diff --git a/docs/fix-record-2026-09-26.md b/docs/fix-record-2026-09-26.md
new file mode 100644
index 0000000..3a59e13
--- /dev/null
+++ b/docs/fix-record-2026-09-26.md
@@ -0,0 +1,30 @@
+# 剩余问题修复记录 · 2026-09-26
+
+对应 `docs/remaining-issues-2026-09-26.html` 第三节（F01–F08）的修复交付记录。
+分支：`fix/remaining-issues-2026-09-26`（基线 `0ac0c4e` 起逐项小步提交）。
+
+## 修复状态
+
+| 编号 | 状态 | 提交 | 说明 |
+| --- | --- | --- | --- |
+| F01 · P0 · 符号链接清理 | 已修复 | `36fd5f3` | 根目录/用户目录/目标文件先 lstat 完整校验（拒绝符号链接与非目录/非普通文件），校验失败不删除任何磁盘文件；失败清理只处理本次登记的受控文件并通过 `O_NOFOLLOW` 目录描述符 unlink，目录被换成符号链接时拒绝删除。测试覆盖用户目录链接、目标链接、运行中目录被换成链接（受保护文件字节不变）、失败后重试成功。 |
+| F02 · P1 · 清理失败与孤儿 | 已修复 | `54dc58d` | 清理受阻时记录不含敏感数据的 WARNING（任务编号+文件名）并保留 `job.path` 使残留可追踪；`cleanup_formula_exports` 新增孤儿 `.xlsx` 回收：超过 24h 保留期且不属于未过期成功任务的文件统一回收，不触碰活动任务。测试含 unlink 权限错误注入→告警→恢复权限后回收且其他任务文件不受影响。 |
+| F03 · P1 · 证据弹窗卡加载 | 已修复 | `6cf236a` | 同期间同页点击直接返回；第二页点击当前期间回第一页刷新；失败显示错误+重试按钮；空结果显示空态行。E2E 覆盖重复点击、快速切换、第二页切期间、失败重试、空列表。 |
+| F04 · P1 · 载入定义丢上下文 | 已修复 | `9a5ab26` | 载入定义时以“用户消息+助手确认”把名称/含义/范围/缺失处理/表达式写入可见对话上下文（需勾选同意才发送），修改后仍需重新确认；选择空选项同步清理定义与上下文。服务端单条消息上限 1000→2000 容纳完整定义。E2E 验证模型请求包含原定义且“乘以 100”修改生效。 |
+| F05 · P1 · 定义库并发与审计 | 已修复 | `0aaa87e` | 迁移 0005 增加 `revision`；保存在单事务内锁定用户所属行：配额（50/数据集）原子化、同名更新必须携带匹配 revision（409 冲突）、审计与变更同事务回滚；保存限流 30/h。前端保存同名定义携带 revision 并明示覆盖行为。PostgreSQL 并发配额用例作为独立门禁（本机 SQLite 自动跳过）。 |
+| F06 · P1 · 工作簿可解释性 | 已修复 | `9962573` | 参数页“产品毛利率”改为只读快照（可编辑列全部为“否”），写明调整位置是计算明细 MarginRate 列；参数页新增维度键-显示名完整对照；使用说明新增输入字典（Revenue/Cost/GrossInput/MarginRate/ListAmount 的单位、来源字段、修改位置、缺失含义）与固定结构边界行。 |
+| F07 · P1 · 对账容差 | 已修复 | `649fec4` | 对账检查页新增容差列：人数严格相等、金额 0.01、比率 0.0001，状态公式按 `ABS(差异)<=容差` 判定；自定义列用 `to_decimal` 白名单表达式解释器（不执行代码）计算 Decimal 服务器参考，不再标注“需业务验收”；参数页写明容差规则与“调整需财务负责人批准”。 |
+| F08 · P1 · 撤权与生命周期 | 已修复 | `5ebe83a` | worker 执行前重新核验账号 `is_active` 与 admin/analyst 角色（撤权/停用后排队任务失败、不发布）；导出状态/下载接口补齐角色白名单（降为 viewer 后 403）；`import_worker` 对超过 15 分钟的 processing 任务重新认领（PG 依赖咨询锁立即恢复），崩溃重启不再无限停留。 |
+
+## 本轮完整测试记录（2026-09-26，SQLite 隔离环境）
+
+- **后端：87 passed，7 skipped，90.3 秒；finance 覆盖率 88.50%（85% 门槛通过）。**
+  跳过项：4 项 LibreOffice 重算（本机未安装 soffice，环境门禁，CI 中执行）、1 项 PostgreSQL 并发配额（依赖 PG 行锁，生产库单独验证）、1 项原附件全量 UI（未设置 `REFERENCE_DATASET_NAME`）。
+- **浏览器 E2E：3 passed，1 skipped，21.0 秒。** 新隔离数据库（`var/e2e-test.sqlite3` + 独立 MEDIA_ROOT），覆盖追问、保存定义、关闭重开、载入定义修改后生成（乘以 100）、下载，含 F03 全部交互回归（重复点击/快速切换/失败重试/空列表/翻页）、桌面与手机。
+- Ruff check / ruff format、TypeScript（tsc --noEmit）、Vite 构建、Django check、`makemigrations --check`（迁移一致）、collectstatic：全部通过。
+- `pip-audit`：无已知漏洞。`npm audit --omit=dev`：本机 registry 访问失败（网络受限，CI 中执行），记录为环境限制项。
+
+## 仍开放的事项（与报告一致，非本轮代码范围）
+
+- 第四节功能缺口（网页主汇总表新列、草稿续聊、原附件 332,830 行导出、高基数与可访问性专项等）与第五节外部门禁（目标 Excel 实机、财务口径签字、AI 真实联调、生产基础设施、PostgreSQL/SLA、安全运营）仍按原报告执行，责任人与关闭证据不变。
+- F07 容差数值（金额 0.01 / 比率 0.0001）为工程默认，仍需财务负责人确认或调整后才视为财务口径关闭。
diff --git a/finance/exporting.py b/finance/exporting.py
index 28e5e18..e24fcab 100644
--- a/finance/exporting.py
+++ b/finance/exporting.py
@@ -1,7 +1,9 @@
 import hashlib
 import hmac
+import logging
 import os
 import secrets
+import stat
 import warnings
 from datetime import timedelta
 from decimal import Decimal, InvalidOperation
@@ -19,6 +21,7 @@ from .formula_columns import (
     RESERVED_DATA_HEADERS,
     ensure_available_metrics,
     metrics_used,
+    to_decimal,
     to_excel,
     validate_column,
 )
@@ -28,6 +31,75 @@ FORMULA_VERSION = "3"
 MAX_ROWS = 100_000
 MAX_GROUPS = 200
 
+# 对账比较规则：人数严格相等；金额 0.01；比率 0.0001（一个基点）。
+# 容差属财务口径，调整需财务负责人批准后方可修改。
+RECONCILE_TOLERANCES = {
+    "customers": Decimal("0"),
+    "revenue": Decimal("0.01"),
+    "cost": Decimal("0.01"),
+    "gross": Decimal("0.01"),
+    "self_revenue": Decimal("0.01"),
+    "basket": Decimal("0.0001"),
+    "margin": Decimal("0.0001"),
+    "discount_rate": Decimal("0.0001"),
+    "upgrade_rate": Decimal("0.0001"),
+    "list_basket": Decimal("0.0001"),
+    "roi": Decimal("0.0001"),
+}
+CUSTOM_COLUMN_TOLERANCE = Decimal("0.01")
+
+logger = logging.getLogger(__name__)
+
+
+def _validated_export_directory(path, description):
+    """导出目录必须不存在或为真实目录；符号链接一律拒绝并保持文件原样。"""
+    try:
+        info = os.lstat(path)
+    except FileNotFoundError:
+        path.mkdir(parents=True, mode=0o700)
+        info = os.lstat(path)
+    if stat.S_ISLNK(info.st_mode) or not stat.S_ISDIR(info.st_mode):
+        raise ValueError(f"{description}必须是真实目录，不能是符号链接")
+    os.chmod(path, 0o700)
+    return path
+
+
+def _validated_export_target(user_dir, job):
+    """目标文件若已存在必须是普通文件；符号链接或其他类型一律拒绝覆盖。"""
+    target = user_dir / f"{job.pk}.xlsx"
+    try:
+        info = os.lstat(target)
+    except FileNotFoundError:
+        return target
+    if stat.S_ISLNK(info.st_mode) or not stat.S_ISREG(info.st_mode):
+        raise ValueError("公式导出目标路径已存在且不是普通文件，已拒绝生成")
+    return target
+
+
+def _unlink_created_file(path):
+    """只通过不跟随链接的目录句柄删除受控目录内的普通文件。
+
+    返回 "removed"（已删除）、"missing"（本就不存在）或 "blocked"（无法安全删除，
+    例如目录已被换成符号链接或权限不足），调用方据此保留失败任务与磁盘的一致性。
+    """
+    try:
+        dir_fd = os.open(path.parent, os.O_RDONLY | os.O_NOFOLLOW)
+    except OSError:
+        return "blocked"
+    try:
+        try:
+            info = os.lstat(path.name, dir_fd=dir_fd)
+        except FileNotFoundError:
+            return "missing"
+        if not stat.S_ISREG(info.st_mode):
+            return "blocked"
+        os.unlink(path.name, dir_fd=dir_fd)
+        return "removed"
+    except OSError:
+        return "blocked"
+    finally:
+        os.close(dir_fd)
+
 
