diff --git a/docs/reviews/assets/biz-compare-desktop.png b/docs/reviews/assets/biz-compare-desktop.png
new file mode 100644
index 0000000..17926f3
Binary files /dev/null and b/docs/reviews/assets/biz-compare-desktop.png differ
diff --git a/docs/reviews/assets/biz-compare-error.png b/docs/reviews/assets/biz-compare-error.png
new file mode 100644
index 0000000..e032b23
Binary files /dev/null and b/docs/reviews/assets/biz-compare-error.png differ
diff --git a/docs/reviews/assets/biz-compare-legacy.png b/docs/reviews/assets/biz-compare-legacy.png
new file mode 100644
index 0000000..a374dee
Binary files /dev/null and b/docs/reviews/assets/biz-compare-legacy.png differ
diff --git a/docs/reviews/assets/biz-compare-memory-benchmark.json b/docs/reviews/assets/biz-compare-memory-benchmark.json
new file mode 100644
index 0000000..70290b7
--- /dev/null
+++ b/docs/reviews/assets/biz-compare-memory-benchmark.json
@@ -0,0 +1,29 @@
+{
+  "date": "2026-10-10",
+  "baseline_ref": "afa8e06b2de64100f963ae30b3babfcb5d7009f4",
+  "python": "3.14.3",
+  "dataset": {
+    "rows_per_side": 1000000,
+    "fields": ["id", "neighbor", "state", "afi"],
+    "strata": 50,
+    "unchanged_limit": 5000,
+    "compact_unchanged": true,
+    "interface_transforms": false,
+    "all_rows_unchanged": true
+  },
+  "measurement": "tracemalloc peak allocated during compare_rows; excludes input allocations, DB loading and process overhead",
+  "results": [
+    {
+      "engine": "baseline",
+      "peak_engine_mib": 1144.43,
+      "seconds_with_tracemalloc": 92.4
+    },
+    {
+      "engine": "optimized",
+      "peak_engine_mib": 420.49,
+      "seconds_with_tracemalloc": 63.681
+    }
+  ],
+  "results_and_sample_order_identical": true,
+  "test_result": "1 passed in 157.88s"
+}
diff --git a/docs/reviews/assets/biz-compare-narrow.png b/docs/reviews/assets/biz-compare-narrow.png
new file mode 100644
index 0000000..f6f316a
Binary files /dev/null and b/docs/reviews/assets/biz-compare-narrow.png differ
