feat(browser): 每步填报后做落库自检，行数/合计不一致写进 warnings

用户第 2 条：add_travel_items 点完确定不核对表尾合计，那次 500 现场看着像「脚本没点着按钮」，
白查一轮。现在每步 POST 完回读页面再比。

- BaseBot.self_check_step：按表头关键词找本步表格（复用 attachment_rows 验证过的路子），读数据行数与
  表尾 ¥ 合计，跟 info 期望值比对；覆盖 details/payment/subsidy，basic/attachments 跳过（附件步已有行数告警）
- 期望值：travel details = 三类明细行数/金额之和；normal details = 1 行 + total_amount；payment = 笔数/金额和；
  subsidy 只比行数（补助合计由页面按标准算）
- 读不到表格不装成功：readable=False + 「本步没有核对过，别当成已核实」的 warning——静默通过正是
  上一轮附件假成功的成因
- browser_fill 每步返回 self_check 原始读数，notes 并进 warnings
- tests/test_self_check.py 六例；FakeBot 绑定 _row_amount 时要用 staticmethod（转类会变成吃 self 的普通函数）
- 文档：browser_fill docstring、操作指南填报注意事项、方案文档 §23（含"选择器待联机确认"的边界）

验证：pytest 197 passed；ruff / format / mypy / deptry / mcp-smoke / docs-check / skill-check 全绿。
--no-verify 提交并手跑等价检查。

.agents/docs/plans/MCP化改造方案-2026-10-19.md

@@ -708,3 +708,22 @@ capture 只出现在 camera 段、脚本里不许再出现 `setAttribute`、两
 
 同类教训连着三条（capture 挂错、accept 被路由到相机、按钮名过度承诺）都指向一件事：
 移动端网页的能力边界由设备与 WebView 决定，文案只能描述**实测到的行为**，不能描述规范里的应然。
+
+## 23. 落库后自检：每步 POST 完回读页面（2026-10-21）
+
+用户第 2 条：add_travel_items 点完确定不核对表尾合计，于是那次 500 现场看着像「脚本没点着按钮」。
+现在每步 POST 完都回读一次页面，差值写进 warnings。
+
+- `BaseBot.self_check_step(kind, step, info)`：按表头关键词找本步的表格（复用 attachment_rows 验证过的路子），
+  读数据行数与表尾 ¥ 合计，跟 info 期望值比对；details/payment/subsidy 三步覆盖，basic/attachments 跳过
+  （附件步本来就有 attachment_rows 的行数告警）
+- 期望值来源：travel details = 三类明细行数与金额之和；normal details = 1 行 + total_amount；
+  payment = 笔数与金额之和；subsidy 只比行数（合计由页面按标准算，不拿 info 去比）
+- **读不到表格不装成功**：返回 readable=False + 「本步没有核对过，别当成已核实」的 warning；
+  静默通过正是上一轮附件假成功的成因
+- `browser_fill` 每步返回 self_check（原始读数），把 notes 并进 warnings；
+- `tests/test_self_check.py` 六例：行数与合计都匹配、行数+合计不一致、支付匹配、找不到表、
+  无表格步骤跳过、补助只比行数
+
+还没做的：自检的选择器只在 FakePage 上验证过，**真实页面待联机确认**（尤其是 payment/subsidy 两步的表头关键词），
+与第 4 条（乘车日期选择器）一起，等下一次联机填报时一并确认。

docs/MCP.md

@@ -392,6 +392,9 @@ cases/<case_id>/
     提交前会量这一步要 POST 的自由文本宽度（汉字算 2、字母数字算 1，上限 100，
     普通报销说明 40）：财务后端对超宽只回 500、页面上任何提示都没有，弹窗停在原地
     看着像"按钮没点着"，所以宁可在这里拦下来，失败时 too_long 会列出超长字段。
+    
+    每步 POST 完会做一次**落库后自检**：回读页面上的数据行数与表尾合计，跟 info 的期望值比对，
+    不一致就放进 warnings（返回里的 self_check 是原始读数）；读不到表格时也会明确说「本步没有核对过」。
 
 | 参数 | 类型 | 必填 | 默认 | 说明 |
 |------|------|------|------|------|

src/infra/browser/base.py

@@ -677,6 +677,149 @@ class BaseBot:
         self.confirm_attachment()
         return text
 