 def _excel_number(value):
     if value is None:
@@ -408,7 +480,12 @@ def build_formula_workbook(ds, member, export_context, output):
         ],
         ["ROI 口径", mapping["roi_formula"], "按导入映射保存的公式分支", "否"],
         ["自营产品集合", _text_cell(parameters, ", ".join(products)), "升单率和自营收入口径", "否"],
-        ["产品毛利率", str(rates), "直接毛利缺失时的可编辑计算输入", "是"],
+        [
+            "产品毛利率（快照）",
+            _text_cell(parameters, str(rates)),
+            "只读快照；导出时已按行写入计算明细 MarginRate 列，调整毛利率请在计算明细页按行修改",
+            "否",
+        ],
         [
             "Excel 数值精度",
             "IEEE-754 双精度",
@@ -417,8 +494,23 @@ def build_formula_workbook(ds, member, export_context, output):
         ],
         ["重算范围", f"最多 {MAX_ROWS:,} 行、{MAX_GROUPS} 组", "超限需缩小导出范围", "否"],
         ["兼容性", "SUMIFS/COUNTIFS", "移除动态数组人数公式；仍需目标 Excel 桌面版本验收", "否"],
+        [
+            "对账容差",
+            "人数严格相等；金额 0.01；比率 0.0001",
+            "对账检查页按指标类型使用独立容差；自定义列按金额容差 0.01。容差调整需财务负责人批准",
+            "否",
+        ],
     ]:
         parameters.append(row)
+    for d in dims:
+        parameters.append(
+            [
+                _text_cell(parameters, f"维度 {d['key']}"),
+                _text_cell(parameters, d["name"]),
+                "维度键与显示名对照（只读）；计算明细与原始证据的列名即维度键",
+                "否",
+            ]
+        )
     for definition in custom_columns:
         parameters.append(
             [
@@ -429,10 +521,33 @@ def build_formula_workbook(ds, member, export_context, output):
             ]
         )
     instructions.append(_styled_row(instructions, ["主题", "说明"], green, white))
+
+    def source_name(role):
+        return next((c["name"] for c in mapping["columns"] if c["role"] == role), "未映射")
+
+    editable_inputs = [
+        ("Revenue", "实际收入", "金额", source_name("revenue"), "空值行不参与求和，汇总按现有行计算"),
+        ("Cost", "成本", "金额", source_name("cost"), "任一行缺失时本组成本及 ROI 等依赖指标不可用"),
+        ("GrossInput", "直接毛利", "金额", source_name("gross"), "空值行改按产品毛利率（lookup）计算毛利"),
+        (
+            "MarginRate",
+            "产品毛利率",
+            "小数比率",
+            "导入确认的产品毛利率表",
+            "仅 GrossSource=lookup 的行参与毛利计算；0.2 表示 20%",
+        ),
+        (
+            "ListAmount",
+            "原价金额",
+            "金额",
+            source_name("list_amount"),
+            "用于扣率与原价客单价；缺失则相关指标不可用",
+        ),
+    ]
     for row in [
         ["用途", "公式分析工作簿；本地修改不回写系统。"],
         ["行粒度", "计算明细每行对应一个 Fact，不等同患者就诊数。"],
-        ["维度列", "列名为数据导入映射中的维度键；显示名称见本页分组维度。"],
+        ["维度列", "列名为数据导入映射中的维度键；键与显示名对照见“参数与口径”页的维度行。"],
         ["毛利", "直接毛利优先，否则收入×产品毛利率；无法计算则不可用。毛利贡献不代表净利润。"],
         ["ROI", "按选定口径；缺成本、成本非正或所需指标缺失时不可用。"],
         ["人数", "客户 token 跨行去重；缺客户编码时不可用。"],
@@ -443,6 +558,25 @@ def build_formula_workbook(ds, member, export_context, output):
         ["变化率", "差额/ABS(基期)；基期为0或期间缺失时不可用。"],
         ["精度", "原始 Decimal 以文本留存；Excel 双精度存在细微误差可能。"],
         ["隐私", "客户标识替换为本文件随机 token，文件仍按敏感财务信息保护。"],
+        [
+            "输入字典",
+            "以下输入可在“计算明细”页直接修改，全部汇总公式与对账实时引用；字段、单位、来源与缺失含义逐项列出。",
+        ],
+        *[
+            [
+                _text_cell(instructions, f"输入 {header}（{label}）"),
+                _text_cell(
+                    instructions,
+                    f"单位：{unit}；来源：{source}；修改位置：计算明细页 {header} 列；缺失含义：{missing}。",
+                ),
+            ]
+            for header, label, unit, source, missing in editable_inputs
+        ],
+        [
+            "固定结构",
+            "Period、维度列、CustomerToken、GrossSource、SelfOperated、CustomerPresent、RowCount 与“人数去重”页均为固定快照；"
+            "修改或增删这些列/行会使汇总、去重与对账失真，需要改变结构时请重新导出。",
+        ],
     ]:
         instructions.append(row)
     for definition in custom_columns:
@@ -546,11 +680,18 @@ def build_formula_workbook(ds, member, export_context, output):
     reconcile.append(
         _styled_row(
             reconcile,
-            ["指标", "分组", "期间侧", "服务器 Decimal 参考", "公式结果", "差异", "状态"],
+            ["指标", "分组", "期间侧", "服务器 Decimal 参考", "公式结果", "差异", "容差", "状态"],
             green,
             white,
         )
     )
+
+    def _reconcile_status(row_number):
+        return (
+            f'=IF(D{row_number}="","无参考",IF(ISNUMBER(F{row_number}),'
+            f'IF(ABS(F{row_number})<=G{row_number},"通过","待复核"),"待复核"))'
+        )
+
     reconcile_row = 2
     for label in labels:
         for side_name, periods, col in (("本期", ctx["current"], "C"), ("基期", ctx["baseline"], "D")):
@@ -567,21 +708,25 @@ def build_formula_workbook(ds, member, export_context, output):
                         "" if value is None else str(value),
                         f"='分析汇总'!{col}{formula_row}",
                         f'=IFERROR(E{reconcile_row}-D{reconcile_row},"")',
-                        f'=IF(D{reconcile_row}="","无参考",IF(ISNUMBER(F{reconcile_row}),IF(ABS(F{reconcile_row})<0.01,"通过","待复核"),"待复核"))',
+                        float(RECONCILE_TOLERANCES[key]),
+                        _reconcile_status(reconcile_row),
                     ]
                 )
                 reconcile_row += 1
             for index, definition in enumerate(custom_columns):
                 formula_row = metric_rows[(label, f"custom_{index}")]
+                # 自定义列用 Decimal 白名单解释器计算服务器参考，不再依赖业务人工验收
+                reference_value = to_decimal(definition["expression"], reference)
                 reconcile.append(
                     [
                         _text_cell(reconcile, definition["name"]),
                         _text_cell(reconcile, label),
                         side_name,
-                        "自定义公式无服务器参考",
+                        "" if reference_value is None else str(reference_value),
                         f"='分析汇总'!{col}{formula_row}",
-                        "",
-                        "需业务验收",
+                        f'=IFERROR(E{reconcile_row}-D{reconcile_row},"")',
+                        float(CUSTOM_COLUMN_TOLERANCE),
+                        _reconcile_status(reconcile_row),
                     ]
                 )
                 reconcile_row += 1
@@ -594,41 +739,51 @@ def build_formula_workbook(ds, member, export_context, output):
 
 def process_formula_export(export_id):
     job = FormulaExport.objects.select_related("dataset", "user").get(pk=export_id)
-    temporary = None
     target = None
+    created = []
     try:
-        root = Path(settings.MEDIA_ROOT) / "formula-exports"
-        if root.is_symlink():
-            raise ValueError("公式导出目录不能是符号链接")
-        root.mkdir(parents=True, exist_ok=True, mode=0o700)
-        root.chmod(0o700)
-        target = root / str(job.user_id) / f"{job.pk}.xlsx"
-        target.parent.mkdir(parents=True, exist_ok=True, mode=0o700)
-        if target.parent.is_symlink() or target.is_symlink():
-            raise ValueError("公式导出路径不能是符号链接")
-        target.parent.chmod(0o700)
-        temporary = target.with_name(f"{job.pk}.{secrets.token_hex(8)}.tmp")
         member = Membership.objects.get(user=job.user, organization=job.dataset.organization)
+        # 执行阶段重新核验角色与账号状态：撤权或停用后排队中的任务不得继续发布敏感工作簿。
+        if not job.user.is_active:
+            raise PermissionError("账号已停用，公式导出已拒绝执行")
+        if member.role not in ("admin", "analyst"):
+            raise PermissionError("账号角色已变更，公式导出已拒绝执行")
         if set(member.departments) != set(job.departments):
             raise PermissionError("科室授权已变更，请重新提交导出")
+        root = _validated_export_directory(Path(settings.MEDIA_ROOT) / "formula-exports", "公式导出目录")
+        user_dir = _validated_export_directory(root / str(job.user_id), "公式导出用户目录")
+        target = _validated_export_target(user_dir, job)
+        # 路径完整验证通过后才把目标登记为可清理文件；此前任何失败都不会删除磁盘文件。
+        created.append(target)
+        temporary = target.with_name(f"{job.pk}.{secrets.token_hex(8)}.tmp")
         descriptor = os.open(temporary, os.O_WRONLY | os.O_CREAT | os.O_EXCL, 0o600)
+        created.append(temporary)
         with os.fdopen(descriptor, "wb") as file:
             stats = build_formula_workbook(job.dataset, member, job.context, file)
-        temporary.replace(target)
+        os.replace(temporary, target)
+        created = [target]
         job.path, job.status, job.error = str(target), "succeeded", ""
         job.finished_at = timezone.now()
         job.expires_at = timezone.now() + timedelta(hours=24)
         job.save(update_fields=["path", "status", "error", "finished_at", "expires_at"])
         return stats
     except Exception as exc:
-        for path in (temporary, target):
-            try:
-                if path is not None and not path.is_symlink():
-                    path.unlink(missing_ok=True)
-            except OSError:
-                pass
-        job.status, job.error, job.path = "failed", str(exc)[:500], ""
-        job.finished_at, job.expires_at = timezone.now(), None
+        job.status, job.error, job.finished_at, job.expires_at = (
+            "failed",
+            str(exc)[:500],
+            timezone.now(),
+            None,
+        )
+        # 只清理本次登记且确认位于受控目录内的文件；目录被换成符号链接时拒绝删除。
+        leftover = [path for path in created if _unlink_created_file(path) == "blocked"]
+        if leftover:
+            # 日志只含任务编号与文件名，不包含敏感数据；残留文件由清理命令按保留期回收。
+            logger.warning(
+                "公式导出任务 %s 失败后清理文件受阻，待清理命令回收：%s",
+                job.pk,
+                "、".join(path.name for path in leftover),
+            )
+        job.path = str(target) if target in leftover else ""
         job.save(update_fields=["status", "error", "path", "finished_at", "expires_at"])
         raise
 
diff --git a/finance/formula_columns.py b/finance/formula_columns.py
index 9981ceb..72e22b2 100644
--- a/finance/formula_columns.py
+++ b/finance/formula_columns.py
@@ -1,4 +1,5 @@
 import math
+from decimal import Decimal
 
 from .analysis import METRICS
 
@@ -107,3 +108,27 @@ def to_excel(node, references):
     if "number" in node:
         return str(node["number"])
     return f"({to_excel(node['left'], references)}{OPERATORS[node['operator']]}{to_excel(node['right'], references)})"
+
+
+def to_decimal(node, values):
+    """按已校验的白名单表达式树用 Decimal 计算服务器参考值。
+
+    只遍历 metric/number/四则运算节点，不执行任何代码；依赖指标缺失或除零时
+    返回 None，与 Excel 侧 #N/A 的语义一致。
+    """
+    if "metric" in node:
+        return values[node["metric"]]
+    if "number" in node:
+        return Decimal(str(node["number"]))
+    left, right = to_decimal(node["left"], values), to_decimal(node["right"], values)
+    if left is None or right is None:
+        return None
+    if node["operator"] == "add":
+        return left + right
+    if node["operator"] == "subtract":
+        return left - right
+    if node["operator"] == "multiply":
+        return left * right
+    if right == 0:
+        return None
+    return left / right
diff --git a/finance/formula_views.py b/finance/formula_views.py
index 898b741..402a09b 100644
--- a/finance/formula_views.py
+++ b/finance/formula_views.py
@@ -3,14 +3,17 @@ from pathlib import Path
 
 import httpx
 from django.conf import settings
+from django.db import IntegrityError, transaction
 from django.http import FileResponse, JsonResponse
 from django.utils import timezone
 
 from .analysis import METRICS, context, dimensions
 from .formula_columns import available_metrics, ensure_available_metrics, validate_column
-from .models import FormulaDefinition, FormulaExport
+from .models import FormulaDefinition, FormulaExport, Membership
 from .views import APIError, audit, body, dataset, endpoint, limit
 
+MAX_DEFINITIONS_PER_DATASET = 50
+
 
 def design_column(messages, ds, ctx):
     if not (settings.AI_ENABLED and settings.OPENAI_API_KEY and settings.OPENAI_MODEL):
@@ -22,7 +25,8 @@ def design_column(messages, ds, ctx):
         or set(item) != {"role", "content"}
         or item["role"] not in ("user", "assistant")
         or not isinstance(item["content"], str)
-        or len(item["content"]) > 1000
+        # 2000 字符容纳载入已保存定义后的完整对话上下文
+        or len(item["content"]) > 2000
         for item in messages
     ):
         raise APIError("对话消息内容无效")
@@ -197,6 +201,7 @@ def formula_definitions(request, dataset_id):
                     {
                         "id": str(item.pk),
                         "definition": item.definition,
+                        "revision": item.revision,
                         "updated_at": item.updated_at.isoformat(),
                     }
                     for item in items
@@ -204,28 +209,54 @@ def formula_definitions(request, dataset_id):
             }
         )
 