diff --git a/docs/reviews/biz-compare-review-2026-10-10.md b/docs/reviews/biz-compare-review-2026-10-10.md
new file mode 100644
index 0000000..f3150e1
--- /dev/null
+++ b/docs/reviews/biz-compare-review-2026-10-10.md
@@ -0,0 +1,120 @@
+# 业务对比检查与优化记录
+
+检查日期：2026-10-10。范围：网络管理 → 割接对比 → 业务对比，包括前端任务/批次/结果界面、Python 与 PostgreSQL 对比路径、结果分页与搜索、短期列表缓存。
+
+本次直接修改项目代码并添加回归测试，没有连接设备执行采集，也没有修改生产数据库。浏览器截图与交互测试使用固定 API 响应。
+
+## 已修复的问题
+
+### 1. 搜索会改变业务判定（P1）
+
+原实现分别按搜索条件过滤割接前、后的原表，再重新对比。例如 `state=up → down`，搜索 `up` 后只剩前侧，原本的“变化”就可能变成“删除”。重复身份键的业务还可能在筛选后错位配对。
+
+现在：已存储的判定直接在结果表分页筛选；需要回查未保存全量明细的成功行时，先找到候选身份键，再补齐两侧的完整分组，按原有配对规则对比，最后筛选。达到读取上限的分组整体跳过，并返回截断标记，避免产生虚假的新增或删除。
+
+### 2. PostgreSQL 搜索漏用部分模板过滤条件（P1）
+
+原搜索只下推 SQL 支持的过滤规则，对 `age_timer` 等不支持的规则直接忽略，可能将静态 ARP 等不属于模板范围的数据纳入结果。现在不支持下推的规则走流式读取并在 Python 中完整过滤。
+
+### 3. 快速切换时旧请求覆盖当前任务（P1）
+
+前后采集任务的批次列表、打开任务、加载历史结果以及轮询缺少完整的过期保护。现在对批次请求使用清理标记，对任务与结果使用请求序号；关闭弹窗、切换路由或打开其他任务后，旧响应不能覆盖当前界面。
+
+### 4. 结果加载错误容易被误读（P2）
+
+原查询失败时保留上一组明细，界面已经显示新筛选条件；切到未完成的表后，加载状态也可能无法结束。现在加载和失败都有独立状态，错误时清除旧明细并提供重试入口；切换表/页签后清理加载状态。
+
+### 5. LLDP 搜索漏查（P2）
+
+LLDP 数据存于 `biz_state_lldp_neighbor`，原搜索只读 `biz_state_metric_row`。现在原表回查与精简结果的字段搜索都支持 LLDP 表。
+
+### 6. 其他已修复边界
+
+- 恰好命中读取上限时，原非 PostgreSQL 路径错误地报告“已截断”；现在多读一条判断。
+- 无效或有歧义的表标识可能落到其他表；现在返回 `sheet_not_found`。
+- `%`、`_` 等搜索字符按字面值匹配，避免不同查询路径含义不一致。
+- 关闭最后一项检查后，空数组被解释为“启用全部”；现在阻止关闭最后一项并提示至少保留一项。
+- 两个 HeroUI Input 使用了 `isDisabled`，实际需要 `disabled`；过滤值与容差输入的禁用状态已修正。
+- 列表缓存缺少并发合并，慢请求可以覆盖强制刷新结果或在失效后重新填充缓存；现在共享请求、保护失效状态并抑制过期的后台通知。
+- 清理中英文文案中的重复键，解决相关类型检查错误。
+
+## 性能改动与实际收益范围
+
+- 合并两套轮询，下一轮在本轮请求完成后启动。模拟浏览器测试中，运行任务在 5.2 秒内请求历史列表和详情各两次，没有重复轮询，也不会因慢请求持续叠加。
+- 已保存的失败、新增、变化等结果在数据库中筛选并分页，避免每次翻页重新加载和对比原始批次。回归测试明确验证这一查询路径不调用 `compare_rows`。
+- 同一前后采集任务的批次请求共用短期缓存，列表后台刷新合并同一请求；6 项缓存测试覆盖并发、强制刷新、失败重试与失效竞态。
+- 模板查找使用 Map，避免列表搜索与绘制逐项线性查找模板。
+- 修改筛选后先回到第一页，避免先请求旧页码再重复请求第一页。
+
+这些改动减少了重复工作。以下后续基准使用合成数据，没有测量生产环境延迟。
+
+## 后续优化（2026-10-10）
+
+### 历史执行配置与精简行还原
+
