数据目录 F1-A1：对话提案确认关联关系时不再抹掉基数和覆盖率——覆盖率是剖析测得的，一律沿用；基数没写时沿用已有的，新增的按目标表的主键或唯一约束推算，推算不了留空并在提案这一项上说明（PatchChange.note）。以前剖析测得多对一的关系经提案确认后基数变空，一对多关联后重复计算的检查静默失效

backend/app/data/catalog.py

@@ -746,6 +746,8 @@ class PatchChange:
     - after：保存后的值（码值是补充后的完整对照）；value：要交回保存的原样取值（码值只有补充的那几个，
       409 之后在最新的目录上重新补）；
     - state：change 值有变化；confirm 值相同、但还不是已确认（保存即确认）；same 已经是这个值且已确认。
+    - note：服务端对这一项的说明，给人看；没有时为 None，as_dict 里也不出现。目前只有一种：新增的关联关系没写
+      基数、又推算不出来（_relation_after）。
     """
 
     path: str
@@ -755,10 +757,14 @@ class PatchChange:
     value: Any
     reason: str
     state: str
+    note: str | None = None
 
     def as_dict(self) -> dict[str, Any]:
-        return {"path": self.path, "before": self.before, "before_status": self.before_status, "after": self.after,
-                "value": self.value, "reason": self.reason, "state": self.state}
+        out = {"path": self.path, "before": self.before, "before_status": self.before_status, "after": self.after,
+               "value": self.value, "reason": self.reason, "state": self.state}
+        if self.note:
+            out["note"] = self.note
+        return out
 
 
 @dataclass(frozen=True)
@@ -785,7 +791,11 @@ def _formula_like(key: Slot, value: Any) -> bool:
 
 def _patch_relation(table: str, raw: Any, path_id: str | None, tables: Mapping[str, Any] | None,
                     columns: set[str] | None) -> tuple[dict[str, Any] | None, str | None]:
-    """关联关系的取值 → (关系的值部分, 问题)。值写两端（columns、to_table、to_columns），基数可选。"""
+    """关联关系的取值 → (关系的值部分, 问题)。值写两端（columns、to_table、to_columns），基数可选。
+
+    返回的是提案原样的取值：基数没写就是 None，覆盖率一律 None（模型写了也不认）。保存后的值由 _relation_after
+    并上目录里已有的测量结果再算。
+    """
     if not isinstance(raw, Mapping):
         return None, "关联关系的值应写明本表字段、目标表和目标字段"
     to_table = raw.get("to_table")
@@ -809,6 +819,35 @@ def _patch_relation(table: str, raw: Any, path_id: str | None, tables: Mapping[s
     return value, None
 
 
+def _relation_after(table: str, value: Mapping[str, Any], prev: Mapping[str, Any] | None,
+                    tables: Mapping[str, Any] | None) -> tuple[dict[str, Any], str | None]:
+    """提案里一条关系保存后的值：(值, 给人看的说明)。
+
+    以前直接拿提案的取值整份替换：覆盖率写成空、基数取提案原值（模型多半不写）。数据剖析测得多对一、覆盖率 100%
+    的关系，经对话「确认一下」就丢了基数，一对多关联后重复计算的检查跟着失效。所以：
+    - 覆盖率是数据剖析测出来的，一律沿用目录里这条关系的（包括被驳回的那条留下的），没有就是空；
+    - 基数：提案写了用提案的；没写时，目录里已有这条关系就沿用它的；新增的按目标表的主键或唯一约束推算
+      （目标列是键，本表这几列也是键就是一对一，否则多对一，同外键约束的 _cardinality）；推算不了就留空，
+      并说明为什么空着。
+    """
+    out = dict(value)
+    out["coverage"] = prev.get("coverage") if prev is not None else None
+    if out.get("cardinality") is not None:
+        return out, None
+    if prev is not None:
+        out["cardinality"] = prev.get("cardinality")
+        return out, None
+    meta = tables.get(table) if tables is not None else None
+    target = tables.get(str(out.get("to_table"))) if tables is not None else None
+    if not isinstance(meta, Mapping) or not isinstance(target, Mapping):
+        return out, "没有写明基数，表结构里也查不到两端的主键和唯一约束，基数暂时空着；一对多关联后重复计算的检查用不上这条关系"
+    if not _is_key(target, list(out.get("to_columns") or [])):
+        return out, (f"没有写明基数，目标字段不是 {out.get('to_table')} 的主键或唯一约束，推算不出，基数暂时空着；"
+                     "一对多关联后重复计算的检查用不上这条关系，可以在数据目录里补上基数")
+    out["cardinality"] = _cardinality(meta, list(out.get("columns") or []))
+    return out, None
+
+
 def _merge_codes(before: Mapping[str, str], value: Mapping[str, str]) -> dict[str, str]:
     """码值的补充：给出的码加进去或改掉原来的含义。含义为空表示「只知道有这个码、含义待填写」（数据剖析的码值
     候选就是这样），所以提案里含义为空的码不覆盖已有的含义，只在原来没有这个码时加进去。"""
@@ -833,6 +872,8 @@ def plan_patch(notes: Mapping[str, Any] | None, changes: Any, *, table: str,
     - 码值是补充：在现有（没被驳回的）码值上加或改给出的码，不整份替换——「status=9 表示作废」不该把已确认的
       其余码值删掉。含义为空的码（数据剖析写进来的「含义待填写」）可以被补上含义；提案里含义为空的码只在原来
       没有这个码时加进去，不会把已有的含义抹掉（_merge_codes）。
+    - 关联关系不抹掉测量结果：覆盖率沿用目录里的，基数没写时沿用或推算（_relation_after）。两端、基数、覆盖率都和
+      目录里一样时只是确认，来源不变。
     - 看起来像计算公式的文字不收（_FORMULA）：公式进口径卡。
     - 同一个路径出现两次只认第一次。
     """
@@ -904,8 +945,10 @@ def plan_patch(notes: Mapping[str, Any] | None, changes: Any, *, table: str,
         prev = slots.get(key)
         live = prev if prev is not None and prev.get("status") != "rejected" else None
         before = _value_of(key, live) if live is not None else None
-        after = value
-        if key[0] == "c" and key[2] == "codes" and isinstance(before, dict):
+        after, note = value, None
+        if key[0] == "r":
+            after, note = _relation_after(table, value, prev, tables)
+        elif key[0] == "c" and key[2] == "codes" and isinstance(before, dict):
             after = _merge_codes(before, value)
         if before is not None and before == after:
             state = "same" if live.get("status") == "confirmed" else "confirm"
@@ -913,7 +956,7 @@ def plan_patch(notes: Mapping[str, Any] | None, changes: Any, *, table: str,
             state = "change"
         out.append(PatchChange(path=norm, before=copy.deepcopy(before),
                                before_status=live.get("status") if live is not None else None,
-                               after=after, value=value, reason=reason, state=state))
+                               after=after, value=value, reason=reason, state=state, note=note))
     return PatchPlan(out, problems)
 
 
@@ -938,7 +981,10 @@ def apply_patch(notes: Mapping[str, Any] | None, changes: Any, *, table: str,
             slots[key] = {**prev, "status": "confirmed", "updated_at": at}
             continue
         if key[0] == "r":
-            slots[key] = {"id": key[1], **change.after, "source": "human", "status": "confirmed", "updated_at": at}
+            rel = {"id": key[1], **change.after, "source": "human", "status": "confirmed", "updated_at": at}
+            if prev is not None and prev.get("note"):
+                rel["note"] = prev["note"]       # 覆盖率沿用了，说明它怎么测的那段备注也留着
+            slots[key] = rel
         else:
             slots[key] = {"value": change.after, "source": "human", "status": "confirmed", "updated_at": at}
     out = _assemble(slots)
@@ -1006,11 +1052,16 @@ def _comment_is_label(text: str) -> bool:
     return len(text) <= _LABEL_MAX and not _SENTENCE_MARKS.search(text)
 
 
-def _cardinality(meta: Mapping[str, Any], columns: list[str]) -> str:
-    """本表这几列指向别的表时的基数：这几列恰好是本表的主键或某个唯一约束，就是一对一，否则多对一。"""
-    wanted = {c.lower() for c in columns}
+def _is_key(meta: Mapping[str, Any], columns: list[str]) -> bool:
+    """这几列恰好是这张表的主键或某个唯一约束（不分大小写、不计顺序）。"""
+    wanted = {str(c).lower() for c in columns}
     keys = [meta.get("primary_key") or []] + [u for u in meta.get("unique") or [] if isinstance(u, list)]
-    return "one_to_one" if any(wanted == {c.lower() for c in k} for k in keys if k) else "many_to_one"
+    return bool(wanted) and any(wanted == {str(c).lower() for c in k} for k in keys if k)
+
+
+def _cardinality(meta: Mapping[str, Any], columns: list[str]) -> str:
+    """本表这几列指向别的表（的键）时的基数：这几列恰好是本表的主键或某个唯一约束，就是一对一，否则多对一。"""
+    return "one_to_one" if _is_key(meta, columns) else "many_to_one"
 
 
 def _relation(table: str, columns: list[str], to_table: str, to_columns: list[str], source: str, *,