-    definition = validate_column(body(request).get("definition"))
+    data = body(request)
+    definition = validate_column(data.get("definition"))
     try:
         ensure_available_metrics(ds, definition)
     except ValueError as exc:
         raise APIError(str(exc)) from exc
     normalized_name = definition["name"].strip().casefold()
-    existing = FormulaDefinition.objects.filter(
-        dataset=ds, user=request.user, normalized_name=normalized_name
-    ).first()
-    if not existing and FormulaDefinition.objects.filter(dataset=ds, user=request.user).count() >= 50:
-        raise APIError("每个数据集最多保存 50 个公式定义")
-    item, created = FormulaDefinition.objects.update_or_create(
-        dataset=ds,
-        user=request.user,
-        normalized_name=normalized_name,
-        defaults={"definition": definition},
-    )
-    audit(request, "formula_definition.save", item.pk, {"dataset": str(ds.pk), "name": definition["name"]})
+    limit(f"formula-definition:{request.user.pk}", 30, 3600)
+    with transaction.atomic():
+        # 锁定稳定的用户所属行序列化同一用户的并发保存，使配额检查与写入原子化。
+        Membership.objects.select_for_update().get(pk=request.member.pk)
+        existing = (
+            FormulaDefinition.objects.select_for_update()
+            .filter(dataset=ds, user=request.user, normalized_name=normalized_name)
+            .first()
+        )
+        if existing:
+            # 同名更新按 revision 做乐观并发控制，避免两个窗口相互静默覆盖。
+            if data.get("revision") != existing.revision:
+                raise APIError("定义已在其他窗口更新，请刷新后重试", 409)
+            existing.definition, existing.revision = definition, existing.revision + 1
+            existing.save(update_fields=["definition", "revision", "updated_at"])
+            item, created = existing, False
+        else:
+            if (
+                FormulaDefinition.objects.filter(dataset=ds, user=request.user).count()
+                >= MAX_DEFINITIONS_PER_DATASET
+            ):
+                raise APIError(f"每个数据集最多保存 {MAX_DEFINITIONS_PER_DATASET} 个公式定义")
+            try:
+                item = FormulaDefinition.objects.create(
+                    dataset=ds,
+                    user=request.user,
+                    normalized_name=normalized_name,
+                    definition=definition,
+                )
+            except IntegrityError as exc:
+                raise APIError("定义保存发生并发冲突，请重试", 409) from exc
+            created = True
+        # 审计与定义变更在同一事务内提交；审计失败时变更一并回滚。
+        audit(
+            request, "formula_definition.save", item.pk, {"dataset": str(ds.pk), "name": definition["name"]}
+        )
     return JsonResponse(
         {
             "id": str(item.pk),
             "definition": item.definition,
+            "revision": item.revision,
             "updated_at": item.updated_at.isoformat(),
         },
         status=201 if created else 200,
@@ -241,8 +272,9 @@ def formula_definition_delete(request, dataset_id, definition_id):
     if not item:
         raise APIError("公式定义不存在", 404)
     name = item.definition.get("name", "")
-    audit(request, "formula_definition.delete", item.pk, {"dataset": str(ds.pk), "name": name})
-    item.delete()
+    with transaction.atomic():
+        item.delete()
+        audit(request, "formula_definition.delete", item.pk, {"dataset": str(ds.pk), "name": name})
     return JsonResponse({"deleted": True})
 
 
@@ -292,7 +324,7 @@ def _owned_job(request, dataset_id, export_id):
     return job
 
 
-@endpoint(("GET",))
+@endpoint(("GET",), roles=("admin", "analyst"))
 def formula_export_status(request, dataset_id, export_id):
     job = _owned_job(request, dataset_id, export_id)
     return JsonResponse(
@@ -305,7 +337,7 @@ def formula_export_status(request, dataset_id, export_id):
     )
 
 
-@endpoint(("GET",))
+@endpoint(("GET",), roles=("admin", "analyst"))
 def formula_export_download(request, dataset_id, export_id):
     job = _owned_job(request, dataset_id, export_id)
     if job.status != "succeeded" or not job.path or not job.expires_at or job.expires_at <= timezone.now():
diff --git a/finance/management/commands/cleanup_formula_exports.py b/finance/management/commands/cleanup_formula_exports.py
index 692e1aa..61cb83f 100644
--- a/finance/management/commands/cleanup_formula_exports.py
+++ b/finance/management/commands/cleanup_formula_exports.py
@@ -1,4 +1,5 @@
 import time
+import uuid
 from datetime import timedelta
 from pathlib import Path
 
@@ -9,9 +10,11 @@ from django.utils import timezone
 
 from finance.models import FormulaExport
 
+ORPHAN_RETENTION_HOURS = 24
+
 
 class Command(BaseCommand):
-    help = "Remove expired formula workbooks and stale temporary files"
+    help = "Remove expired formula workbooks and reclaim orphan files"
 
     def add_arguments(self, parser):
         parser.add_argument("--loop", action="store_true")
@@ -19,8 +22,8 @@ class Command(BaseCommand):
 
     def handle(self, *args, **options):
         while True:
-            jobs, files = self.clean_expired()
-            self.stdout.write(f"Expired {jobs} jobs; removed {files} files")
+            jobs, files, orphans = self.clean_expired()
+            self.stdout.write(f"Expired {jobs} jobs; removed {files} files; reclaimed {orphans} orphan files")
             if not options["loop"]:
                 return
             time.sleep(max(60, options["interval"]))
@@ -31,7 +34,7 @@ class Command(BaseCommand):
             raise ValueError("公式导出目录不能是符号链接")
         root = export_root.resolve()
         now = timezone.now()
-        jobs = files = 0
+        jobs = files = orphans = 0
 
         for job in FormulaExport.objects.filter(status="succeeded", expires_at__lte=now).iterator():
             path = Path(job.path) if job.path else None
@@ -61,5 +64,41 @@ class Command(BaseCommand):
                         files += 1
                     except OSError as exc:
                         self.stderr.write(f"Could not remove stale temporary export {path.name}: {exc}")
+                orphans += self._reclaim_orphan_workbooks(user_dir, now)
 
-        return jobs, files
+        return jobs, files, orphans
+
+    def _reclaim_orphan_workbooks(self, user_dir, now):
+        """回收超过保留期且不再属于任何有效导出任务的 .xlsx 孤儿文件。
+
+        有效任务指状态 succeeded、路径一致且尚未过期的任务；失败任务残留的文件和任务行
+        已被删除（如数据集级联删除）的文件在保留期后回收，绝不触碰活动任务的文件。
+        """
+        retention = (now - timedelta(hours=ORPHAN_RETENTION_HOURS)).timestamp()
+        removed = 0
+        for path in user_dir.glob("*.xlsx"):
+            try:
+                if path.is_symlink() or not path.is_file() or path.stat().st_mtime > retention:
+                    continue
+                job_id = uuid.UUID(path.stem)
+            except (OSError, ValueError):
+                continue
+            job = FormulaExport.objects.filter(pk=job_id).first()
+            if (
+                job is not None
+                and job.status == "succeeded"
+                and job.path == str(path)
+                and job.expires_at is not None
+                and job.expires_at > now
+            ):
+                continue
+            try:
+                path.unlink()
+                removed += 1
+                if job is not None and job.path == str(path):
+                    job.path = ""
+                    job.save(update_fields=["path"])
+                    self.stdout.write(f"Reclaimed leftover workbook of job {job.pk}")
+            except OSError as exc:
+                self.stderr.write(f"Could not reclaim orphan export {path.stem}: {exc}")
+        return removed
diff --git a/finance/management/commands/import_worker.py b/finance/management/commands/import_worker.py
index 469b76b..f11ada7 100644
--- a/finance/management/commands/import_worker.py
+++ b/finance/management/commands/import_worker.py
@@ -1,13 +1,19 @@
 import time
+from datetime import timedelta
 
 from django.core.management.base import BaseCommand
 from django.db import connection, transaction
+from django.db.models import Q
 from django.utils import timezone
 
 from finance.exporting import process_formula_export
 from finance.importing import process_dataset
 from finance.models import Dataset, FormulaExport
 
+# 无咨询锁的数据库（如 SQLite）上等待该时限后才重新认领 processing 任务，
+# 避免与仍在运行的 worker 抢任务；PostgreSQL 由咨询锁保证单 worker，可立即恢复。
+STALE_PROCESSING_SECONDS = 15 * 60
+
 
 class Command(BaseCommand):
     help = "Process durable import and formula-export jobs"
@@ -30,30 +36,35 @@ class Command(BaseCommand):
                 with connection.cursor() as cursor:
                     cursor.execute("SELECT pg_advisory_unlock(782146291)")
 
-    def run_jobs(self, options, recover):
+    def run_jobs(self, options, recover_immediately):
         while True:
-            processed = self._next_import(recover)
+            processed = self._next_import(recover_immediately)
             if not processed:
-                processed = self._next_export(recover)
+                processed = self._next_export(recover_immediately)
             if options["once"]:
                 return
             if not processed:
                 time.sleep(2)
 
-    def _claim(self, model, states):
+    def _claim(self, model, immediate_recovery):
+        stale_before = timezone.now() - timedelta(seconds=STALE_PROCESSING_SECONDS)
+        claimable = Q(status="queued") | Q(status="processing", created_at__lt=stale_before)
+        if immediate_recovery:
+            claimable = Q(status="queued") | Q(status="processing")
         with transaction.atomic():
-            qs = model.objects.filter(status__in=states).order_by("created_at")
+            qs = model.objects.filter(claimable).order_by("created_at")
             if connection.features.has_select_for_update_skip_locked:
                 qs = qs.select_for_update(skip_locked=True)
             item = qs.first()
             if not item:
                 return None
-            updated = model.objects.filter(pk=item.pk, status__in=states).update(status="processing")
-            return item if updated else None
+            # CAS 认领 queued 任务互斥；processing 仅在超时或咨询锁保护下重认领，
+            # 重复发布对同一目标路径写入相同内容，不会产生重复文件。
+            claimed = model.objects.filter(pk=item.pk, status=item.status).update(status="processing")
+            return item if claimed else None
 
-    def _next_import(self, recover):
-        states = ["queued", "processing"] if recover else ["queued"]
-        item = self._claim(Dataset, states)
+    def _next_import(self, recover_immediately):
+        item = self._claim(Dataset, recover_immediately)
         if not item:
             return False
         Dataset.objects.filter(pk=item.pk).update(started_at=timezone.now())
@@ -61,9 +72,8 @@ class Command(BaseCommand):
         self.stdout.write(f"{result.pk}: {result.status}, {result.row_count} rows")
         return True
 
-    def _next_export(self, recover):
-        states = ["queued", "processing"] if recover else ["queued"]
-        item = self._claim(FormulaExport, states)
+    def _next_export(self, recover_immediately):
+        item = self._claim(FormulaExport, recover_immediately)
         if not item:
             return False
         try:
diff --git a/finance/migrations/0005_formuladefinition_revision.py b/finance/migrations/0005_formuladefinition_revision.py
new file mode 100644
index 0000000..0695b28
--- /dev/null
+++ b/finance/migrations/0005_formuladefinition_revision.py
@@ -0,0 +1,17 @@
+# Generated by Django 5.2.17 on 2026-09-25 17:30
+
+from django.db import migrations, models
+
+
+class Migration(migrations.Migration):
+    dependencies = [
+        ("finance", "0004_formuladefinition"),
+    ]
+
+    operations = [
+        migrations.AddField(
+            model_name="formuladefinition",
+            name="revision",
+            field=models.PositiveIntegerField(default=1),
+        ),
+    ]
diff --git a/finance/models.py b/finance/models.py
index 1de5920..e84caf7 100644
--- a/finance/models.py
+++ b/finance/models.py
@@ -121,6 +121,7 @@ class FormulaDefinition(models.Model):
     user = models.ForeignKey(settings.AUTH_USER_MODEL, on_delete=models.CASCADE)
     normalized_name = models.CharField(max_length=320)
     definition = models.JSONField()
+    revision = models.PositiveIntegerField(default=1)
     created_at = models.DateTimeField(auto_now_add=True)
     updated_at = models.DateTimeField(auto_now=True)
 
diff --git a/frontend/src/dialogs.tsx b/frontend/src/dialogs.tsx
index 0e52997..6e5e96b 100644
--- a/frontend/src/dialogs.tsx
+++ b/frontend/src/dialogs.tsx
@@ -1,5 +1,12 @@
 import { useEffect, useRef, useState } from "react";
-import { ChevronLeft, ChevronRight, Download, Save, Trash2, X } from "lucide-react";
+import {
+  ChevronLeft,
+  ChevronRight,
+  Download,
+  Save,
+  Trash2,
+  X,
+} from "lucide-react";
 import { api } from "./api";
 import {
   labels,
@@ -24,6 +31,7 @@ type FormulaExpression =
 type SavedFormulaDefinition = {
   id: string;
   definition: Record<string, unknown>;
+  revision: number;
   updated_at: string;
 };
 
