fix(skill-evolution): strip non-user-configurable fields from per-user overrides

load_user_overrides 只保留 interval_hours/stale_after_days/archive_after_days，
过滤掉老文件中的 enabled、model_name 等字段，防止旧数据覆盖全局开关。

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

offline-backend-20260512/backend/packages/harness/deerflow/config/skill_evolution_runtime.py

@@ -135,8 +135,18 @@ def resolve_user_path(user_id: str) -> Path:
     return get_paths().user_dir(user_id) / _FILENAME
 
 
+# Only these curator sub-fields may be overridden per user.
+# All other fields (enabled, model_name, timeouts, etc.) are global-only.
+_USER_ALLOWED_CURATOR_KEYS = {"interval_hours", "stale_after_days", "archive_after_days"}
+
+
 def load_user_overrides(user_id: str) -> dict[str, Any] | None:
-    """Read the per-user overrides file for ``user_id`` (None if missing/invalid)."""
+    """Read the per-user overrides file for ``user_id`` (None if missing/invalid).
+
+    Only the three curator threshold keys are honoured; any other keys
+    (e.g. legacy ``enabled`` or ``model_name``) are silently stripped so that
+    old files written by previous versions cannot override global-only settings.
+    """
     path = resolve_user_path(user_id)
     if not path.exists():
         return None
@@ -154,7 +164,14 @@ def load_user_overrides(user_id: str) -> dict[str, Any] | None:
     if not isinstance(data, dict):
         logger.warning("User skill_evolution overrides at %s must be a JSON object", path)
         return None
-    return data
+    # Strip top-level keys that are not user-configurable (e.g. legacy "enabled").
+    curator_raw = data.get("curator")
+    if not isinstance(curator_raw, dict):
+        return None
+    filtered_curator = {k: v for k, v in curator_raw.items() if k in _USER_ALLOWED_CURATOR_KEYS}
+    if not filtered_curator:
+        return None
+    return {"curator": filtered_curator}
 
 
 def save_user_overrides(user_id: str, payload: dict[str, Any]) -> Path: