报告里照抄查询结果中文本格的名字（「冲压5线」「KX-205 线」）时，名字里的数字不再算裸数字：目录里每次查询的快照中、文本格里带数字的值（至少一个数字、至少一个不是数字的字，本身不是数也不是带单位的数），在正文里一字不差地出现、前后不紧挨着数字的那几段，查裸数字时涂白。组装和出口复核用同一条规则；只认快照里真有的值，不做模糊匹配，同一个数单独写出来照样要出处，数值格不受影响

backend/app/engine/evidence.py

@@ -61,7 +61,7 @@ from app.data.provenance_types import DOC_PROVENANCE
 from app.data.tabular import UNSHAPED_NOTE
 from app.engine.expressions import CellError, cell_value, column_kind, locate_cell
 from app.engine.guards import FALLBACK_CONTEXT
-from app.engine.issuance import _tolerance, extract_numbers, number_allowance
+from app.engine.issuance import _CN_COUNT_UNIT, _tolerance, extract_numbers, number_allowance
 from app.engine.labels import option_label
 
 DOC_SCHEMA = "agentlab.report/1"
@@ -2872,7 +2872,8 @@ def compose_doc(
                     unit["segments"] = _link_entities(unit["segments"], catalog, index)
         all_segments = [s for block in out_blocks for unit in block["units"] for s in unit["segments"]]
     # 裸数字单独切成片段：前端要在那几个字底下画「无出处」
-    refs = [tuple(s["span"]) for s in all_segments if s.get("ref")] + _header_masks(out_blocks, catalog)
+    refs = [tuple(s["span"]) for s in all_segments if s.get("ref")] + _header_masks(out_blocks, catalog) \
+        + _name_masks(markdown, catalog, snaps)
     syntax = [tuple(s["span"]) for s in all_segments if s["kind"] == "structural"]
     bare = _bare_numbers(markdown, refs, allow_numbers, syntax)
     for block in out_blocks:
@@ -2953,6 +2954,52 @@ def _header_masks(blocks: list[dict[str, Any]], catalog: dict[str, Any]) -> list
     return out
 
 
+#: 文本格的值本身是数、或者是带单位的数（「120台」「3.5万元」「45%」「约 30 天」「205L」）：不算名字（_name_masks），照抄进
+#: 正文的那个数照样要出处。单位认出具校验数量词那一套（issuance._CN_COUNT_UNIT），再加百分号、万亿千百和 1–3 个拉丁字母
+_UNIT_VALUE = re.compile(
+    r"^(?:约|共|近|超|逾|合计|总计|≈|~)?\s*[+-]?(?:\d[\d,]*(?:\.\d+)?|\.\d+)\s*"
+    r"(?:[%％‰]|[万亿千百]?(?:" + _CN_COUNT_UNIT.pattern + r"|[A-Za-z]{1,3}))?\s*$")
+
+
+def _text_names(catalog: dict[str, Any], snaps: _Snapshots) -> set[str]:
+    """目录里每次查询的快照中，文本格里带数字的值（「冲压5线」「KX-205 线」）：至少有一个数字、至少有一个不是数字的字，
+    本身不是数、也不是带单位的数（_UNIT_VALUE）。按写作目录里显示的样子（压成一行）收。取不回的快照跳过。"""
+    names: set[str] = set()
+    for alias, entry in catalog.items():
+        if not isinstance(entry, dict) or entry.get("kind") != "query":
+            continue
+        snapshot, why = snaps.get(alias, entry.get("artifact"))
+        if why or not isinstance(snapshot, dict):
+            continue
+        columns = [str(c) for c in snapshot.get("columns") or []]
+        kinds = [column_kind(snapshot, c) for c in columns]
+        for row in snapshot.get("rows") or []:
+            cells = row if isinstance(row, list) else [row.get(c) for c in columns] if isinstance(row, dict) else []
+            for raw, kind in zip(cells, kinds):
+                if not isinstance(raw, str) or cell_value(raw, kind) is not raw:
+                    continue
+                text = _clean(raw)
+                if any(c.isdigit() for c in text) and not text.isdigit() and not _UNIT_VALUE.match(text):
+                    names.add(text)
+    return names
+
+
+def _name_masks(markdown: str, catalog: dict[str, Any], snaps: _Snapshots) -> list[tuple[int, int]]:
+    """正文里一字不差地写出了查询结果中文本格的值（_text_names）的那几段：名字里的数字不是结论数字，写作者照着查询结果
+    写名字，不该因为名字里有个 5 被打回重写。只认快照里真有的值、整段一字不差，前后紧挨着数字的不算（「冲压15线」不是
+    「冲压5线」）；同一个数单独写出来照样要出处。组装和复核用同一份目录，放过的范围一致。"""
+    out: list[tuple[int, int]] = []
+    for name in _text_names(catalog, snaps) if any(c.isdigit() for c in markdown) else ():
+        at = markdown.find(name)
+        while at >= 0:
+            end = at + len(name)
+            before, after = markdown[at - 1:at], markdown[end:end + 1]
+            if not (before.isdigit() or before == "." or after.isdigit()):
+                out.append((at, end))
+            at = markdown.find(name, at + 1)
+    return out
+
+
 def _split_bare(segments: list[dict[str, Any]], bare: list[Any]) -> list[dict[str, Any]]:
     out: list[dict[str, Any]] = []
     for seg in segments:
@@ -3219,7 +3266,7 @@ def verify_doc(
     if tiled and pos != len(markdown):
         violations.append(_violation("segment_mismatch", "正文末尾有一段不属于任何片段，文档可能被修改过"))
 
