feat(gateway): worker 池主备分组派发，主全挂才用备机

- pool.py: WorkerState 增加 primary 标记；acquire_worker 分级派发——
  主组有在线只在主组内挑空闲（全忙排队，不退备组），主组全部下线才用备机
- config.py: 新增可选 primary_workers 配置（须为 workers 子集，自动过滤+告警）；
  不配置/空列表 = 全部视为主 worker，兼容旧行为
- config.json(不入 git): 3 台 GPU worker 设为主，max_global_concurrency 1→3
- app.py: /health 增加 workers_primary_total / workers_primary_healthy
- tests: 新增 4 条主备派发单测（主优先/主全挂用备/主全忙不走备/无配置兼容）
- docs(实现说明): 补充 primary_workers 配置说明，改配置需重启网关生效

tests/test_gateway_pool.py

@@ -152,3 +152,85 @@ def test_cooldown_seconds_falls_back_to_60(monkeypatch):
     from gateway import config as gw_config
     monkeypatch.setattr(gw_config, "_config_cache", None)  # get_config 会抛 RuntimeError
     assert pool._cooldown_seconds() == 60.0
+
+
+# ---------------------------------------------------------------------------
+# 主备分组派发（primary_workers）
+# ---------------------------------------------------------------------------
+
+_DISPATCH_CFG = {
+    "dispatch": {
+        "queue_wait_seconds": 0.05,  # 测试用：快速超时
+    }
+}
+
+
+def _make_pool(specs):
+    """specs: [(url, online, busy, primary), ...] → 构造 pool._workers。"""
+    return {
+        url: WorkerState(url=url, online=online, busy=busy, primary=primary)
+        for url, online, busy, primary in specs
+    }
+
+
+def _run_acquire(workers):
+    async def run():
+        pool._pool_condition = asyncio.Condition()
+        orig = pool._workers
+        pool._workers = workers
+        try:
+            return await pool.acquire_worker(_DISPATCH_CFG)
+        finally:
+            pool._workers = orig
+    return asyncio.run(run())
+
+
+def test_primary_preferred_over_backup():
+    """主 worker 在线时空闲主优先于备 worker。"""
+    workers = _make_pool([
+        ("http://backup:1", True, False, False),
+        ("http://primary:1", True, False, True),
+    ])
+    w = _run_acquire(workers)
+    assert w.url == "http://primary:1"
+    assert w.busy is True
+
+
+def test_backup_used_only_when_all_primary_down():
+    """主 worker 全部下线时才用备 worker。"""
+    workers = _make_pool([
+        ("http://backup:1", True, False, False),
+        ("http://primary:1", False, False, True),
+        ("http://primary:2", False, False, True),
+    ])
+    w = _run_acquire(workers)
+    assert w.url == "http://backup:1"
+
+
+def test_backup_not_used_when_primary_busy():
+    """主 worker 在线但全忙时不退到备 worker（等待后超时拒绝）。"""
+    workers = _make_pool([
+        ("http://backup:1", True, False, False),
+        ("http://primary:1", True, True, True),
+    ])
+
+    async def run():
+        pool._pool_condition = asyncio.Condition()
+        try:
+            await pool.acquire_worker(_DISPATCH_CFG)
+        except pool.NoWorkerAvailable:
+            return True
+        return False
+
+    assert asyncio.run(run()) is True, "主组全忙时不得使用备 worker"
+    assert workers["http://backup:1"].busy is False
+
+
+def test_no_primary_config_all_are_primary():
+    """未配置 primary_workers 时全部视为主 worker（兼容旧行为）。"""
+    workers = _make_pool([
+        ("http://w1:1", True, False, True),
+        ("http://w2:1", True, False, True),
+    ])
+    w = _run_acquire(workers)
+    assert w.url == "http://w1:1"