backend/tests/test_catalog_semantics.py

@@ -0,0 +1,151 @@
+"""数据目录核心语义的回归（2026-10 完整性审查 F1）。
+
+- A1：经对话提案确认一条关系，不抹掉数据剖析测得的基数和覆盖率；新增关系没写基数时按目标表的键推算，
+  推算不了就留空并在提案的这一项上说明。否则扇出检查（fanout_sum）看不到基数，静默失效。
+- A2：在目录页去掉推断或外键来的关系即驳回，保留编号；只有人工新建的关系才真正删除。检查器、关系图
+  （助手挑表也用它）、重新起草三处都不能再出现被驳回的关系。
+- A4：码值只补了一部分时不当作完整清单。「码值不在码值表中」只对标了「已列全」的码值报；「已列全」只有人工
+  明确确认、或者剖析没截断地看完全表取值时才有。
+- B2：隔天重新剖析、数据没变，目录不升版本（剖析写的备注里只有日期变了不算变化）。
+- B3：计算公式在人工整份提交、模型起草两条路径上也拦下（以前只拦提案）。
+- 一致性：日期列的判定、表名的匹配、拆词各只有一份。
+
+数据和名称全部虚构（tests/fixtures/catalog/scenic.py）。
+"""
+from __future__ import annotations
+
+import sqlite3
+
+import pytest
+
+from app.data import catalog, catalog_profile, sqlcheck
+from app.db.base import SessionLocal
+from app.db.models import DataSource
+from tests.fixtures.catalog import scenic, scenic_notes
+from tests.fixtures.catalog.fake_model import FakeModel
+from tests.test_catalog_patch import _schema
+from tests.test_catalog_profile import _notes, _on, _relation, _table, _url, client, make_source  # noqa: F401
+
+AT = "2026-10-04T08:00:00+00:00"
+DAY1, DAY2, DAY3 = "2026-10-01T08:00:00+00:00", "2026-10-02T08:00:00+00:00", "2026-10-03T08:00:00+00:00"
+
+
+async def _checker(sid: str) -> sqlcheck.SqlChecker | None:
+    async with SessionLocal() as session:
+        row = await session.get(DataSource, sid)
+        return (await sqlcheck.load_checkers(session, [row])).get(row.name)
+
+
+async def _found(sid: str, sql: str) -> list[tuple]:
+    checker = await _checker(sid)
+    return [(c.code, c.level, c.table, c.column) for c in (checker.check(sql) if checker else [])]
+
+
+async def _graph(sid: str) -> dict[str, list[catalog.JoinEdge]]:
+    """助手挑表用的关系图：库里的目录原样（含驳回项）加表结构现推。"""
+    async with SessionLocal() as session:
+        row = await session.get(DataSource, sid)
+        entries = await catalog.read_catalog(session, sid)
+    return catalog.relation_graph({t: e.notes for t, e in entries.items()}, schema_cache=row.schema_cache)
+
+
+async def _put(client, sid: str, table: str, notes: dict, version: int):
+    return await client.put(f"/api/datasources/{sid}/catalog/{table}", json={"notes": notes, "if_version": version})
+
+
+def _profile_rel(table: str, columns: list[str], to_table: str, to_columns: list[str], *,
+                 cardinality: str | None = "many_to_one", coverage: float | None = 1.0,
+                 status: str = "verified") -> dict:
+    return {"id": catalog.relation_id(table, columns, to_table, to_columns), "columns": columns,
+            "to_table": to_table, "to_columns": to_columns, "cardinality": cardinality, "coverage": coverage,
+            "source": "profile", "status": status,
+            "note": "数据剖析（2026-10-01）：子表抽样 8 个不同键值，父表对上 8 个，覆盖率 100%；被指向列是主键。"}
+
+
+# ==========================================================================
+# A1：提案不抹掉基数和覆盖率
+# ==========================================================================
+
+
+async def test_patch_confirming_a_profiled_relation_keeps_cardinality_and_coverage(scenic_db):
+    tables = (await _schema(scenic_db))["tables"]
+    gate = _profile_rel("visits", ["gate_id"], "gates", ["id"])
+    notes = {"relations": [gate]}
+    value = {"columns": ["gate_id"], "to_table": "gates", "to_columns": ["id"]}
+    for path in ("relations.new", f"relations.{gate['id']}"):
+        [change] = catalog.plan_patch(notes, [{"path": path, "value": value}], table="visits", tables=tables).changes
+        assert change.after["cardinality"] == "many_to_one" and change.after["coverage"] == 1.0
+        # 两端、基数、覆盖率都和目录里的一样：保存即确认，来源不变
+        assert change.state == "confirm"
+    out = catalog.apply_patch(notes, [{"path": "relations.new", "value": value}], table="visits", tables=tables,
+                              at=AT)
+    [rel] = out["relations"]
+    assert (rel["source"], rel["status"], rel["cardinality"], rel["coverage"]) == (
+        "profile", "confirmed", "many_to_one", 1.0)
+    assert rel["note"] == gate["note"]