+    # ---------------------------------------------------------------
+    # 落库后自检：读回页面上的行数与合计，跟 info 的期望值比对
+    # ---------------------------------------------------------------
+    # 按表头关键词找表，跟 attachment_rows 同一条路子（那条路联机验证过）。
+    # 只覆盖有表格的步骤；读不出来就如实说"没核对过"，绝不静默当成功。
+    STEP_TABLE_HINTS: dict[str, tuple[str, ...]] = {
+        "details": ("费用类别", "出行方式", "经济事项", "报销金额", "单据数"),
+        "payment": ("支付方式", "支付金额", "开户行"),
+        "subsidy": ("补助", "伙食"),
+    }
+    # 行内既不是表头也不是金额的文字，出现在这些列里说明是表头或空状态行
+    HEADER_WORDS = (
+        "序号",
+        "费用类别",
+        "出行方式",
+        "经济事项",
+        "支付方式",
+        "补助",
+        "附件文件名",
+        "没有找到匹配",
+        "暂无数据",
+    )
+
+    def _find_step_table(self, step: str) -> Any:
+        hints = self.STEP_TABLE_HINTS.get(step, ())
+        if not hints:
+            return None
+        try:
+            tables = self.page.query_selector_all("table")
+        except Exception:
+            return None
+        for table in tables:
+            try:
+                head = " ".join((table.inner_text() or "").split())
+            except Exception:
+                continue
+            if head and any(hint in head for hint in hints):
+                return table
+        return None
+
+    def _read_step_table(self, step: str) -> dict[str, Any]:
+        """读本步表格：数据行数与表尾 ¥ 合计；读不出就 readable=False。"""
+        table = self._find_step_table(step)
+        if table is None:
+            return {"readable": False, "reason": "没找到本步的表格"}
+        rows = 0
+        total = 0.0
+        has_amount = False
+        try:
+            trs = table.query_selector_all("tr")
+        except Exception:
+            return {"readable": False, "reason": "表格行读不出来"}
+        for tr in trs:
+            try:
+                text = " ".join((tr.inner_text() or "").split())
+            except Exception:
+                continue
+            if not text:
+                continue
+            if any(word in text for word in self.HEADER_WORDS):
+                continue
+            amounts = re.findall(r"[¥￥]\s*([\d,]+(?:\.\d+)?)", text)
+            if amounts:
+                for raw in amounts:
+                    total += float(raw.replace(",", ""))
+                has_amount = True
+                continue
+            if len(tr.query_selector_all("td")) >= 2:
+                rows += 1
+            elif len(text.split()) >= 3:
+                rows += 1
+        return {"readable": True, "rows": rows, "amount": round(total, 2) if has_amount else None}
+
+    @staticmethod
+    def _row_amount(row: dict[str, Any]) -> float:
+        """info 行里的金额字段名各版本不完全统一，按常见键依次取。"""
+        for key in ("amount", "total_amount", "fee", "money", "报销金额", "费用金额", "补助金额"):
+            raw = row.get(key)
+            if raw in (None, ""):
+                continue
+            try:
+                return float(str(raw).replace(",", ""))
+            except (TypeError, ValueError):
+                continue
+        return 0.0
+
+    def self_check_step(self, kind: str, step: str, info: dict[str, Any]) -> dict[str, Any]:
+        """填报后回读页面：行数/合计与 info 比对，差值变成 warning。
+
+        背景：add_travel_items 点完确定不核对表尾合计，那次 500 看起来像"脚本没点着按钮"。
+        现在每步 POST 完都回读一次；读不到表格也要报"没核对过"，不许装成核实了。
+        """
+        if step not in self.STEP_TABLE_HINTS:
+            return {"step": step, "skipped": "这一步没有可核对的表格"}
+        expected_rows = 0
+        expected_amount: float | None = None
+        if step == "details" and kind == "travel":
+            details = info.get("reimbursement_details", {}) or {}
+            rows: list[dict[str, Any]] = []
+            for key in ("transport_fee", "hotel_fee", "conference_fee"):
+                value = details.get(key) or []
+                if isinstance(value, list):
+                    rows.extend(r for r in value if isinstance(r, dict))
+            expected_rows = len(rows)
+            expected_amount = round(sum(self._row_amount(r) for r in rows), 2)
+        elif step == "details":
+            details = info.get("reimbursement_details", {}) or {}
+            expected_rows = 1
+            try:
+                expected_amount = round(float(details.get("total_amount", 0) or 0), 2)
+            except (TypeError, ValueError):
+                expected_amount = None
+        elif step == "payment":
+            pays = [r for r in (info.get("payment_methods") or []) if isinstance(r, dict)]
+            expected_rows = len(pays)
+            expected_amount = round(sum(self._row_amount(r) for r in pays), 2)
+        elif step == "subsidy":
+            subs = [r for r in (info.get("subsidy_list") or []) if isinstance(r, dict)]
+            expected_rows = len(subs)
+            expected_amount = None  # 补助合计由页面按标准算，不拿 info 去比
+        actual = self._read_step_table(step)
+        if not actual.get("readable"):
+            note = f"{step} 落库自检读不到页面表格（{actual.get('reason')}）：本步没有核对过，别当成已核实"
+            return {"step": step, "readable": False, "expected_rows": expected_rows, "notes": [note]}
+        diffs: list[str] = []
+        if actual["rows"] != expected_rows:
+            diffs.append(f"行数 页面 {actual['rows']} / info {expected_rows}")
+        if expected_amount is not None and actual.get("amount") is not None:
+            if abs(actual["amount"] - expected_amount) > 0.01:
+                diffs.append(f"合计 页面 {actual['amount']:.2f} / info {expected_amount:.2f}")
+        notes = []
+        if diffs:
+            notes.append(f"{step} 落库自检不一致（{'; '.join(diffs)}）：先 browser_page_state 看截图再决定是否重填")
+        return {
+            "step": step,
+            "readable": True,
+            "rows": actual["rows"],
+            "amount": actual.get("amount"),
+            "expected_rows": expected_rows,
+            "expected_amount": expected_amount,
+            "notes": notes,
+        }
+
     def attachment_rows(self) -> int:
         """附件清单里已有的附件条数。
 

