diff --git a/tools/import_thirdparty_catalog_csv.py b/tools/import_thirdparty_catalog_csv.py
index 4ac1212..5385e1d 100644
--- a/tools/import_thirdparty_catalog_csv.py
+++ b/tools/import_thirdparty_catalog_csv.py
@@ -32,6 +32,7 @@ def product_from(row):
 def sku_from(row):
     return {'sku_id':row['shopee_sku_id'],'goods_id':row['shopee_goods_id'],'spec_raw':row['spec_raw'],'color':row['color'],'size':row['size'],'advice':row['advice'],'parse_ok':row['parse_ok']=='true','sku_code':row['shopee_sku_code']}
 def pdd_from(row): return {'goods_id':row['pdd_goods_id'],'url':row['pdd_goods_url'],'title':'','shop_name':'','dimensions':[],'skus':[]}
+def spec_key(raw): return ' '.join(raw.split())
 
 def groups(root):
     """流式读 CSV；来源按商品连续导出，同商品不会一次占用整个文件。"""
@@ -49,19 +50,30 @@ def groups(root):
 def make_group(rows):
     first=rows[0]; skus=[]; specs=set(); pdds={}; associations=[]
     for row in rows:
-        if row['spec_raw'] and row['spec_raw'] not in specs:
-            specs.add(row['spec_raw']); skus.append(sku_from(row))
+        key=spec_key(row['spec_raw'])
+        if key and key not in specs:
+            specs.add(key); skus.append(sku_from(row))
         if row['pdd_importable']=='true' and row['pdd_goods_id'] not in pdds:
             pdds[row['pdd_goods_id']]=pdd_from(row)
             associations.append({'shopee_goods_id':first['shopee_goods_id'],'pdd_goods_id':row['pdd_goods_id']})
     return product_from(first),skus,pdds,associations
 
-def batches(root, observed_at, dry_run):
+def batches(root, observed_at, dry_run, keep_existing):
     selected=[]; seen_pdd={}; sku_count=0
     def emit():
         nonlocal selected,seen_pdd,sku_count
-        products=[g[0] for g in selected]; skus=[s for g in selected for s in g[1]]
-        pdds=list(seen_pdd.values()); assocs=[a for g in selected for a in g[3]]
+        # 同一商品可能在不同源 CSV 再次出现；接口禁止批内重复身份，
+        # 所以在真正组装请求时按四类业务键再去重一次。
+        product_by_id={g[0]['goods_id']:g[0] for g in selected}
+        sku_by_key={}
+        association_by_key={}
+        for group in selected:
+            for sku in group[1]: sku_by_key.setdefault((sku['goods_id'],spec_key(sku['spec_raw'])),sku)
+            for association in group[3]:
+                if association['shopee_goods_id'] not in keep_existing:
+                    association_by_key.setdefault((association['shopee_goods_id'],association['pdd_goods_id']),association)
+        products=list(product_by_id.values()); skus=list(sku_by_key.values())
+        pdds=list(seen_pdd.values()); assocs=list(association_by_key.values())
         seed=json.dumps([products,skus,pdds,assocs],ensure_ascii=False,sort_keys=True,separators=(',',':')).encode()
         payload={'schema_version':1,'batch_id':'thirdparty-csv-v1-'+hashlib.sha256(seed).hexdigest()[:24],'observed_at':observed_at,'update_policy':'fill_missing','dry_run':dry_run,'shopee_products':products,'shopee_skus':skus,'pdd_products':pdds,'associations':assocs}
         if len(json.dumps(payload,ensure_ascii=False,separators=(',',':')).encode()) > MAX_BYTES:
@@ -75,18 +87,19 @@ def batches(root, observed_at, dry_run):
     if selected: yield emit()
 
 def main():
-    p=argparse.ArgumentParser(); p.add_argument('csv_dir'); p.add_argument('--base-url',required=True); p.add_argument('--observed-at',default='2026-08-20T00:00:00+08:00'); mode=p.add_mutually_exclusive_group(required=True); mode.add_argument('--dry-run',action='store_true'); mode.add_argument('--apply',action='store_true'); p.add_argument('--report',required=True); a=p.parse_args()
+    p=argparse.ArgumentParser(); p.add_argument('csv_dir'); p.add_argument('--base-url',required=True); p.add_argument('--observed-at',default='2026-08-20T00:00:00+08:00'); mode=p.add_mutually_exclusive_group(required=True); mode.add_argument('--dry-run',action='store_true'); mode.add_argument('--apply',action='store_true'); p.add_argument('--report',required=True); p.add_argument('--keep-existing-association',action='append',default=[],help='保留该蝦皮商品的数据库既有关联，不写来源新关联；可重复传入'); a=p.parse_args()
     token=os.environ.get('CMAUTOBUY_CATALOG_TOKEN','').strip()
     if not token: error('缺少环境变量 CMAUTOBUY_CATALOG_TOKEN')
     root=Path(a.csv_dir)
     endpoint=a.base_url.rstrip('/')+'/api/v1/integrations/catalog/batches'
     total=Counter(); conflicts=[]; count=0
-    for payload in batches(root,a.observed_at,a.dry_run):
+    keep_existing={value.strip() for value in a.keep_existing_association if value.strip()}
+    for payload in batches(root,a.observed_at,a.dry_run,keep_existing):
         count+=1; response=request(endpoint,token,payload)
         if response.get('status') not in {'previewed','succeeded'}: error('接口未返回预期状态')
         total.update(response.get('counts',{})); conflicts.extend(response.get('conflicts',[]))
         print(f"[{count}] {response['status']} {payload['batch_id']}")
-    Path(a.report).write_text(json.dumps({'mode':'preview' if a.dry_run else 'apply','batches':count,'counts':total,'association_conflicts':conflicts},ensure_ascii=False,indent=2),encoding='utf-8')
+    Path(a.report).write_text(json.dumps({'mode':'preview' if a.dry_run else 'apply','batches':count,'counts':total,'kept_existing_association_goods_ids':sorted(keep_existing),'association_conflicts':conflicts},ensure_ascii=False,indent=2),encoding='utf-8')
     print(f"完成：{count} 批；报告：{a.report}")
 if __name__=='__main__':
     try: main()