+
+
+async def test_patch_may_change_cardinality_but_never_coverage(scenic_db):
+    tables = (await _schema(scenic_db))["tables"]
+    gate = _profile_rel("visits", ["gate_id"], "gates", ["id"], coverage=0.97)
+    value = {"columns": ["gate_id"], "to_table": "gates", "to_columns": ["id"], "cardinality": "one_to_one",
+             "coverage": 0.2}                               # 覆盖率是测量值：模型写了也不认
+    [change] = catalog.plan_patch({"relations": [gate]}, [{"path": "relations.new", "value": value}],
+                                  table="visits", tables=tables).changes
+    assert change.state == "change"
+    assert (change.after["cardinality"], change.after["coverage"]) == ("one_to_one", 0.97)
+    out = catalog.apply_patch({"relations": [gate]}, [{"path": "relations.new", "value": value}], table="visits",
+                              tables=tables, at=AT)
+    [rel] = out["relations"]
+    assert (rel["source"], rel["status"], rel["cardinality"], rel["coverage"]) == (
+        "human", "confirmed", "one_to_one", 0.97)
+    assert rel["note"] == gate["note"]                      # 覆盖率留着，说明它的剖析备注也留着
+
+
+async def test_new_relation_without_cardinality_is_inferred_from_target_keys(scenic_db):
+    tables = (await _schema(scenic_db))["tables"]
+
+    def plan(table: str, value: dict) -> catalog.PatchChange:
+        [change] = catalog.plan_patch({}, [{"path": "relations.new", "value": value}], table=table,
+                                      tables=tables).changes
+        return change
+
+    # 目标是主键、本表这一列不唯一：多对一
+    gate = plan("visits", {"columns": ["gate_id"], "to_table": "gates", "to_columns": ["id"]})
+    assert (gate.after["cardinality"], gate.after["coverage"], gate.note) == ("many_to_one", None, None)
+    assert "note" not in gate.as_dict()
+    # 目标是唯一约束、本表这一列也唯一（票号有唯一索引）：一对一
+    ticket = plan("visits", {"columns": ["ticket_no"], "to_table": "orders", "to_columns": ["order_no"]})
+    assert ticket.after["cardinality"] == "one_to_one"
+    # 目标列不是主键也不是唯一约束：推算不了，基数留空，这一项上说明
+    park = plan("visits", {"columns": ["park_id"], "to_table": "gates", "to_columns": ["park_id"]})
+    assert park.after["cardinality"] is None
+    assert park.note and "基数" in park.note and park.as_dict()["note"] == park.note
+    # 不知道表结构（尚未探查）：不推算，也说明
+    [blind] = catalog.plan_patch({}, [{"path": "relations.new", "value": {
+        "columns": ["gate_id"], "to_table": "gates", "to_columns": ["id"]}}], table="visits", tables=None).changes
+    assert blind.after["cardinality"] is None and blind.note
+
+
+async def test_confirming_a_profiled_relation_by_patch_keeps_the_fanout_check(client, make_source, scenic_db):
+    """审查探针 A1：剖析得出多对一、覆盖率 1.0，经对话提案确认这条关系之后，扇出检查照样报 error。"""
+    sql = "SELECT COUNT(g.id) AS n FROM gates g JOIN visits v ON v.gate_id = g.id"
+    sid = await make_source(scenic_db, options=_on(), draft=["visits", "gates"])
+    r = await client.post(_url(sid), json={"tables": ["visits"]})
+    assert r.status_code == 200, r.text
+    detail = await _notes(client, sid, "visits")
+    rel = _relation(detail["notes"], "gate_id")
+    assert (rel["source"], rel["status"], rel["cardinality"], rel["coverage"]) == (
+        "profile", "verified", "many_to_one", 1.0)
+    assert ("fanout_sum", "error", "gates", "id") in await _found(sid, sql)
+
+    changes = [{"path": "relations.new", "value": {"columns": ["gate_id"], "to_table": "gates", "to_columns": ["id"]},
+                "reason": "用户说 gate_id 对应 gates.id"}]
+    r = await client.post(f"/api/datasources/{sid}/catalog/visits/patch",
+                          json={"changes": changes, "if_version": detail["version"]})
+    assert r.status_code == 200, r.text
+    after = _relation(r.json()["notes"], "gate_id")
+    assert (after["status"], after["cardinality"], after["coverage"]) == ("confirmed", "many_to_one", 1.0)
+    assert ("fanout_sum", "error", "gates", "id") in await _found(sid, sql)
+
+