feat: ocg_router 并发模式——X-OCG-Mode: concurrent 时 round-robin 锁 key 轮询(每次请求轮着用key,避免并发压同一key触发429/401), 非并发模式保持按用量选最空闲; 测试验证key1→key4→key5→key6轮询

gateway/scripts/ocg_router.py

@@ -70,6 +70,8 @@ _key_failure_reason: dict[str, str] = {}  # key_id → 最近失败原因（冷
 _route_stats: dict[str, int] = {}  # key_id → hit_count
 _total_hits = 0
 _current_key_id = ""
+# 2026-08-13 并发模式：round-robin 轮询索引（锁 key 轮着来）
+_rr_index = 0
 _state_lock = threading.Lock()
 
 
@@ -202,6 +204,35 @@ def pick_key():
     return best
 
 
+def pick_key_round_robin():
+    """并发模式：round-robin 轮询选 key（锁 key 轮着来）。
+    每次请求用下一个 key，避免并发压同一 key（2026-08-13 老莫设计）。
+    跳过不可用 key（失败冷却中/未订阅）。"""
+    global _rr_index
+    if not _keys:
+        return None
+    with _state_lock:
+        n = len(_keys)
+        if n == 0:
+            return None
+        # 轮询：从 _rr_index 开始找下一个可用 key
+        for offset in range(n):
+            idx = (_rr_index + offset) % n
+            k = _keys[idx]
+            kid = k["key_id"]
+            # 跳过不可用（失败冷却中/未订阅）
+            if time.time() < _key_failures.get(kid, 0):
+                continue
+            usage = _key_usage.get(kid, {})
+            if not usage.get("subscribed", True):
+                continue
+            # 选中：更新索引到下一个
+            _rr_index = (idx + 1) % n
+            return k
+        # 全部不可用 → 返回第一个（兜底）
+        return _keys[0]
+
+
 def mark_key_failed(kid, reason="", cooldown=FAIL_COOLDOWN_SEC):
     """标记 key 故障，进入冷却期。记录失败原因供冷却期报错。"""
     until = time.time() + cooldown
@@ -423,10 +454,37 @@ class RouterHandler(BaseHTTPRequestHandler):
             except (json.JSONDecodeError, UnicodeDecodeError) as e:
                 log.warning("failed to parse request body (len=%d): %s", len(body), e)
 
+        # 2026-08-13 并发模式：header X-OCG-Mode: concurrent → round-robin 锁 key 轮询
+        concurrent_mode = self.headers.get("X-OCG-Mode", "").lower() == "concurrent"
+
         # 尝试所有 key（按用量排序），直到成功
         with _state_lock:
             usage_snapshot = dict(_key_usage)
             failures_snapshot = dict(_key_failures)
+            if concurrent_mode:
+                # 并发模式：round-robin 轮询（锁 key 轮着来）
+                ranked_keys = []
+                n = len(_keys)
+                global _rr_index
+                for offset in range(n):
+                    idx = (_rr_index + offset) % n
+                    k = _keys[idx]
+                    kid = k["key_id"]
+                    if time.time() < failures_snapshot.get(kid, 0):
+                        continue
+                    usage = usage_snapshot.get(kid, {})
+                    if not usage.get("subscribed", True):
+                        continue
+                    ranked_keys.append(k)
+                # 更新轮询索引
+                if ranked_keys:
+                    first_kid = ranked_keys[0]["key_id"]
+                    for i, k in enumerate(_keys):
+                        if k["key_id"] == first_kid:
+                            _rr_index = (i + 1) % len(_keys)
+                            break
+                log.info("concurrent mode: round-robin, next_idx=%d, pool=%d", _rr_index, len(ranked_keys))
+            else:
                 ranked_keys = sorted(
                     list(_keys),
                     key=lambda k: _key_health_score(k["key_id"])