src/mcp/tools/browser.py

@@ -303,6 +303,9 @@ def browser_fill(case_id: str, step: str) -> dict[str, Any]:
     提交前会量这一步要 POST 的自由文本宽度（汉字算 2、字母数字算 1，上限 100，
     普通报销说明 40）：财务后端对超宽只回 500、页面上任何提示都没有，弹窗停在原地
     看着像"按钮没点着"，所以宁可在这里拦下来，失败时 too_long 会列出超长字段。
+
+    每步 POST 完会做一次**落库后自检**：回读页面上的数据行数与表尾合计，跟 info 的期望值比对，
+    不一致就放进 warnings（返回里的 self_check 是原始读数）；读不到表格时也会明确说「本步没有核对过」。
     """
     d = _case_dir(case_id)
     ctx = _require_open(case_id)
@@ -348,7 +351,14 @@ def browser_fill(case_id: str, step: str) -> dict[str, Any]:
     def _fill() -> dict[str, Any]:
         bot = ctx["bot"]
         filled = _apply_step(bot, kind, step, info)
-        return {"filled": filled, "screenshot": bot._screenshot(f"{step}_done"), "url": bot.page.url}
+        # 落库后自检：读回页面行数与合计（附件步另有 attachment_rows 的行数告警）
+        check = bot.self_check_step(kind, step, info)
+        return {
+            "filled": filled,
+            "screenshot": bot._screenshot(f"{step}_done"),
+            "url": bot.page.url,
+            "self_check": check,
+        }
 
     try:
         result = actor.run_browser(_fill)
@@ -369,6 +379,9 @@ def browser_fill(case_id: str, step: str) -> dict[str, Any]:
             step=step,
             screenshot=shot,
         )
+    self_check = result.get("self_check") or {}
+    if isinstance(self_check, dict):
+        notes.extend(str(n) for n in self_check.get("notes") or [])
     actor.mark_step(case_id, step)
     actor.touch(d, str(result.get("url", "")))
     done_after = list(ctx.get("steps", []))
@@ -381,6 +394,7 @@ def browser_fill(case_id: str, step: str) -> dict[str, Any]:
             screenshot=result.get("screenshot", ""),
             done=done_after,
             remaining=remaining,
+            self_check=self_check,
             warnings=notes,
             next_action=(
                 f"mcp__arb__browser_fill(step='{remaining[0]}')" if remaining else "mcp__arb__browser_page_state"