@@ -34,6 +42,19 @@ function describeExpression(expression: FormulaExpression): string {
   return `(${describeExpression(expression.left)} ${operators[expression.operator]} ${describeExpression(expression.right)})`;
 }
 
+function definitionSummary(definition: Record<string, unknown>): string {
+  const text = [
+    `已载入已保存定义「${String(definition.name)}」`,
+    `含义：${String(definition.meaning)}`,
+    `适用范围：${String(definition.scope)}`,
+    `缺失/零分母处理：${String(definition.missing)}`,
+    `计算表达式：${JSON.stringify(definition.expression)}`,
+  ].join("；");
+  return text.length <= 2000
+    ? text
+    : `${text.slice(0, 2000)}…（内容过长已截断）`;
+}
+
 export function Modal({
   title,
   onClose,
@@ -349,7 +370,8 @@ export function EvidenceDialog({
     }>(),
     [side, setSide] = useState<"current" | "baseline">("current"),
     [page, setPage] = useState(0),
-    [error, setError] = useState("");
+    [error, setError] = useState(""),
+    [reloadKey, setReloadKey] = useState(0);
   useEffect(() => {
     let active = true;
     setResult(undefined);
@@ -364,8 +386,14 @@ export function EvidenceDialog({
     return () => {
       active = false;
     };
-  }, [ds.id, ctx, side, page]);
+  }, [ds.id, ctx, side, page, reloadKey]);
   function selectSide(value: "current" | "baseline") {
+    if (value === side) {
+      // 重复点击当前期间：仅在不在第一页时回到第一页触发刷新，
+      // 同期间同页直接返回，避免清空列表后停在“加载中”。
+      if (page !== 0) setPage(0);
+      return;
+    }
     setResult(undefined);
     setPage(0);
     setSide(value);
@@ -401,7 +429,8 @@ export function EvidenceDialog({
         {user.role !== "admin" && <p>客户原始标识不包含在明细中。</p>}
         {error && (
           <p role="alert" className="error">
-            {error}
+            {error}{" "}
+            <button onClick={() => setReloadKey((n) => n + 1)}>重试</button>
           </p>
         )}
         <div className="table-scroll">
@@ -435,6 +464,11 @@ export function EvidenceDialog({
                   <td>{String(r.gross ?? "")}</td>
                 </tr>
               ))}
+              {result && result.rows.length === 0 && (
+                <tr>
+                  <td colSpan={3 + ds.dimensions.length}>本期范围内没有数据</td>
+                </tr>
+              )}
             </tbody>
           </table>
         </div>
@@ -769,14 +803,25 @@ export function FormulaExportDialog({
     setBusy(true);
     setError("");
     try {
+      // 同名保存按服务端匹配规则覆盖旧版本；带上已知版本的 revision，
+      // 其他窗口先更新过则服务端返回 409，避免静默覆盖。
+      const normalizedName = (value: string) => value.trim().toLowerCase();
+      const existing = definitions.find(
+        (item) =>
+          normalizedName(String(item.definition.name)) ===
+          normalizedName(String(column.name)),
+      );
       const saved = await api<SavedFormulaDefinition>(
         `/datasets/${ds.id}/formula-definitions`,
         "POST",
-        { definition: column },
+        { definition: column, revision: existing?.revision },
       );
       setDefinitions((items) =>
         [...items.filter((item) => item.id !== saved.id), saved].sort((a, b) =>
-          String(a.definition.name).localeCompare(String(b.definition.name), "zh-CN"),
+          String(a.definition.name).localeCompare(
+            String(b.definition.name),
+            "zh-CN",
+          ),
         ),
       );
       setSelectedDefinition(saved.id);
@@ -789,9 +834,25 @@ export function FormulaExportDialog({
   function loadDefinition(id: string) {
     setSelectedDefinition(id);
     const saved = definitions.find((item) => item.id === id);
-    if (!saved) return;
+    if (!saved) {
+      // 选择空选项时同步清理当前定义，避免下拉选择与预览不一致
+      setColumn(null);
+      setMessages([]);
+      setQuestions([]);
+      return;
+    }
     setColumn(saved.definition);
-    setMessages([]);
+    // 载入定义后把原定义显式放进可见、经同意的对话上下文，
+    // 使“把刚才公式改成乘以 100”这类指代能关联到已载入表达式；
+    // 修改结果仍需重新确认，不会静默沿用旧定义。
+    setMessages([
+      { role: "user", content: definitionSummary(saved.definition) },
+      {
+        role: "assistant",
+        content:
+          "已载入该定义作为当前基准。请描述要调整的内容，我会在该定义基础上生成新的列定义，仍需你确认后再生成。",
+      },
+    ]);
     setQuestions([]);
     setDraft("");
   }
@@ -862,6 +923,11 @@ export function FormulaExportDialog({
             </button>
           </div>
         )}
+        {definitions.length > 0 && !task && (
+          <p className="muted">
+            保存同名定义会覆盖原版本；若已在其他窗口被修改，保存将提示刷新后重试。
+          </p>
+        )}
         <label className="check">
           <input
             type="checkbox"
diff --git a/tests/e2e/formula-export.spec.ts b/tests/e2e/formula-export.spec.ts
index ede33c5..9fa4ae3 100644
--- a/tests/e2e/formula-export.spec.ts
+++ b/tests/e2e/formula-export.spec.ts
@@ -59,11 +59,22 @@ test("新增列：补充信息触发追问并重新确认后生成公式工作
                 meaning: "实际收入减成本",
                 scope: "所选分组与期间",
                 missing: "结果不可用（#N/A）；零分母不可用",
-                expression: {
-                  operator: "subtract",
-                  left: { metric: "revenue" },
-                  right: { metric: "cost" },
-                },
+                expression:
+                  designCalls === 5
+                    ? {
+                        operator: "multiply",
+                        left: {
+                          operator: "subtract",
+                          left: { metric: "revenue" },
+                          right: { metric: "cost" },
+                        },
+                        right: { number: 100 },
+                      }
+                    : {
+                        operator: "subtract",
+                        left: { metric: "revenue" },
+                        right: { metric: "cost" },
+                      },
               },
             },
     });
@@ -123,11 +134,33 @@ test("新增列：补充信息触发追问并重新确认后生成公式工作
   const reopened = page.getByRole("dialog", { name: "生成公式工作簿" });
   await reopened.getByLabel("已保存定义").selectOption({ label: "收入成本差" });
   await expect(reopened).toContainText("计算表达式：(实际收入 − 成本)");
-  await reopened
-    .getByRole("button", { name: "确认并生成公式工作簿" })
-    .click();
+  // F04 回归：载入定义后，原定义作为可见上下文进入对话
+  await expect(reopened).toContainText("已载入已保存定义「收入成本差」");
+  await expect(reopened).toContainText("我：已载入已保存定义");
+  await expect(reopened).toContainText("助手：已载入该定义作为当前基准");
+  // F04 回归：选择空选项时同步清理当前定义与上下文，选择状态与预览一致
+  await reopened.getByLabel("已保存定义").selectOption({ label: "选择定义" });
+  await expect(reopened.getByText("计算表达式：(实际收入 − 成本)")).toHaveCount(
+    0,
+  );
+  await expect(reopened.getByText("已载入已保存定义")).toHaveCount(0);
+
+  await reopened.getByLabel("已保存定义").selectOption({ label: "收入成本差" });
+  await reopened.getByRole("checkbox").check();
+  await reopened.getByLabel("列计算描述").fill("把刚才公式改成乘以 100");
+  await reopened.getByRole("button", { name: "发送描述" }).click();
+  await expect(reopened).toContainText(
+    "计算表达式：((实际收入 − 成本) × 100)",
+  );
+  // 模型请求必须包含载入的原定义上下文和新的修改描述
+  expect(sentMessages.at(0)?.content).toContain(
+    "已载入已保存定义「收入成本差」",
+  );
+  expect(sentMessages.at(0)?.content).toContain('"operator":"subtract"');
+  expect(sentMessages.at(-1)?.content).toBe("把刚才公式改成乘以 100");
+  await reopened.getByRole("button", { name: "确认并生成公式工作簿" }).click();
   await expect(reopened.getByRole("status")).toContainText("生成完成", {
     timeout: 30000,
   });
-  expect(designCalls).toBe(4);
+  expect(designCalls).toBe(5);
 });
diff --git a/tests/e2e/workspace.spec.ts b/tests/e2e/workspace.spec.ts
index 20f26af..ddef609 100644
--- a/tests/e2e/workspace.spec.ts
+++ b/tests/e2e/workspace.spec.ts
@@ -30,7 +30,9 @@ test("桌面：真实导入、异常、下钻、证据、设置、Copilot、导
   });
   await uploaded;
   const dialog = page.getByRole("dialog");
-  await expect(dialog.getByRole("heading", { name: "确认导入字段" })).toBeVisible();
+  await expect(
+    dialog.getByRole("heading", { name: "确认导入字段" }),
+  ).toBeVisible();
   await dialog.getByRole("button", { name: "确认并导入" }).click();
   await expect(
     page.getByRole("heading", { name: "集团经营总览", exact: true }),
@@ -85,14 +87,27 @@ test("桌面：真实导入、异常、下钻、证据、设置、Copilot、导
   await expect(evidenceDialog).toContainText("1 条");
   await expect(evidenceDialog).toContainText("2026-08");
   await expect(evidenceDialog).toContainText("CSV");
-  await evidenceDialog.getByRole("button", { name: "基期", exact: true }).click();
+  await evidenceDialog
+    .getByRole("button", { name: "基期", exact: true })
+    .click();
   await expect(evidenceDialog).toContainText("2025-08");
   await evidenceDialog.getByLabel("关闭", { exact: true }).click();
   const evidenceOffsets: number[] = [];
+  let evidenceFailOnce = false;
+  let evidenceEmpty = false;
   await page.route("**/api/datasets/*/evidence", async (route) => {
     const request = route.request().postDataJSON() as { offset?: number };
     const offset = request.offset || 0;
+    if (evidenceFailOnce) {
+      evidenceFailOnce = false;
+      await route.fulfill({ status: 500, json: { error: "服务暂不可用" } });
+      return;
+    }
     evidenceOffsets.push(offset);
+    if (evidenceEmpty) {
+      await route.fulfill({ json: { total: 0, sheet: "CSV", rows: [] } });
+      return;
+    }
     const count = offset === 0 ? 50 : 1;
     await route.fulfill({
       json: {
@@ -111,10 +126,51 @@ test("桌面：真实导入、异常、下钻、证据、设置、Copilot、导
   });
   await page.getByLabel("查看医生A证据", { exact: true }).click();
   await expect(evidenceDialog).toContainText("第 1-50 条，共 51 条");
-  await evidenceDialog.getByRole("button", { name: "下一页", exact: true }).click();
+  await evidenceDialog
+    .getByRole("button", { name: "下一页", exact: true })
+    .click();
   await expect(evidenceDialog).toContainText("第 51-51 条，共 51 条");
   await expect(evidenceDialog).toContainText("2026-09");
   expect(evidenceOffsets).toEqual([0, 50]);
+  // F03 回归：第二页时重复点击当前期间应回到第一页并重新加载
+  await evidenceDialog
+    .getByRole("button", { name: "本期", exact: true })
+    .click();
+  await expect(evidenceDialog).toContainText("第 1-50 条，共 51 条");
+  expect(evidenceOffsets).toEqual([0, 50, 0]);
+  // F03 回归：第一页重复点击当前期间不清空列表、不重新请求、不停在“加载中”
+  await evidenceDialog
+    .getByRole("button", { name: "本期", exact: true })
+    .click();
+  await expect(evidenceDialog).toContainText("第 1-50 条，共 51 条");
+  await expect(evidenceDialog).not.toContainText("加载中");
+  expect(evidenceOffsets).toEqual([0, 50, 0]);
+  // F03 回归：快速本期/基期切换最终停留在当前期间的数据
+  await evidenceDialog
+    .getByRole("button", { name: "基期", exact: true })
+    .click();
+  await evidenceDialog
+    .getByRole("button", { name: "本期", exact: true })
+    .click();
+  await expect(evidenceDialog).toContainText("第 1-50 条，共 51 条");
+  await expect(evidenceDialog).not.toContainText("加载中");
+  // F03 回归：请求失败显示错误并可重试恢复
+  evidenceFailOnce = true;
+  await evidenceDialog
+    .getByRole("button", { name: "下一页", exact: true })
+    .click();
+  await expect(evidenceDialog).toContainText("服务暂不可用");
+  await expect(evidenceDialog).toContainText("加载失败");
+  await evidenceDialog.getByRole("button", { name: "重试" }).click();
+  await expect(evidenceDialog).toContainText("第 51-51 条，共 51 条");
+  // F03 回归：空结果显示空态而不是过期数据或加载中
+  evidenceEmpty = true;
+  await evidenceDialog
+    .getByRole("button", { name: "上一页", exact: true })
+    .click();
+  await expect(evidenceDialog).toContainText("本期范围内没有数据");
+  await expect(evidenceDialog).toContainText("共 0 条");
+  await expect(evidenceDialog).not.toContainText("加载中");
   await evidenceDialog.getByLabel("关闭", { exact: true }).click();
   await page
     .getByRole("navigation", { name: "分析路径" })
@@ -164,38 +220,53 @@ test("桌面：真实导入、异常、下钻、证据、设置、Copilot、导
   await page.route("**/formula-design", async (route) => {
     designCalls++;
     await route.fulfill({
-      json: designCalls === 1
-        ? { complete: false, questions: ["缺失值和零分母如何处理？"], column: null }
-        : {
-            complete: true,
-            questions: [],
-            column: {
-              name: "收入成本差",
-              meaning: "实际收入减成本",
-              scope: "所选分组与期间",
-              missing: "结果不可用（#N/A）；零分母不可用",
-              expression: {
-                operator: "subtract",
-                left: { metric: "revenue" },
-                right: { metric: "cost" },
+      json:
+        designCalls === 1
+          ? {
+              complete: false,
+              questions: ["缺失值和零分母如何处理？"],
+              column: null,
+            }
+          : {
+              complete: true,
+              questions: [],
+              column: {
+                name: "收入成本差",
+                meaning: "实际收入减成本",
+                scope: "所选分组与期间",
+                missing: "结果不可用（#N/A）；零分母不可用",
+                expression: {
+                  operator: "subtract",
+                  left: { metric: "revenue" },
+                  right: { metric: "cost" },
+                },
               },
             },
-          },
     });
   });
   await page.getByRole("button", { name: "生成公式工作簿" }).click();
   const formulaDialog = page.getByRole("dialog", { name: "生成公式工作簿" });
   await formulaDialog.getByRole("checkbox").check();
-  await formulaDialog.getByLabel("列计算描述").fill("新增收入成本差，缺失或零分母显示不可用");
+  await formulaDialog
+    .getByLabel("列计算描述")
+    .fill("新增收入成本差，缺失或零分母显示不可用");
   await formulaDialog.getByRole("button", { name: "发送描述" }).click();
   await expect(formulaDialog).toContainText("缺失值和零分母如何处理？");
-  await formulaDialog.getByLabel("列计算描述").fill("缺失项和零分母都按不可用处理");
+  await formulaDialog
+    .getByLabel("列计算描述")
+    .fill("缺失项和零分母都按不可用处理");
   await formulaDialog.getByRole("button", { name: "发送描述" }).click();
-  await formulaDialog.getByRole("button", { name: "确认并生成公式工作簿" }).click();
-  await expect(formulaDialog.getByRole("status")).toContainText("生成完成", { timeout: 30000 });
+  await formulaDialog
+    .getByRole("button", { name: "确认并生成公式工作簿" })
+    .click();
+  await expect(formulaDialog.getByRole("status")).toContainText("生成完成", {
+    timeout: 30000,
+  });
   const formulaDownload = page.waitForEvent("download");
   await formulaDialog.getByRole("button", { name: "下载公式工作簿" }).click();
-  expect((await formulaDownload).suggestedFilename()).toBe("traceiq-formula-analysis.xlsx");
+  expect((await formulaDownload).suggestedFilename()).toBe(
+    "traceiq-formula-analysis.xlsx",
+  );
   await formulaDialog.getByLabel("关闭").click();
 });
 test("手机：登录、总览、设置、Copilot 布局无横向溢出", async ({ page }) => {
diff --git a/tests/test_assistant_api.py b/tests/test_assistant_api.py
index 62d9895..deea87e 100644
--- a/tests/test_assistant_api.py
+++ b/tests/test_assistant_api.py
@@ -168,16 +168,21 @@ def test_formula_definition_library_is_private_validated_and_audited(workspace):
     saved = client.post(url, data={"definition": definition}, content_type="application/json")
     assert saved.status_code == 201
     definition_id = saved.json()["id"]
+    assert saved.json()["revision"] == 1
     assert client.get(url).json()["items"][0]["definition"] == definition
 
