fix(agent): P0.6 guard human corrections

agent_v0/server_v1/replay_learner.py

@@ -194,8 +194,33 @@ class ReplayLearner:
             or target_spec.get("expected_window_before", "")
             or (target_spec.get("context_hints") or {}).get("window_title", "")
         )
-        x_pct = correction.get("x_pct", 0.0)
-        y_pct = correction.get("y_pct", 0.0)
+        x_pct = correction.get("x_pct")
+        y_pct = correction.get("y_pct")
+        last_click = correction.get("last_click")
+        if (x_pct is None or y_pct is None) and isinstance(last_click, dict):
+            x_pct = last_click.get("x_pct")
+            y_pct = last_click.get("y_pct")
+
+        try:
+            x_pct_f = float(x_pct)
+            y_pct_f = float(y_pct)
+        except (TypeError, ValueError):
+            logger.warning(
+                "[APPRENTISSAGE] Correction humaine non persistée : "
+                "aucune coordonnée clic exploitable pour '%s'",
+                by_text,
+            )
+            return
+
+        if not (0.0 < x_pct_f <= 1.0 and 0.0 < y_pct_f <= 1.0):
+            logger.warning(
+                "[APPRENTISSAGE] Correction humaine non persistée : "
+                "coordonnées hors bornes pour '%s' (%.4f, %.4f)",
+                by_text,
+                x_pct_f,
+                y_pct_f,
+            )
+            return
 
         # Enregistrer dans le JSONL d'apprentissage
         outcome = ActionOutcome(
@@ -218,8 +243,8 @@ class ReplayLearner:
                 stored = memory_record_success(
                     window_title=window_title,
                     target_spec=target_spec,
-                    x_pct=float(x_pct),
-                    y_pct=float(y_pct),
+                    x_pct=x_pct_f,
+                    y_pct=y_pct_f,
                     method="human_supervised",
                     confidence=1.0,
                 )