diff --git a/backend/app/engine/nodes/metrics.py b/backend/app/engine/nodes/metrics.py
index b13d9878..9e41cd4f 100644
--- a/backend/app/engine/nodes/metrics.py
+++ b/backend/app/engine/nodes/metrics.py
@@ -33,6 +33,7 @@ from app.engine.expressions import (
     substitute,
 )
 from app.engine.labels import field_label
+from app.engine.merge_query import MERGE_SOURCE
 from app.engine.state import GraphState, template_context
 
 
@@ -393,13 +394,29 @@ def _result_artifact(path: str, tctx: dict[str, Any]) -> str | None:
     return None
 
 
-def _error_checks(artifact: str) -> list[dict[str, Any]]:
+def _error_checks(artifact: str, _seen: set[str] | None = None) -> list[dict[str, Any]]:
+    """这份查询快照对照数据目录查出的 error 级问题。
+
+    合并查询的快照没有自己的 checks（它在内存 SQLite 上执行，不对照数据目录），但记着 inputs：顺着往下找。
+    不然源库那条查询一对多关联后重复计算了，经过一次合并，指标就看不出来、照样完整出具。
+    """
+    seen = _seen if _seen is not None else set()
+    if artifact in seen:
+        return []
+    seen.add(artifact)
     try:
         snapshot = artifact_store.load(artifact)
     except Exception:  # noqa: BLE001 - 快照读不出来、被改过：这里不报，证据层另有说法
         return []
-    checks = snapshot.get("checks") if isinstance(snapshot, dict) else None
-    return [c for c in checks if isinstance(c, dict) and c.get("level") == "error"] if isinstance(checks, list) else []
+    if not isinstance(snapshot, dict):
+        return []
+    checks = snapshot.get("checks")
+    found = [c for c in checks if isinstance(c, dict) and c.get("level") == "error"] if isinstance(checks, list) else []
+    inputs = snapshot.get("inputs") if snapshot.get("source") == MERGE_SOURCE else None
+    for entry in inputs if isinstance(inputs, list) else []:
+        if isinstance(entry, dict) and isinstance(entry.get("artifact"), str):
+            found += _error_checks(entry["artifact"], seen)
+    return found
 
 
 def _sql_check_reason(failed: list[tuple[str, list[dict[str, Any]]]]) -> str:
diff --git a/backend/tests/test_merge_sqlcheck.py b/backend/tests/test_merge_sqlcheck.py
new file mode 100644
index 00000000..1be800ea
--- /dev/null
+++ b/backend/tests/test_merge_sqlcheck.py
@@ -0,0 +1,60 @@
+"""合并查询把输入查询的 SQL 检查结果传到指标上。
+
+口径卡判「来源查询有没有 error 级问题」读的是来源快照上的 checks（data/sqlcheck.py 在数据源查询时写）。
+合并查询在内存 SQLite 上执行，不对照数据目录检查，它的快照没有 checks——于是源库那条查询明明一对多关联后
+重复计算了，经过一次合并，指标就看不出来了，照样完整出具。合并快照记着 inputs（每个输入的快照 id），
+顺着它往下找才对得上。
+"""
+from __future__ import annotations
+
+import app.engine.nodes.metrics as metrics_mod
+from app.engine.merge_query import MERGE_SOURCE
+from app.engine.nodes.metrics import _error_checks
+
+FANOUT = {"code": "fanout_sum", "level": "error", "message": "「订单」关联「订单明细」是一对多，求和会重复计算"}
+NOTICE = {"code": "missing_valid_filter", "level": "warning", "message": "没有过滤作废记录"}
+
+
+def _query(rows, checks=None):
+    snap = {"columns": ["日期", "门店", "销售额"], "rows": rows, "row_count": len(rows), "truncated": False}
+    return {**snap, "checks": checks} if checks else snap
+
+
+def _merged(inputs):
+    return {"columns": ["日期", "门店", "销售额"], "rows": [["2026-05-01", "S01", 10]], "row_count": 1,
+            "truncated": False, "sql": "SELECT * FROM s JOIN v USING (日期, 门店)", "source": MERGE_SOURCE,
+            "inputs": [{"alias": a, "node_id": f"q_{a}", "artifact": art, "rows": 1} for a, art in inputs]}
+
+
+def _store(monkeypatch, snapshots: dict):
+    monkeypatch.setattr(metrics_mod.artifact_store, "load", lambda artifact: snapshots.get(artifact))
+
+
+def test_an_error_in_a_merge_input_reaches_the_metric(monkeypatch):
+    _store(monkeypatch, {
+        "sales": _query([["2026-05-01", "S01", 10]], [FANOUT, NOTICE]),
+        "visits": _query([["2026-05-01", "S01", 3]]),
+        "merged": _merged([("s", "sales"), ("v", "visits")]),
+    })
+    assert [c["code"] for c in _error_checks("merged")] == ["fanout_sum"]
+
+
+def test_clean_inputs_leave_the_merge_clean(monkeypatch):
+    _store(monkeypatch, {"sales": _query([["2026-05-01", "S01", 10]], [NOTICE]),
+                         "visits": _query([["2026-05-01", "S01", 3]]),
+                         "merged": _merged([("s", "sales"), ("v", "visits")])})
+    assert _error_checks("merged") == []
+
+
+def test_merges_of_merges_are_followed_and_cycles_do_not_hang(monkeypatch):
+    _store(monkeypatch, {
+        "sales": _query([["2026-05-01", "S01", 10]], [FANOUT]),
+        "inner": _merged([("s", "sales")]),
+        "outer": _merged([("m", "inner"), ("self", "outer")]),     # 自引用：工件内容寻址做不出来，读坏的库也不能卡死
+    })
+    assert [c["code"] for c in _error_checks("outer")] == ["fanout_sum"]
+
+
+def test_a_plain_query_snapshot_is_read_as_before(monkeypatch):
+    _store(monkeypatch, {"q": _query([["2026-05-01", "S01", 10]], [FANOUT, NOTICE])})
+    assert [c["code"] for c in _error_checks("q")] == ["fanout_sum"]