     updated = client.post(
         url,
-        data={"definition": {**definition, "name": "  收入成本差  ", "meaning": "更新说明"}},
+        data={
+            "definition": {**definition, "name": "  收入成本差  ", "meaning": "更新说明"},
+            "revision": 1,
+        },
         content_type="application/json",
     )
     assert updated.status_code == 200
     assert FormulaDefinition.objects.filter(dataset=ds, user=ds.created_by).count() == 1
     assert updated.json()["definition"]["meaning"] == "更新说明"
+    assert updated.json()["revision"] == 2
 
     other = Client()
     other_user = ds.created_by.__class__.objects.create_user(
diff --git a/tests/test_formula_cleanup.py b/tests/test_formula_cleanup.py
index 19d5475..33acc9d 100644
--- a/tests/test_formula_cleanup.py
+++ b/tests/test_formula_cleanup.py
@@ -1,4 +1,5 @@
 import os
+import uuid
 from datetime import timedelta
 
 import pytest
@@ -27,6 +28,12 @@ def cleanup_root(settings, tmp_path):
     return tmp_path / "formula-exports", dataset, user
 
 
+def _aged(path, hours=25):
+    stamp = (timezone.now() - timedelta(hours=hours)).timestamp()
+    os.utime(path, (stamp, stamp))
+    return path
+
+
 def test_cleanup_expires_jobs_and_removes_files(cleanup_root):
     root, dataset, user = cleanup_root
     path = root / str(user.pk) / "expired.xlsx"
@@ -55,8 +62,7 @@ def test_cleanup_removes_only_stale_temporary_files(cleanup_root):
     stale, recent = user_dir / "stale.tmp", user_dir / "recent.tmp"
     stale.touch()
     recent.touch()
-    old = timezone.now().timestamp() - 25 * 3600
-    os.utime(stale, (old, old))
+    _aged(stale)
 
     call_command("cleanup_formula_exports")
 
@@ -81,3 +87,71 @@ def test_cleanup_does_not_delete_files_outside_export_root(cleanup_root, tmp_pat
     job.refresh_from_db()
     assert job.status == "expired"
     assert outside.read_bytes() == b"keep"
+
+
+def test_cleanup_reclaims_failed_job_leftover_after_retention(cleanup_root):
+    root, dataset, user = cleanup_root
+    job = FormulaExport.objects.create(dataset=dataset, user=user, status="failed")
+    path = root / str(user.pk) / f"{job.pk}.xlsx"
+    path.parent.mkdir(parents=True)
+    path.write_bytes(b"leftover")
+    job.path = str(path)
+    job.save(update_fields=["path"])
+    _aged(path)
+
+    call_command("cleanup_formula_exports")
+
+    job.refresh_from_db()
+    assert not path.exists()
+    assert job.path == ""
+
+
+def test_cleanup_keeps_workbook_of_unexpired_succeeded_job(cleanup_root):
+    root, dataset, user = cleanup_root
+    job = FormulaExport.objects.create(
+        dataset=dataset,
+        user=user,
+        status="succeeded",
+        expires_at=timezone.now() + timedelta(hours=1),
+    )
+    path = root / str(user.pk) / f"{job.pk}.xlsx"
+    path.parent.mkdir(parents=True)
+    path.write_bytes(b"active")
+    job.path = str(path)
+    job.save(update_fields=["path"])
+    _aged(path)
+
+    call_command("cleanup_formula_exports")
+
+    assert path.read_bytes() == b"active"
+    job.refresh_from_db()
+    assert job.path == str(path)
+
+
+def test_cleanup_reclaims_orphan_whose_job_row_was_deleted(cleanup_root):
+    root, _, user = cleanup_root
+    user_dir = root / str(user.pk)
+    user_dir.mkdir(parents=True)
+    orphan = user_dir / f"{uuid.uuid4()}.xlsx"
+    orphan.write_bytes(b"orphan")
+    recent = user_dir / f"{uuid.uuid4()}.xlsx"
+    recent.write_bytes(b"recent orphan")
+    _aged(orphan)
+
+    call_command("cleanup_formula_exports")
+
+    assert not orphan.exists()
+    assert recent.exists()
+
+
+def test_cleanup_ignores_non_uuid_workbook_names(cleanup_root):
+    root, _, user = cleanup_root
+    user_dir = root / str(user.pk)
+    user_dir.mkdir(parents=True)
+    odd = user_dir / "notes.xlsx"
+    odd.write_bytes(b"keep")
+    _aged(odd)
+
+    call_command("cleanup_formula_exports")
+
+    assert odd.exists()
diff --git a/tests/test_formula_columns.py b/tests/test_formula_columns.py
index d9a9f8b..8fa9146 100644
--- a/tests/test_formula_columns.py
+++ b/tests/test_formula_columns.py
@@ -97,8 +97,11 @@ def test_formula_workbook_exposes_custom_metric_as_recalculable_column(formula_d
     custom = workbook["自定义指标"]
     assert custom["B1"].value == "收入成本差（本期）"
     assert custom["B2"].value == f"='分析汇总'!C{metric_row}"
-    assert "含义：实际收入减成本" in workbook["使用说明"]["B12"].value
-    assert "范围：按所选分组及本期、基期计算" in workbook["使用说明"]["B12"].value
+    custom_note = next(
+        str(row[1]) for row in workbook["使用说明"].iter_rows(values_only=True) if row[0] == "收入成本差"
+    )
+    assert "含义：实际收入减成本" in custom_note
+    assert "范围：按所选分组及本期、基期计算" in custom_note
 
 
 def test_custom_expression_rejects_formula_text_and_unknown_metrics():
@@ -200,6 +203,54 @@ def test_design_column_preserves_clarification_branch_and_full_conversation(form
     assert payload["input"][1:] == messages
 
