fix(concurrency): 消除 model 覆盖的模块级变量中转(并发串味)

R4 首版实现用模块级 _ACTIVE_MODEL_OVERRIDE 中转 model override,
理由是不想改动 run_specialists 的签名(既有测试会 mock 它)。

但 backtest.py:225 会 asyncio.gather 并发 8 场预测(Semaphore(8)),
每场都调用 predict_match_multi —— 模块级变量会被并发调用互相覆盖,
导致 A 场的预测用上 B 场的模型。这是静默的正确性缺陷。

改为: run_specialists 新增 model_override 形参,一路显式下传;
删除模块级变量。同步更新 6 处 mock 签名。

新增 3 个守卫并做变异验证:
  - test_r4_no_module_level_model_override_global (源码级,可判别)
  - test_r4_run_specialists_accepts_model_override_parameter
  - test_r4_dispatch_passes_override_to_specialists
变异测试: 重新引入全局变量方案 → 两个守卫变红;还原 → 全绿。

注: 曾尝试写并发行为测试,但真实 run_specialists 会访问数据库,
测试环境下不稳定(ConnectionRefusedError),会是 flaky 的假证据,
故改用清晰的源码级判别守卫,并在注释中说明原因。

240 passed / 1 skipped; tsc 0 error; vite build 成功。

tests/test_multi_agent_cutoff.py

@@ -43,7 +43,7 @@ class TestMultiAgentCutoffPropagation:
 
         captured_before = []
 
-        async def mock_run_specialists(header, *, version, before=None):
+        async def mock_run_specialists(header, *, version, before=None, model_override=None):
             captured_before.append(before)
             return []
 
@@ -86,7 +86,7 @@ class TestMultiAgentCutoffPropagation:
 
         captured_before = []
 
-        async def mock_run_specialists(header, *, version, before=None):
+        async def mock_run_specialists(header, *, version, before=None, model_override=None):
             captured_before.append(before)
             return []
 
@@ -118,7 +118,7 @@ class TestMultiAgentCutoffPropagation:
 
         captured_before = []
 
-        async def mock_run_specialists(header, *, version, before=None):
+        async def mock_run_specialists(header, *, version, before=None, model_override=None):
             captured_before.append(before)
             return []