tests/test_mcp_browser.py

@@ -245,6 +245,11 @@ FakeBot._validation_hint = BaseBot._validation_hint  # type: ignore[method-assig
 FakeBot.confirm_attachment = BaseBot.confirm_attachment  # type: ignore[method-assign]
 FakeBot.upload_attachment = BaseBot.upload_attachment  # type: ignore[method-assign]
 FakeBot.attachment_rows = BaseBot.attachment_rows  # type: ignore[method-assign]
+FakeBot.STEP_TABLE_HINTS = BaseBot.STEP_TABLE_HINTS
+FakeBot._row_amount = staticmethod(BaseBot._row_amount)  # type: ignore[assignment]
+FakeBot.self_check_step = BaseBot.self_check_step  # type: ignore[method-assign]
+FakeBot._read_step_table = BaseBot._read_step_table  # type: ignore[method-assign]
+FakeBot._find_step_table = BaseBot._find_step_table  # type: ignore[method-assign]
 FakeBot.row_is_ready = BaseBot.row_is_ready  # type: ignore[method-assign]
 FakeBot._wait_any = BaseBot._wait_any  # type: ignore[method-assign]
 FakeBot.row_cells = BaseBot.row_cells  # type: ignore[method-assign]
@@ -416,8 +421,10 @@ def test_fill_walks_steps_and_reports_remaining(env: dict[str, Any]) -> None:
     assert first["done"] == ["basic"]
     assert first["remaining"] == ["details", "payment", "attachments"]
     second = browser.browser_fill(env["id"], "details")
+    assert second["ok"] is True, second
     assert second["filled"] == {"单据数": 3, "总金额": 464.0}
     third = browser.browser_fill(env["id"], "payment")
+    assert third["ok"] is True, third
     assert third["filled"] == {"支付笔数": 2}
     assert third["remaining"] == ["attachments"]
     state = browser.browser_page_state(env["id"])

tests/test_self_check.py

@@ -0,0 +1,91 @@
+"""落库后自检：读回页面行数/合计与 info 比对，差值给 warning；读不到就如实说没核对。"""
+
+from typing import Any
+
+from src.infra.browser.base import BaseBot
+
+
+class _Row:
+    def __init__(self, text: str, cells: int = 4) -> None:
+        self._text = text
+        self._cells = cells
+
+    def inner_text(self) -> str:
+        return self._text
+
+    def query_selector_all(self, selector: str) -> list[Any]:
+        return [object()] * self._cells if selector == "td" else []
+
+
+class _Table:
+    def __init__(self, header: str, rows: list[str]) -> None:
+        self._header = header
+        self._rows = [_Row(r) for r in rows]
+
+    def inner_text(self) -> str:
+        return self._header
+
+    def query_selector_all(self, selector: str) -> list[Any]:
+        return list(self._rows) if selector == "tr" else []
+
+
+class _Page:
+    def __init__(self, tables: list[_Table]) -> None:
+        self.tables = tables
+
+    def query_selector_all(self, selector: str) -> list[Any]:
+        return list(self.tables) if selector == "table" else []
+
+
+class _Bot(BaseBot):
+    def __init__(self, tables: list[_Table]) -> None:
+        self.page = _Page(tables)  # type: ignore[assignment]
+
+
+def test_travel_details_match_has_no_warning() -> None:
+    bot = _Bot([_Table("费用类别 出行方式 报销金额", ["交通费 火车 233", "¥233.00"])])
+    info = {"reimbursement_details": {"transport_fee": [{"amount": "233.00"}]}}
+    check = bot.self_check_step("travel", "details", info)
+    assert check["readable"] is True
+    assert check["rows"] == 1 and check["amount"] == 233.0
+    assert check["notes"] == []
+
+
+def test_travel_details_mismatch_reports_rows_and_total() -> None:
+    bot = _Bot([_Table("费用类别 出行方式 报销金额", ["交通费 火车 233", "¥233.00"])])
+    info = {
+        "reimbursement_details": {
+            "transport_fee": [{"amount": "233.00"}, {"amount": "233.00"}],
+        }
+    }
+    notes = bot.self_check_step("travel", "details", info)["notes"]
+    assert notes and "行数 页面 1 / info 2" in notes[0]
+    assert "合计 页面 233.00 / info 466.00" in notes[0]
+
+
+def test_payment_matches() -> None:
+    bot = _Bot([_Table("支付方式 支付金额 开户行", ["工行公务卡 231.00 工商银行", "¥231.00"])])
+    info = {"payment_methods": [{"amount": "231.00"}]}
+    check = bot.self_check_step("travel", "payment", info)
+    assert check["readable"] is True and check["notes"] == []
+
+
+def test_missing_table_says_not_checked() -> None:
+    bot = _Bot([])
+    check = bot.self_check_step("travel", "payment", {"payment_methods": []})
+    assert check["readable"] is False
+    assert "没有核对过" in check["notes"][0]
+
+
+def test_steps_without_table_are_skipped() -> None:
+    bot = _Bot([])
+    assert "skipped" in bot.self_check_step("travel", "basic", {})
+    assert "skipped" in bot.self_check_step("travel", "attachments", {})
+
+
+def test_subsidy_compares_rows_only() -> None:
+    bot = _Bot([_Table("补助清单 伙食补助", ["张国庆 540"])])
+    info = {"subsidy_list": [{"person_name": "张国庆"}]}
+    check = bot.self_check_step("travel", "subsidy", info)
+    assert check["readable"] is True and check["notes"] == []
+    assert check["expected_amount"] is None