 
+def test_design_column_accepts_loaded_definition_context(formula_data, settings):
+    ds, _ = formula_data
+    settings.AI_ENABLED = True
+    settings.OPENAI_API_KEY = "test-key"
+    settings.OPENAI_MODEL = "test-model"
+    context = "已载入已保存定义「收入成本差」；计算表达式：" + '{"operator":"subtract",' * 45
+    assert 1000 < len(context) <= 2000
+    messages = [
+        {"role": "user", "content": context},
+        {"role": "assistant", "content": "已载入该定义作为当前基准。"},
+        {"role": "user", "content": "把刚才公式改成乘以 100"},
+    ]
+    clarification = {
+        "status": "completed",
+        "output": [
+            {
+                "content": [
+                    {
+                        "type": "output_text",
+                        "text": '{"complete":false,"questions":["要乘在整个表达式上吗？"],"column":null}',
+                    }
+                ]
+            }
+        ],
+    }
+    with patch("finance.formula_views.httpx.Client") as http:
+        http.return_value.__enter__.return_value.post.return_value = type(
+            "Response",
+            (),
+            {"raise_for_status": lambda self: None, "json": lambda self: clarification},
+        )()
+        result = design_column(
+            messages,
+            ds,
+            {"current": ["2026-08"], "baseline": ["2025-08"], "filters": {}, "group": "department"},
+        )
+    assert result["complete"] is False
+    payload = http.return_value.__enter__.return_value.post.call_args.kwargs["json"]
+    assert payload["input"][1:] == messages
+
+    with pytest.raises(Exception, match="对话消息内容无效"):
+        design_column(
+            [{"role": "user", "content": "y" * 2001}],
+            ds,
+            {"current": ["2026-08"], "baseline": ["2025-08"], "filters": {}, "group": "department"},
+        )
+
+
 def test_design_column_returns_only_validated_formula_definition(formula_data, settings):
     ds, _ = formula_data
     settings.AI_ENABLED = True
diff --git a/tests/test_formula_definitions_concurrency.py b/tests/test_formula_definitions_concurrency.py
new file mode 100644
index 0000000..d0c8483
--- /dev/null
+++ b/tests/test_formula_definitions_concurrency.py
@@ -0,0 +1,123 @@
+from unittest.mock import patch
+
+import pytest
+from django.db import connection
+
+from finance.models import FormulaDefinition
+
+pytestmark = pytest.mark.django_db
+
+
+def column_definition(name="收入成本差"):
+    return {
+        "name": name,
+        "meaning": "实际收入减成本",
+        "scope": "所选分组与期间",
+        "missing": "结果不可用（#N/A）；零分母不可用",
+        "expression": {
+            "operator": "subtract",
+            "left": {"metric": "revenue"},
+            "right": {"metric": "cost"},
+        },
+    }
+
+
+def _post(client, url, definition, revision=None):
+    return client.post(
+        url,
+        data={"definition": definition, **({"revision": revision} if revision is not None else {})},
+        content_type="application/json",
+    )
+
+
+def test_same_name_update_requires_matching_revision(workspace):
+    ds, _, client, *_ = workspace
+    url = f"/api/datasets/{ds.pk}/formula-definitions"
+
+    saved = _post(client, url, column_definition())
+    assert saved.status_code == 201
+    item_id, revision = saved.json()["id"], saved.json()["revision"]
+
+    # 缺失或过期的 revision 都不能静默覆盖其他窗口的修改
+    assert _post(client, url, column_definition()).status_code == 409
+    assert _post(client, url, column_definition(), revision=999).status_code == 409
+
+    updated = _post(client, url, column_definition(), revision=revision)
+    assert updated.status_code == 200
+    assert updated.json()["revision"] == revision + 1
+
+    # 基于旧版本的再次更新必须被拒绝，两个标签页不互相静默丢失
+    stale = _post(client, url, column_definition(), revision=revision)
+    assert stale.status_code == 409
+    current = FormulaDefinition.objects.get(pk=item_id)
+    assert current.revision == revision + 1
+    # 列表接口暴露当前 revision 供前端携带
+    listed = client.get(url).json()["items"]
+    assert listed[0]["revision"] == revision + 1
+
+
+def test_definition_quota_is_enforced_and_release_allows_new_save(workspace):
+    ds, _, client, *_ = workspace
+    url = f"/api/datasets/{ds.pk}/formula-definitions"
+    for index in range(50):
+        FormulaDefinition.objects.create(
+            dataset=ds,
+            user=ds.created_by,
+            normalized_name=f"定义{index:02d}",
+            definition=column_definition(f"定义{index:02d}"),
+        )
+
+    rejected = _post(client, url, column_definition("第 51 个不同名定义"))
+    assert rejected.status_code == 400
+    assert "50" in rejected.json()["error"]
+    assert FormulaDefinition.objects.filter(dataset=ds).count() == 50
+
+    # 同名更新不受配额限制；删除后名额释放可再保存新名称
+    assert _post(client, url, column_definition("定义00"), revision=1).status_code == 200
+    FormulaDefinition.objects.filter(normalized_name="定义00").delete()
+    assert _post(client, url, column_definition("第 51 个不同名定义")).status_code == 201
+
+
+def test_audit_failure_rolls_back_definition_change(workspace):
+    ds, _, client, *_ = workspace
+    url = f"/api/datasets/{ds.pk}/formula-definitions"
+
+    with patch("finance.formula_views.audit", side_effect=RuntimeError("audit unavailable")):
+        with pytest.raises(RuntimeError, match="audit unavailable"):
+            _post(client, url, column_definition())
+
+    assert FormulaDefinition.objects.filter(dataset=ds).count() == 0
+
+
+def test_concurrent_distinct_saves_cannot_exceed_quota_on_postgres(workspace):
+    if connection.vendor != "postgresql":
+        pytest.skip("并发配额依赖 PostgreSQL 行锁语义，生产库单独验证")
+    import threading
+
+    from django.test import Client
+
+    ds, member, *_ = workspace
+    user = ds.created_by
+    for index in range(49):
+        FormulaDefinition.objects.create(
+            dataset=ds,
+            user=user,
+            normalized_name=f"定义{index:02d}",
+            definition=column_definition(f"定义{index:02d}"),
+        )
+
+    barrier = threading.Barrier(4)
+
+    def save(index):
+        client = Client()
+        assert client.login(username=user.username, password="Finance-TestPassword991!")
+        barrier.wait()
+        _post(client, f"/api/datasets/{ds.pk}/formula-definitions", column_definition(f"并发{index}"))
+
+    threads = [threading.Thread(target=save, args=(index,)) for index in range(4)]
+    for thread in threads:
+        thread.start()
+    for thread in threads:
+        thread.join()
+
+    assert FormulaDefinition.objects.filter(dataset=ds).count() == 50
diff --git a/tests/test_formula_export_reliability.py b/tests/test_formula_export_reliability.py
index 0530117..9376041 100644
--- a/tests/test_formula_export_reliability.py
+++ b/tests/test_formula_export_reliability.py
@@ -1,3 +1,4 @@
+import os
 import stat
 from decimal import Decimal
 from io import BytesIO
@@ -198,6 +199,170 @@ def test_formula_export_worker_records_path_preparation_failure(workspace, setti
     assert job.status == "failed"
 
 
+def _queued_job(ds):
+    return FormulaExport.objects.create(
+        dataset=ds,
+        user=ds.created_by,
+        context={"start": "2026-08", "end": "2026-08", "comparison": "yoy", "group": "department"},
+    )
+
+
+def test_failed_export_rejects_symlinked_user_directory_without_touching_files(workspace, settings):
+    from finance.exporting import process_formula_export
+
+    ds, *_ = workspace
+    job = _queued_job(ds)
+    root = Path(settings.MEDIA_ROOT) / "formula-exports"
+    root.mkdir(parents=True)
+    protected = Path(settings.MEDIA_ROOT) / "protected"
+    protected.mkdir()
+    victim = protected / f"{job.pk}.xlsx"
+    victim.write_bytes(b"protected workbook")
+    (root / str(job.user_id)).symlink_to(protected, target_is_directory=True)
+
+    with pytest.raises(ValueError, match="符号链接"):
+        process_formula_export(job.pk)
+
+    job.refresh_from_db()
+    assert job.status == "failed"
+    assert not job.path
+    assert victim.read_bytes() == b"protected workbook"
+    assert list(protected.glob("*.tmp")) == []
+
+
+def test_failed_export_rejects_symlinked_target_and_keeps_link(workspace, settings):
+    from finance.exporting import process_formula_export
+
+    ds, *_ = workspace
+    job = _queued_job(ds)
+    user_dir = Path(settings.MEDIA_ROOT) / "formula-exports" / str(job.user_id)
+    user_dir.mkdir(parents=True)
+    protected_file = Path(settings.MEDIA_ROOT) / "protected.xlsx"
+    protected_file.write_bytes(b"protected workbook")
+    (user_dir / f"{job.pk}.xlsx").symlink_to(protected_file)
+
+    with pytest.raises(ValueError, match="普通文件"):
+        process_formula_export(job.pk)
+
+    job.refresh_from_db()
+    assert job.status == "failed"
+    assert protected_file.read_bytes() == b"protected workbook"
+    assert (user_dir / f"{job.pk}.xlsx").is_symlink()
+    assert list(user_dir.glob("*.tmp")) == []
+
+
+def test_failed_export_refuses_cleanup_when_parent_directory_swapped_to_symlink(
+    workspace, settings, monkeypatch
+):
+    from finance.exporting import process_formula_export
+
+    ds, *_ = workspace
+    job = _queued_job(ds)
+    root = Path(settings.MEDIA_ROOT) / "formula-exports"
+    user_dir = root / str(job.user_id)
+    user_dir.mkdir(parents=True)
+    protected = Path(settings.MEDIA_ROOT) / "protected"
+    protected.mkdir()
+    victim = protected / f"{job.pk}.xlsx"
+    victim.write_bytes(b"protected workbook")
+
+    def swap_and_fail(*args, **kwargs):
+        args[-1].write(b"partial")
+        user_dir.rename(user_dir.with_name("swapped"))
+        user_dir.symlink_to(protected, target_is_directory=True)
+        raise ValueError("controlled failure after path validation")
+
+    monkeypatch.setattr("finance.exporting.build_formula_workbook", swap_and_fail)
+    with pytest.raises(ValueError, match="controlled failure"):
+        process_formula_export(job.pk)
+
+    job.refresh_from_db()
+    assert job.status == "failed"
+    assert victim.read_bytes() == b"protected workbook"
+    assert user_dir.is_symlink()
+    # 本次创建的临时文件残留在被换名的原目录中，由清理命令按保留期回收。
+    assert list((root / "swapped").glob(f"{job.pk}.*.tmp"))
+
+
+def test_formula_export_retry_after_failure_succeeds(workspace, monkeypatch):
+    from finance.exporting import process_formula_export
+
+    ds, *_ = workspace
+    job = _queued_job(ds)
+
+    def fail(*args, **kwargs):
+        args[-1].write(b"partial")
+        raise ValueError("controlled export failure")
+
+    monkeypatch.setattr("finance.exporting.build_formula_workbook", fail)
+    with pytest.raises(ValueError, match="controlled export failure"):
+        process_formula_export(job.pk)
+    monkeypatch.undo()
+
+    stats = process_formula_export(job.pk)
+    job.refresh_from_db()
+    assert stats["rows"] == 5
+    assert job.status == "succeeded"
+    assert Path(job.path).read_bytes()[:2] == b"PK"
+    assert not list(Path(job.path).parent.glob(f"{job.pk}.*.tmp"))
+
+
+def test_cleanup_failure_is_logged_and_reclaimed_after_permissions_restored(
+    workspace, settings, monkeypatch, caplog
+):
+    import logging as logging_module
+    from datetime import timedelta
+
+    from django.core.management import call_command
+    from django.utils import timezone
+
+    from finance.exporting import process_formula_export
+
+    ds, *_ = workspace
+    job = _queued_job(ds)
+    user_dir = Path(settings.MEDIA_ROOT) / "formula-exports" / str(job.user_id)
+    user_dir.mkdir(parents=True)
+    target = user_dir / f"{job.pk}.xlsx"
+    target.write_bytes(b"stale workbook")
+    other = FormulaExport.objects.create(
+        dataset=ds,
+        user=ds.created_by,
+        context={"start": "2026-08", "end": "2026-08", "comparison": "yoy", "group": "department"},
+        status="succeeded",
+        expires_at=timezone.now() + timedelta(hours=1),
+    )
+    kept = user_dir / f"{other.pk}.xlsx"
+    other.path = str(kept)
+    other.save(update_fields=["path"])
+    kept.write_bytes(b"active workbook")
+
+    def fail(*args, **kwargs):
+        args[-1].write(b"partial")
+        os.chmod(user_dir, 0o500)
+        raise ValueError("controlled export failure")
+
+    monkeypatch.setattr("finance.exporting.build_formula_workbook", fail)
+    with pytest.raises(ValueError, match="controlled export failure"):
+        process_formula_export(job.pk)
+    job.refresh_from_db()
+    assert job.status == "failed"
+    assert job.path == str(target)
+    warnings = [r for r in caplog.records if r.levelno == logging_module.WARNING]
+    assert warnings and all(str(job.pk) in record.getMessage() for record in warnings)
+
+    os.chmod(user_dir, 0o700)
+    old = (timezone.now() - timedelta(hours=25)).timestamp()
+    for leftover in [target, *user_dir.glob(f"{job.pk}.*.tmp"), kept]:
+        os.utime(leftover, (old, old))
+    call_command("cleanup_formula_exports")
+
+    job.refresh_from_db()
+    assert not target.exists()
+    assert not list(user_dir.glob(f"{job.pk}.*.tmp"))
+    assert job.path == ""
+    assert kept.read_bytes() == b"active workbook"
+
+
 def test_formula_workbook_rejects_unclassified_label_collision(workspace):
     from decimal import Decimal
 