-    masks += _header_masks(list(doc.get("blocks") or []), catalog)
+    masks += _header_masks(list(doc.get("blocks") or []), catalog) + _name_masks(markdown, catalog, snaps)
     for token in _bare_numbers(markdown, masks, allow_numbers, syntax):
         owner = next(((sid, uid) for s, e, sid, uid in owners if s <= token.start < e), (None, None))
         violations.append(_violation(

backend/tests/test_report_text_cell_numbers.py

@@ -0,0 +1,73 @@
+"""查询结果里文本格的名字（「冲压5线」「KX-205 线」）照抄进报告时，名字里的数字不算裸数字。
+
+部署上撞过：查询结果的产线名里带数字，写作者照着写「冲压5线的产量 [[v:…]]」，名字里的那个 5 被当成没有出处的数字打回
+重写，重写还是同一个名字。现在：写作目录的查询结果里、文本格的值（一字不差的子串；值里至少有一个不是数字的字，本身
+不是数、也不是带单位的数）出现在正文里时，那一段不查裸数字。只认快照里真有的值，不做模糊匹配：同一个数单独写出来照样
+要出处；数值格不受影响。数据是虚构的。
+"""
+from __future__ import annotations
+
+import pytest
+
+from app.core import artifact_store
+from app.engine.evidence import build_catalog, compose_doc, verify_doc
+
+SNAP = {
+    "columns": ["产线", "产量", "备注"],
+    "column_types": {"产线": "text", "产量": "number", "备注": "text"},
+    "rows": [
+        ["冲压5线", 1520, "120台"],          # 带单位的数：不算名字
+        ["KX-205 线", 830, "45%"],
+        ["焊装 2 线", 64, "3.5万元"],
+    ],
+    "row_count": 3, "truncated": False, "elapsed_ms": 2,
+    "sql": "SELECT line AS 产线, SUM(qty) AS 产量, MAX(memo) AS 备注 FROM output GROUP BY line",
+    "source": "plant",
+}
+
+
+def _store(snap: dict) -> str:
+    text = artifact_store.canonical_json(snap)
+    artifact = artifact_store.content_hash(text)
+    path = artifact_store._path_of(artifact)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(text, encoding="utf-8")
+    return artifact
+
+
+@pytest.fixture
+def catalog():
+    art = _store(SNAP)
+    ledger = [{"kind": "query", "node_id": "fetch", "exec": 1, "artifact": art, "via": "b" * 64, "call_id": "call_1",
+               "tool": "db_query__plant", "source": "plant", "columns": SNAP["columns"], "rows": 3, "truncated": False}]
+    return build_catalog(nodes={}, ledger=ledger, inputs={})
+
+
+def _uncited(doc) -> list[str]:
+    return [v["text"] for v in doc["violations"] if v["code"] == "uncited_number"]
+
+
+def test_a_number_inside_a_name_from_the_results_is_allowed(catalog):
+    text = "冲压5线的产量是 [[v:Q1.r0.产量]]，KX-205 线是 [[v:Q1.r1.产量]]，焊装 2 线是 [[v:Q1.r2.产量]]。[[see:Q1]]"
+    doc = compose_doc(text, catalog)
+    assert _uncited(doc) == [], doc["violations"]
+    assert verify_doc(doc, catalog)["ok"], "出口复核按同一条规则：组装时放过的，复核也放过"
+
+
+def test_the_same_number_on_its_own_still_needs_a_citation(catalog):
+    doc = compose_doc("冲压5线有 5 个班组，产量 [[v:Q1.r0.产量]]。[[see:Q1]]", catalog)
+    assert _uncited(doc) == ["5"]
+    # 名字只认一字不差的：快照里没有「冲压15线」，15 照样要出处；名字前后紧挨着数字的也不算那个名字
+    assert _uncited(compose_doc("冲压15线的产量 [[v:Q1.r0.产量]]。", catalog)) == ["15"]
+    assert _uncited(compose_doc("KX-2050 线的产量 [[v:Q1.r1.产量]]。", catalog)) == ["2050"]
+    assert _uncited(compose_doc("焊装 2 线以外还有 2 线。", catalog)) == ["2"]
+
+
+def test_numbers_and_numbers_with_units_in_text_cells_are_not_names(catalog):
+    doc = compose_doc("备注里写着 120台、45%、3.5万元。", catalog)
+    assert _uncited(doc) == ["120", "45%", "3.5万"], doc["violations"]
+
+
+def test_numeric_cells_are_unaffected(catalog):
+    # 产量格的 1520 是数值格：照抄成裸数字照样违规
+    assert _uncited(compose_doc("冲压5线的产量是 1520。", catalog)) == ["1520"]