour les icônes sans texte (disquette, loupe, etc.)
- Cascade : text match → anchor match → VLM

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

agent_v0/server_v1/api_stream.py

@@ -3528,6 +3528,74 @@ def _resolve_by_som(
                 len(exact_matches), anchor_label,
             )
 
+    # ── 2.7. Fallback : template matching anchor vs éléments SomEngine ──
+    # Pour les icônes sans texte : comparer le crop de référence contre
+    # chaque région YOLO détectée par SomEngine.
+    anchor_b64 = target_spec.get("anchor_image_base64", "")
+    if anchor_b64 and not anchor_label:
+        try:
+            import cv2
+            import numpy as np
+
+            # Décoder l'anchor
+            anc_bytes = base64.b64decode(anchor_b64)
+            anc_array = np.frombuffer(anc_bytes, dtype=np.uint8)
+            anc_img = cv2.imdecode(anc_array, cv2.IMREAD_GRAYSCALE)
+
+            # Charger le screenshot en OpenCV
+            screenshot_cv = cv2.imread(screenshot_path, cv2.IMREAD_GRAYSCALE)
+
+            if anc_img is not None and screenshot_cv is not None:
+                best_elem = None
+                best_score = 0.0
+                anc_h, anc_w = anc_img.shape[:2]
+
+                for elem in som_result.elements:
+                    x1, y1, x2, y2 = elem.bbox
+                    # Agrandir la zone de 20% pour tolérer les différences
+                    margin_x = int((x2 - x1) * 0.2)
+                    margin_y = int((y2 - y1) * 0.2)
+                    rx1 = max(0, x1 - margin_x)
+                    ry1 = max(0, y1 - margin_y)
+                    rx2 = min(screenshot_cv.shape[1], x2 + margin_x)
+                    ry2 = min(screenshot_cv.shape[0], y2 + margin_y)
+                    region = screenshot_cv[ry1:ry2, rx1:rx2]
+
+                    if region.shape[0] < anc_h or region.shape[1] < anc_w:
+                        continue
+
+                    res = cv2.matchTemplate(region, anc_img, cv2.TM_CCOEFF_NORMED)
+                    _, score, _, _ = cv2.minMaxLoc(res)
+                    if score > best_score:
+                        best_score = score
+                        best_elem = elem
+
+                if best_elem and best_score >= 0.6:
+                    elapsed = time.time() - t0
+                    cx_norm, cy_norm = best_elem.center_norm
+                    logger.info(
+                        "SoM resolve ANCHOR : match crop '#%d' score=%.3f → (%.4f, %.4f) en %.1fs",
+                        best_elem.id, best_score, cx_norm, cy_norm, elapsed,
+                    )
+                    return {
+                        "resolved": True,
+                        "method": "som_anchor_match",
+                        "x_pct": round(cx_norm, 6),
+                        "y_pct": round(cy_norm, 6),
+                        "matched_element": {
+                            "label": best_elem.label or f"icon #{best_elem.id}",
+                            "type": best_elem.source,
+                            "role": "som_anchor_match",
+                            "confidence": best_score,
+                            "som_id": best_elem.id,
+                        },
+                        "score": best_score,
+                    }
+        except ImportError:
+            pass
+        except Exception as e:
+            logger.debug("SoM anchor match erreur : %s", e)
+
     # ── 3. Sauvegarder l'image annotée SoM temporairement ──
     if som_result.som_image is None:
         logger.debug("SoM resolve : pas d'image annotée, skip VLM")