diff --git a/tests/test_formula_reconcile_guard.py b/tests/test_formula_reconcile_guard.py
index a4b9131..da6b6f4 100644
--- a/tests/test_formula_reconcile_guard.py
+++ b/tests/test_formula_reconcile_guard.py
@@ -20,11 +20,12 @@ def test_reconcile_status_guards_non_numeric_difference(workspace):
         output,
     )
     workbook = load_workbook(BytesIO(output.getvalue()), data_only=False)
+    status_cells = [row[7] for row in workbook["对账检查"].iter_rows(min_row=2)]
     formulas = [
-        row[6].value
-        for row in workbook["对账检查"].iter_rows(min_row=2)
-        if isinstance(row[6].value, str) and row[6].value.startswith("=IF(D")
+        cell.value for cell in status_cells if isinstance(cell.value, str) and cell.value.startswith("=IF(D")
     ]
     assert formulas
     assert all("ISNUMBER(F" in formula for formula in formulas)
     assert all('"通过"' in formula for formula in formulas)
+    # 容差列独立成列，状态按容差比较而不是统一 0.01
+    assert all(row[6].value is not None for row in workbook["对账检查"].iter_rows(min_row=2))
diff --git a/tests/test_formula_reconcile_tolerances.py b/tests/test_formula_reconcile_tolerances.py
new file mode 100644
index 0000000..cc959bf
--- /dev/null
+++ b/tests/test_formula_reconcile_tolerances.py
@@ -0,0 +1,236 @@
+import os
+import shutil
+import subprocess
+from decimal import Decimal
+from io import BytesIO
+
+import pytest
+from openpyxl import load_workbook
+
+from finance.exporting import CUSTOM_COLUMN_TOLERANCE, RECONCILE_TOLERANCES, build_formula_workbook
+from finance.formula_columns import to_decimal
+
+pytestmark = pytest.mark.django_db
+
+
+def custom_column(name="收入成本差"):
+    return {
+        "name": name,
+        "meaning": "实际收入减成本",
+        "scope": "按所选分组及本期、基期计算",
+        "missing": "结果不可用（#N/A）；零分母不可用",
+        "expression": {
+            "operator": "subtract",
+            "left": {"metric": "revenue"},
+            "right": {"metric": "cost"},
+        },
+    }
+
+
+def test_to_decimal_evaluator_golden_samples():
+    values = {
+        "revenue": Decimal("999999999999.99"),
+        "cost": Decimal("0.0000001"),
+        "customers": Decimal("7"),
+    }
+    # 大数与多层表达式：((收入 - 收入) + 成本×2) ÷ (成本×2) = 1，抵消后不丢失精度
+    cancellation = {
+        "operator": "divide",
+        "left": {
+            "operator": "add",
+            "left": {
+                "operator": "subtract",
+                "left": {"metric": "revenue"},
+                "right": {"metric": "revenue"},
+            },
+            "right": {"operator": "multiply", "left": {"metric": "cost"}, "right": {"number": 2}},
+        },
+        "right": {"operator": "multiply", "left": {"metric": "cost"}, "right": {"number": 2}},
+    }
+    assert to_decimal(cancellation, values) == Decimal("1")
+    # 极小值与比率
+    assert to_decimal({"metric": "cost"}, values) == Decimal("0.0000001")
+    assert to_decimal(
+        {"operator": "divide", "left": {"metric": "revenue"}, "right": {"metric": "customers"}}, values
+    ) == Decimal("999999999999.99") / Decimal("7")
+    # 除零与缺失指标返回 None（对应 Excel #N/A）
+    assert (
+        to_decimal({"operator": "divide", "left": {"metric": "revenue"}, "right": {"number": 0}}, values)
+        is None
+    )
+    assert (
+        to_decimal(
+            {"operator": "add", "left": {"metric": "cost"}, "right": {"metric": "customers"}},
+            {**values, "customers": None},
+        )
+        is None
+    )
+    # 负数与常数组合
+    assert to_decimal(
+        {"operator": "subtract", "left": {"number": -2.5}, "right": {"metric": "cost"}}, values
+    ) == Decimal("-2.5") - Decimal("0.0000001")
+
+
+def test_reconcile_uses_independent_tolerances_per_metric_kind(workspace):
+    ds, member, *_ = workspace
+    output = BytesIO()
+    build_formula_workbook(
+        ds,
+        member,
+        {"start": "2026-08", "end": "2026-08", "comparison": "yoy", "group": "department"},
+        output,
+    )
+    workbook = load_workbook(BytesIO(output.getvalue()), data_only=False)
+    sheet = workbook["对账检查"]
+    rows = list(sheet.iter_rows(values_only=True))
+    assert list(rows[0]) == [
+        "指标",
+        "分组",
+        "期间侧",
+        "服务器 Decimal 参考",
+        "公式结果",
+        "差异",
+        "容差",
+        "状态",
+    ]
+
+    def find(metric, side="本期"):
+        return next(row for row in rows[1:] if row[0] == metric and row[2] == side)
+
+    assert find("消费人数")[6] == 0  # 人数严格相等
+    assert find("实际收入")[6] == pytest.approx(0.01)
+    assert find("毛利贡献")[6] == pytest.approx(0.01)
+    assert find("毛利率")[6] == pytest.approx(0.0001)
+    assert find("ROI")[6] == pytest.approx(0.0001)
+    assert find("客单价")[6] == pytest.approx(0.0001)
+    # 状态公式按容差列比较（<=G），不再统一 <0.01
+    status = next(
+        cell.value for cell in sheet[2] if isinstance(cell.value, str) and cell.value.startswith("=IF(D")
+    )
+    assert "<=G2" in status
+    # 参数页写明对账规则，容差调整需财务批准
+    parameters = {str(row[0]): row for row in workbook["参数与口径"].iter_rows(values_only=True) if row[0]}
+    assert "人数严格相等" in parameters["对账容差"][1]
+    assert parameters["对账容差"][3] == "否"
+
+
+def test_custom_columns_reconcile_against_decimal_reference(workspace):
+    ds, member, *_ = workspace
+    output = BytesIO()
+    build_formula_workbook(
+        ds,
+        member,
+        {
+            "start": "2026-08",
+            "end": "2026-08",
+            "comparison": "yoy",
+            "group": "department",
+            "custom_columns": [custom_column()],
+        },
+        output,
+    )
+    workbook = load_workbook(BytesIO(output.getvalue()), data_only=False)
+    rows = list(workbook["对账检查"].iter_rows(values_only=True))
+    custom_rows = [row for row in rows[1:] if row[0] == "收入成本差"]
+    assert custom_rows
+    pediatrics = next(row for row in custom_rows if row[1] == "儿科" and row[2] == "本期")
+    assert Decimal(pediatrics[3]) == Decimal("60")  # 100 - 40
+    assert pediatrics[6] == pytest.approx(float(CUSTOM_COLUMN_TOLERANCE))
+    status = next(
+        row[7]
+        for row in workbook["对账检查"].iter_rows(min_row=2, values_only=True)
+        if row[0] == "收入成本差" and row[1] == "儿科" and row[2] == "本期"
+    )
+    assert "<=G" in status and "通过" in status
+    # 旧的占位文案不再出现
+    assert all("自定义公式无服务器参考" not in str(row[3]) for row in custom_rows)
+    assert all(row[7] != "需业务验收" for row in custom_rows)
+
+
+def test_custom_column_reference_missing_when_dependency_unavailable(workspace):
+    from finance.models import Fact
+
+    ds, member, *_ = workspace
+    # 外科 2026-08 成本置空 → cost 不可用 → 收入成本差参考为 None → 无参考
+    Fact.objects.filter(dataset=ds, period="2026-08", dimensions__department="外科").update(cost=None)
+    output = BytesIO()
+    build_formula_workbook(
+        ds,
+        member,
+        {
+            "start": "2026-08",
+            "end": "2026-08",
+            "comparison": "yoy",
+            "group": "department",
+            "custom_columns": [custom_column()],
+        },
+        output,
+    )
+    workbook = load_workbook(BytesIO(output.getvalue()), data_only=False)
+    surgery = next(
+        row
+        for row in workbook["对账检查"].iter_rows(min_row=2, values_only=True)
+        if row[0] == "收入成本差" and row[1] == "外科" and row[2] == "本期"
+    )
+    assert surgery[3] in ("", None)  # 无服务器参考
+    assert "无参考" in surgery[7]
+
+
+def test_reconcile_tolerances_cover_every_metric():
+    from finance.analysis import METRICS
+
+    assert set(RECONCILE_TOLERANCES) == set(METRICS)
+    assert RECONCILE_TOLERANCES["customers"] == 0
+    assert all(
+        tolerance in (Decimal("0.01"), Decimal("0.0001"))
+        for key, tolerance in RECONCILE_TOLERANCES.items()
+        if key != "customers"
+    )
+
+
+def test_reconcile_passes_in_libreoffice_with_per_metric_tolerances(workspace, tmp_path):
+    executable = os.environ.get("SOFFICE_BIN") or shutil.which("soffice")
+    if not executable:
+        pytest.skip("set SOFFICE_BIN to run the LibreOffice recalculation check")
+    ds, member, *_ = workspace
+    source, output = tmp_path / "source", tmp_path / "out"
+    source.mkdir()
+    output.mkdir()
+    workbook_path = source / "reconcile.xlsx"
+    with workbook_path.open("wb") as file:
+        build_formula_workbook(
+            ds,
+            member,
+            {
+                "start": "2026-08",
+                "end": "2026-08",
+                "comparison": "yoy",
+                "group": "department",
+                "custom_columns": [custom_column()],
+            },
+            file,
+        )
+    profile = tmp_path / "profile"
+    result = subprocess.run(
+        [
+            executable,
+            "--headless",
+            f"-env:UserInstallation={profile.as_uri()}",
+            "--convert-to",
+            "xlsx",
+            "--outdir",
+            str(output),
+            str(workbook_path),
+        ],
+        capture_output=True,
+        text=True,
+        timeout=120,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr or result.stdout
+    values = load_workbook(output / workbook_path.name, data_only=True)
+    statuses = [
+        (row[0], row[1], row[2], row[7]) for row in values["对账检查"].iter_rows(min_row=2, values_only=True)
+    ]
+    assert statuses
+    assert all(status in ("通过", "无参考") for *_head, status in statuses), statuses
diff --git a/tests/test_formula_workbook_docs.py b/tests/test_formula_workbook_docs.py
new file mode 100644
index 0000000..9542b12
--- /dev/null
+++ b/tests/test_formula_workbook_docs.py
@@ -0,0 +1,155 @@
+from io import BytesIO
+
+import pytest
+from openpyxl import load_workbook
+
+from finance.exporting import build_formula_workbook
+
+pytestmark = pytest.mark.django_db
+
+EDITABLE_INPUT_HEADERS = ["Revenue", "Cost", "GrossInput", "MarginRate", "ListAmount"]
+
+
+def _workbook(workspace, **extra):
+    ds, member, *_ = workspace
+    output = BytesIO()
+    build_formula_workbook(
+        ds,
+        member,
+        {"start": "2026-08", "end": "2026-08", "comparison": "yoy", "group": "department", **extra},
+        output,
+    )
+    return load_workbook(BytesIO(output.getvalue()), data_only=False)
+
+
+def test_parameter_sheet_is_read_only_snapshot(workspace):
+    workbook = _workbook(workspace)
+    rows = list(workbook["参数与口径"].iter_rows(values_only=True))
+    header = list(rows[0])
+    editable = header.index("可编辑")
+    assert all(row[editable] == "否" for row in rows[1:] if row[0])
+
+    margin = next(row for row in rows if row[0] and "产品毛利率" in str(row[0]))
+    assert "只读" in margin[2]
+    assert "MarginRate" in margin[2]
+    assert "计算明细" in margin[2]
+
+    # 维度键与显示名完整对照
+    texts = [str(row[0]) for row in rows if row[0]]
+    assert "维度 department" in texts and "维度 product" in texts
+    department = next(row for row in rows if row[0] == "维度 department")
+    assert department[1] == "科室"
+    product = next(row for row in rows if row[0] == "维度 product")
+    assert product[1] == "产品一级分类"
+
+
+def test_instructions_document_editable_inputs_with_real_dependencies(workspace):
+    workbook = _workbook(workspace)
+    instructions = {
+        str(row[0]): str(row[1]) for row in workbook["使用说明"].iter_rows(values_only=True) if row[0]
+    }
+
+    data_headers = [cell.value for cell in workbook["计算明细"][1]]
+    sources = {
+        "Revenue": "实际收入",
+        "Cost": "成本",
+        "GrossInput": "毛利贡献",
+        "MarginRate": "产品毛利率表",
+        "ListAmount": "应收金额",
+    }
+    for header in EDITABLE_INPUT_HEADERS:
+        key = next(k for k in instructions if k.startswith(f"输入 {header}（"))
+        text = instructions[key]
+        assert f"计算明细页 {header} 列" in text
+        assert sources[header] in text
+        assert header in data_headers
+
+    # 每个可修改输入都被汇总公式真实引用
+    summary_formula_cells = [
+        cell.value
+        for row in workbook["分析汇总"].iter_rows(min_row=6)
+        for cell in row
+        if isinstance(cell.value, str) and cell.value.startswith("=IF")
+    ]
+    joined = "\n".join(summary_formula_cells)
+    for header in ("Revenue", "Cost", "ListAmount"):
+        assert f"Data[{header}]" in joined
+    detail_formula = next(
+        cell.value
+        for row in workbook["计算明细"].iter_rows(min_row=2)
+        for cell in row
+        if isinstance(cell.value, str) and cell.value.startswith("=IF(")
+    )
+    assert "GrossInput" not in detail_formula  # 列名按相对引用参与 GrossCalc
+    assert detail_formula.startswith("=IF(")
+
+    boundary = instructions["固定结构"]
+    assert "人数去重" in boundary and "固定快照" in boundary
+    assert "重新导出" in boundary
+    assert "参数与口径" in instructions["维度列"]
+
+
+def test_margin_rate_lookup_edit_changes_gross_via_detail_recalculation(workspace, tmp_path):
+    import os
+    import shutil
+
+    from finance.models import Fact
+
+    if not (os.environ.get("SOFFICE_BIN") or shutil.which("soffice")):
+        pytest.skip("set SOFFICE_BIN to run the LibreOffice recalculation check")
+
+    ds, member, *_ = workspace
+    # 去掉直接毛利，使毛利全部走 lookup 毛利率
+    Fact.objects.filter(dataset=ds).update(gross=None)
+    mapping = {**ds.mapping, "margin_rates": {"医疗服务": "0.3"}}
+    ds.mapping = mapping
+    ds.save(update_fields=["mapping"])
+
+    source, output = tmp_path / "source", tmp_path / "out"
+    source.mkdir()
+    output.mkdir()
+    workbook_path = source / "lookup.xlsx"
+    with workbook_path.open("wb") as file:
+        build_formula_workbook(
+            ds,
+            member,
+            {"start": "2026-08", "end": "2026-08", "comparison": "yoy", "group": "department"},
+            file,
+        )
+
+    editable = load_workbook(workbook_path)
+    # MarginRate 列在维度列之后：G 列（期间+2维度=3 列后第 4 个保留列）
+    header_row = [cell.value for cell in editable["计算明细"][1]]
+    rate_column = header_row.index("MarginRate") + 1
+    editable["计算明细"].cell(row=2, column=rate_column).value = 0.5
+    editable.save(workbook_path)
+
+    import subprocess
+
+    profile = tmp_path / "profile"
+    result = subprocess.run(
+        [
+            os.environ.get("SOFFICE_BIN") or shutil.which("soffice"),
+            "--headless",
+            f"-env:UserInstallation={profile.as_uri()}",
+            "--convert-to",
+            "xlsx",
+            "--outdir",
+            str(output),
+            str(workbook_path),
+        ],
+        capture_output=True,
+        text=True,
+        timeout=120,
+        check=False,
+    )
+    assert result.returncode == 0, result.stderr or result.stdout
+    values = load_workbook(output / workbook_path.name, data_only=True)
+    summary = values["分析汇总"]
+    gross_row = next(
+        row
+        for row in range(6, summary.max_row + 1)
+        if summary.cell(row, 1).value == "毛利贡献" and summary.cell(row, 2).value == "儿科"
+    )
+    # 2026-08 儿科收入 100，lookup 毛利率改为 0.5 后毛利应为 50
+    assert summary.cell(gross_row, 3).value == pytest.approx(50)
diff --git a/tests/test_worker_paths.py b/tests/test_worker_paths.py
index 1b1bc6d..68c84b3 100644
--- a/tests/test_worker_paths.py
+++ b/tests/test_worker_paths.py
@@ -1,7 +1,10 @@
+from datetime import timedelta
+from pathlib import Path
 from unittest.mock import patch
 
