fix(mining): 修复 Windows 挖掘锁状态误报

锁诊断优先使用当前后端实际持锁对象确认自身归属，避免 Windows 排他锁阻止二次读取文件时误报未知持有者。真实外部占用仍按陈旧判定处理，并拒绝终止当前持锁后端。

验证：83 项相关 pytest 通过，Ruff 和 git diff --check 通过。原生 Windows 安装包仍待实机复验。

Signed-off-by: shy3130 <shy3130@qq.com>

backend/app/api/mining.py

@@ -169,7 +169,10 @@ def get_lock_status(request: Request) -> dict[str, Any]:
     """挖掘锁诊断: 持有者线索 + 是否允许强制释放 (fail-closed 判定)。"""
     from app.services.mining_lock_admin import lock_status
 
-    return lock_status(request.app.state.repo.store.data_dir)
+    return lock_status(
+        request.app.state.repo.store.data_dir,
+        lock=getattr(request.app.state, "mining_process_lock", None),
+    )
 
 
 @router.post("/lock/force-release")

backend/app/services/mining_lock_admin.py

@@ -49,23 +49,30 @@ class MiningLockAdminError(ValueError):
 def lock_status(
     data_dir: Path | str,
     *,
+    lock: MiningProcessLock | None = None,
     self_pid: int | None = None,
     now: float | None = None,
     stale_after: float = STALE_AFTER_SECONDS,
 ) -> dict[str, Any]:
     """汇总挖掘锁状态: 是否被持有、持有者线索、是否允许强制释放。"""
     pid = os.getpid() if self_pid is None else self_pid
-    held = probe_lock_held(data_dir)
+    owns_lock = lock is not None and lock.acquired
+    held = owns_lock or probe_lock_held(data_dir)
     holder = read_lock_holder(data_dir)
     holder_source = "file" if holder else None
-    if held and holder is None:
+    if owns_lock and (holder is None or holder.get("pid") != pid):
+        # Windows denies reads of the locked byte through a second handle,
+        # even in this process. The live owned handle is authoritative.
+        holder = {"pid": pid, "role": "app"}
+        holder_source = None
+    if held and not owns_lock and holder is None:
         # 旧格式/损坏内容 (典型: 旧版本残留后端进程): 从 OS 层面反查持有者,
         # 让这类锁也能走人工确认的强制释放, 而不是让用户手动猜进程。
         located_pid = locate_lock_holder_pid(data_dir)
         if located_pid is not None:
             holder = {"pid": located_pid, "role": "unknown"}
             holder_source = "os-scan"
-    mine = held and bool(holder) and holder.get("pid") == pid
+    mine = owns_lock or (held and bool(holder) and holder.get("pid") == pid)
     holder_pid = holder.get("pid") if holder else None
     holder_alive = _pid_alive(holder_pid) if holder_pid else False
     pid_consistent = (
@@ -109,7 +116,7 @@ def force_release(
     stale_after: float = STALE_AFTER_SECONDS,
 ) -> dict[str, Any]:
     """终止陈旧持有者并让本进程接管锁; 任何不满足条件的情况直接拒绝。"""
-    status = lock_status(data_dir, stale_after=stale_after)
+    status = lock_status(data_dir, lock=lock, stale_after=stale_after)
     if not status["force_releasable"]:
         raise MiningLockAdminError(status["reason"] or "不可强制释放")
     if lock.acquired:

backend/tests/test_mining_lock_admin.py

@@ -301,6 +301,47 @@ def test_lock_endpoints_free_lock(tmp_path):
     assert "未被持有" in resp.json()["detail"]
 
 
+def test_lock_endpoint_recognizes_own_lock_when_metadata_read_is_denied(tmp_path, monkeypatch):
+    """Windows forbids a second handle from reading the locked first byte."""
+    client = _api_client(tmp_path, acquired=True)
+    lock = client.app.state.mining_process_lock
+    lock.acquire()
+    original_read = Path.read_bytes
+
+    def read_bytes(path):
+        if path == lock.path:
+            raise PermissionError("locked byte is unreadable through another handle")
+        return original_read(path)
+
+    monkeypatch.setattr(Path, "read_bytes", read_bytes)
+    monkeypatch.setattr("app.services.mining_lock_admin.locate_lock_holder_pid", lambda _: None)
+    try:
+        body = client.get("/api/backtest/mining/lock").json()
+        assert body["held"] is True
+        assert body["mine"] is True
+        assert body["holder"]["pid"] == os.getpid()
+        assert body["force_releasable"] is False
+        assert "当前后端" in body["reason"]
+        refused = client.post("/api/backtest/mining/lock/force-release")
+        assert refused.status_code == 409
+        assert "当前后端" in refused.json()["detail"]
+        assert lock.acquired is True
+    finally:
+        lock.release()
+
+
+def test_lock_endpoint_does_not_treat_startup_flag_as_current_ownership(tmp_path, monkeypatch):
+    client = _api_client(tmp_path, acquired=True)
+    monkeypatch.setattr("app.services.mining_lock_admin.probe_lock_held", lambda _: True)
+    monkeypatch.setattr("app.services.mining_lock_admin.read_lock_holder", lambda _: None)
+    monkeypatch.setattr("app.services.mining_lock_admin.locate_lock_holder_pid", lambda _: None)
+    body = client.get("/api/backtest/mining/lock").json()
+    assert body["held"] is True
+    assert body["mine"] is False
+    assert body["holder"] is None
+    assert body["force_releasable"] is False
+
+
 @_POSIX
 def test_lock_endpoints_full_recovery_loop(tmp_path):
     proc = _spawn_holder(tmp_path)

docs/mining.md

@@ -110,6 +110,8 @@ run 级的“有效折”统计只描述因子赛道：它按每折被选中的
 
 **挖掘锁冲突的降级与强制释放**：数据目录的 `.mining_process.lock`（flock）保证同一时刻只有一个应用进程负责挖掘；持锁者会把 `{pid, started_at, role}` 写进锁文件（flock 仍是唯一事实，内容只是诊断线索，旧格式 `"0"` 视为持有者未知）。撞锁不再导致后端启动失败——面板以降级模式运行：挖掘启动入口（手动/自动/周度调度）返回 409，守护复验等其他功能不受影响；挖掘页顶部显示持有者横幅（pid、角色、挖掘存储静默时长）。当且仅当满足陈旧判定（持有者信息可读、进程存活且 pid 未被复用、无进行中的挖掘运行、挖掘存储 10 分钟内无写入——全部 fail-closed，任一不满足即拒绝并给出原因）时，可一键「强制停止并接管」：终止持有者进程并由当前后端获取锁，挖掘入口立即恢复。不做启动时自动清理——终止进程是破坏性操作，必须人工确认。
 
+锁状态接口优先使用当前后端实际持有的锁句柄确认归属。Windows 的字节范围排他锁会阻止另一个句柄读取锁文件，即使两个句柄属于同一进程；读取诊断信息失败时，不能把已持锁的当前后端误报成未知残留进程。真正的外部占用仍按上述陈旧判定处理，启动标志本身不能代替实际持锁状态。
+
 每个运行目录包含 manifest、compact summary、最近 256 条事件和四个 Parquet artifact。大面板、完整矩阵对象和逐折宽结果不会通过 worker IPC 返回。
 
 ## 自动运行