feat: résolution serveur pour replay Windows + VLM multi-image + métriques

Feature 4 — Résolution serveur :
- Nouvelle méthode _server_resolve_target() dans executor.py
- Cascade : template local → serveur /resolve_target → VLM local (fallback)
- Popup handling via serveur aussi
- L'agent Windows peut maintenant résoudre les clics via SomEngine+VLM

Feature 5 — VLM multi-image :
- _resolve_by_som() envoie l'anchor crop en 2ème image au VLM
- Le VLM voit les marks numérotés + le crop de l'élément recherché

Feature 6 — Métriques de résolution :
- resolution_method, resolution_score, resolution_elapsed_ms
- Propagés agent → serveur via /replay/result
- Résumé en fin de replay (méthodes, score moyen, temps moyen)

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

agent_v0/server_v1/api_stream.py

@@ -408,6 +408,10 @@ class ReplayResultReport(BaseModel):
     screenshot: Optional[str] = None  # Chemin ou base64 du screenshot post-action
     screenshot_after: Optional[str] = None  # Chemin ou base64 du screenshot APRES l'action
     actual_position: Optional[Dict[str, float]] = None  # {"x": px, "y": py} position réelle du clic
+    # Métriques de résolution visuelle
+    resolution_method: Optional[str] = None  # som_text_match, som_vlm, vlm_quick_find, etc.
+    resolution_score: Optional[float] = None
+    resolution_elapsed_ms: Optional[float] = None
 
 
 class ErrorCallbackConfig(BaseModel):
@@ -2286,6 +2290,9 @@ async def report_action_result(report: ReplayResultReport):
             "actual_position": report.actual_position,
             "retry_count": retry_count,
             "verification": verification.to_dict() if verification else None,
+            "resolution_method": report.resolution_method,
+            "resolution_score": report.resolution_score,
+            "resolution_elapsed_ms": report.resolution_elapsed_ms,
         }
         replay_state["results"].append(result_entry)
 
@@ -2384,6 +2391,30 @@ async def report_action_result(report: ReplayResultReport):
                 f" ({replay_state['retried_actions']} retries, "
                 f"{replay_state['unverified_actions']} non vérifiées)"
             )
+            # Résumé des métriques de résolution visuelle
+            results_with_method = [
+                r for r in replay_state["results"]
+                if r.get("resolution_method")
+            ]
+            if results_with_method:
+                methods_count = {}
+                total_elapsed = 0.0
+                total_score = 0.0
+                for r in results_with_method:
+                    m = r["resolution_method"]
+                    methods_count[m] = methods_count.get(m, 0) + 1
+                    total_elapsed += r.get("resolution_elapsed_ms") or 0
+                    total_score += r.get("resolution_score") or 0
+                avg_elapsed = total_elapsed / len(results_with_method)
+                avg_score = total_score / len(results_with_method)
+                methods_str = ", ".join(
+                    f"{m}={c}" for m, c in sorted(methods_count.items())
+                )
+                logger.info(
+                    f"Replay {replay_state['replay_id']} métriques résolution : "
+                    f"{len(results_with_method)} resolves [{methods_str}] "
+                    f"score_moy={avg_score:.2f} temps_moy={avg_elapsed:.0f}ms"
+                )
 
         # Libérer le GPU pour le worker VLM si le replay est terminé ou en erreur
         if replay_state["status"] in ("completed", "error"):
@@ -3506,12 +3537,26 @@ def _resolve_by_som(
         for e in labeled_elements
     )
 
-    prompt = (
-        f"I'm looking for: {target_desc}\n\n"
-        f"Here are the numbered elements detected on screen:\n{elements_list}\n\n"
-        "Which number is the correct element?\n"
-        'Answer with JSON only: {"mark_id": N, "confidence": 0.9}'
-    )
+    # Multi-image : SoM annotée + anchor crop (si disponible)
+    anchor_b64 = target_spec.get("anchor_image_base64", "")
+    extra_images = [anchor_b64] if anchor_b64 else None
+
+    if extra_images:
+        prompt = (
+            "Image 1 shows the screen with numbered marks on each UI element.\n"
+            "Image 2 shows the element I'm looking for.\n\n"
+            f"Target: {target_desc}\n\n"
+            f"Detected elements:\n{elements_list}\n\n"
+            "Which mark number matches the target element in Image 2?\n"
+            'Answer with JSON only: {"mark_id": N, "confidence": 0.9}'
+        )
+    else:
+        prompt = (
+            f"I'm looking for: {target_desc}\n\n"
+            f"Detected elements:\n{elements_list}\n\n"
+            "Which number is the correct element?\n"
+            'Answer with JSON only: {"mark_id": N, "confidence": 0.9}'
+        )
 
     system_prompt = "You identify UI elements by number. Output JSON only, no explanation."
 
@@ -3523,6 +3568,7 @@ def _resolve_by_som(
             temperature=0.1,
             max_tokens=50,
             force_json=False,
+            extra_images_b64=extra_images,
         )
     except Exception as e:
         logger.warning("SoM resolve : erreur VLM — %s", e)