 import pytest
 from django.core.management import call_command
+from django.utils import timezone
 
 from finance.models import Dataset, FormulaExport
 
@@ -30,3 +33,108 @@ def test_worker_processes_queued_import_and_reports_export_failure(workspace):
         call_command("import_worker", once=True)
     export.refresh_from_db()
     assert export.status == "processing"
+
+
+def _export_job(ds, **kwargs):
+    return FormulaExport.objects.create(
+        dataset=ds,
+        user=ds.created_by,
+        context={"start": "2026-08", "end": "2026-08", "comparison": "yoy", "group": "department"},
+        **kwargs,
+    )
+
+
+def test_worker_refuses_export_after_role_revocation(workspace):
+    from finance.exporting import process_formula_export
+
+    ds, member, *_ = workspace
+    job = _export_job(ds, status="processing")
+    member.role = "viewer"
+    member.save(update_fields=["role"])
+
+    with pytest.raises(PermissionError, match="角色"):
+        process_formula_export(job.pk)
+
+    job.refresh_from_db()
+    assert job.status == "failed"
+    assert not job.path
+    assert not list(Path(ds.file.storage.location).rglob(f"{job.pk}*"))
+
+
+def test_worker_refuses_export_after_account_deactivation(workspace):
+    from finance.exporting import process_formula_export
+
+    ds, _, _, _, _ = workspace
+    user = ds.created_by
+    user.is_active = False
+    user.save(update_fields=["is_active"])
+    job = _export_job(ds, status="processing")
+
+    with pytest.raises(PermissionError, match="停用"):
+        process_formula_export(job.pk)
+
+    job.refresh_from_db()
+    assert job.status == "failed"
+
+
+def test_status_and_download_deny_revoked_viewer(workspace, settings):
+    ds, member, client, *_ = workspace
+    job = _export_job(
+        ds,
+        status="succeeded",
+        finished_at=timezone.now(),
+        expires_at=timezone.now() + timedelta(hours=1),
+    )
+    target = Path(settings.MEDIA_ROOT) / "formula-exports" / str(job.user_id) / f"{job.pk}.xlsx"
+    target.parent.mkdir(parents=True)
+    target.write_bytes(b"PK workbook")
+    job.path = str(target)
+    job.save(update_fields=["path"])
+
+    assert client.get(f"/api/datasets/{ds.pk}/formula-exports/{job.pk}").status_code == 200
+    assert client.get(f"/api/datasets/{ds.pk}/formula-exports/{job.pk}/download").status_code == 200
+
+    member.role = "viewer"
+    member.save(update_fields=["role"])
+    assert client.get(f"/api/datasets/{ds.pk}/formula-exports/{job.pk}").status_code == 403
+    assert client.get(f"/api/datasets/{ds.pk}/formula-exports/{job.pk}/download").status_code == 403
+
+
+def test_half_finished_export_cannot_be_downloaded(workspace, settings):
+    ds, _, client, *_ = workspace
+    for status, code in (("processing", 409), ("failed", 409), ("queued", 409)):
+        job = _export_job(ds, status=status)
+        assert client.get(f"/api/datasets/{ds.pk}/formula-exports/{job.pk}/download").status_code == code, (
+            status
+        )
+    succeeded = _export_job(
+        ds,
+        status="succeeded",
+        path=str(Path(settings.MEDIA_ROOT) / "nowhere.xlsx"),
+        finished_at=timezone.now(),
+        expires_at=timezone.now() + timedelta(hours=1),
+    )
+    assert client.get(f"/api/datasets/{ds.pk}/formula-exports/{succeeded.pk}/download").status_code == 410
+
+
+def test_stale_processing_export_is_reclaimed_after_timeout(workspace):
+    ds, *_ = workspace
+    job = _export_job(ds, status="processing")
+    FormulaExport.objects.filter(pk=job.pk).update(created_at=timezone.now() - timedelta(minutes=20))
+
+    call_command("import_worker", once=True)
+
+    job.refresh_from_db()
+    assert job.status == "succeeded"
+    assert Path(job.path).read_bytes()[:2] == b"PK"
+    assert not list(Path(job.path).parent.glob(f"{job.pk}.*.tmp"))
+
+
+def test_recent_processing_export_is_not_reclaimed(workspace):
+    ds, *_ = workspace
+    job = _export_job(ds, status="processing")
+
+    call_command("import_worker", once=True)
+
+    job.refresh_from_db()
+    assert job.status == "processing"