+新运行在创建时将模板、启用的检查项、字段规则/过滤、接口归一化、端口映射、成功行保存策略和名称存入 `summary_json.config_snapshot`。后台执行、成功行回查、精简行还原以及导出共用该配置；排队之后修改配置，甚至删除原模板/映射，也不会改变这一运行使用的配置。无需数据库结构迁移。
+
+修复了精简存储明细还原时将 `mapped_before` 直接设为原始 `before` 的问题。现在前侧依次归一化和映射，后侧归一化，保持与完整结果的显示及 CSV 一致。ZIP 新增 `config_snapshot.json`，便于核对当时配置；详情接口只返回快照模板与版本，不将完整端口映射加入轮询响应。
+
+页面显示“执行配置已保存”；旧记录显示“旧记录：未保存执行配置”。旧运行无法可靠补造当时的配置，仍保留兼容查询，成功行回查及精简行还原会使用当前配置。快照也不保存原始采集行的副本，若原始批次数据被清理，依赖行 ID 的还原仍会缺少内容。
+
+### Python 大表内存
+
+成功样本按“层内序号、层首次出现顺序”排序，只保留样本上限内的候选和分层状态，避免先积累全部成功行。保留原分层轮询顺序；未截断的小表仍保持采集顺序。不需要接口变换时复用输入行；有变换时创建新行，不修改输入数据。稳定身份键顺序直接使用分组字典，省去再次计算全表键、额外键列表与集合。
+
+每侧 1,000,000 行、50 个邻居分层、全部成功、保存 5,000 条精简样本的测试：优化前引擎新增分配峰值 **1144.43 MiB**，优化后 **420.49 MiB**，下降约 **63.3%**。完整结果与样本顺序的 SHA-256 一致。开启 `tracemalloc` 时耗时为 92.400 秒和 63.681 秒；该计时含内存追踪开销，不代表生产耗时。输入行在追踪开始前创建，峰值不包含输入行、数据库加载或 Python 进程的总内存。
+
+基线为 `afa8e06b2de64100f963ae30b3babfcb5d7009f4` 中的对比引擎。原始指标见 [百万行测试结果](assets/biz-compare-memory-benchmark.json)。身份键分组仍需要随原表行数增长的内存；这次不是完全流式对比。
+
+### 大表导出
+
+导出改为按 `(seq, id)` 游标分批查询、分批还原精简行、逐行将 UTF-8 CSV 写入 ZIP，避免将全部明细与完整 CSV 同时保存在内存。相同 `seq` 的多条结果按 `id` 连续读取，不漏掉跨批次的同行序号记录。保留 BOM、字段顺序及逗号/引号/换行转义。最终压缩包仍由现有接口在内存中返回，压缩文件本身的大小仍计入内存。
+
+## UI 改动
+
+结果区分为当前检查项/全表通过率、结果类型、搜索与辅助操作三层。增加明确的分页范围、清空筛选、加载状态、错误重试和搜索无结果提示。筛选按钮增加按下状态，字段筛选增加展开状态与键盘焦点提示。后续增加了执行配置保存状态及旧记录限制提示。
+
+窄屏检查项导航改为紧凑的横向排列，避免一个检查项占据整行的大块区域。浏览器检查覆盖 1440px、768px、390px；筛选栏没有横向溢出。结果表保留必要的横向滚动。
+
+- [桌面结果区](assets/biz-compare-desktop.png)
+- [768px 结果区](assets/biz-compare-narrow.png)
+- [加载失败与重试](assets/biz-compare-error.png)
+- [旧记录执行配置提示](assets/biz-compare-legacy.png)
+
+## 验证
+
+- Python：87 项通过，1 项百万行基准默认跳过；单独启用百万行基准后通过。新增测试覆盖搜索、历史配置快照、有界抽样与游标导出。
+- 缓存：6 项 Node 测试通过。
+- 浏览器：10 个场景通过，包括配置保存状态、旧记录提示、结果布局、错误清除旧明细、重试、清空筛选、窄屏布局、旧任务响应保护、单套轮询、无运行时异常。
+- Vite 生产打包成功。
+- 本次业务对比页面的类型错误已经修复。全项目 `tsc -b` 仍被 `BizStatePage.tsx`、`BizMigrationPage.tsx` 中原有的类型/未使用变量错误阻塞。
+- 页面 ESLint 原有 34 个错误，主要是 `any` 与现有 React Hook 写法；本次未增加错误。结果请求依赖警告已修正。缓存模块没有 lint 错误。
+- PostgreSQL 路径验证了 SQL 编译与规则降级分支，没有在真实 PostgreSQL 数据库上执行查询或测量执行计划。
+
+可重复执行的检查：
+
+```powershell
+# 项目根目录
+.venv\Scripts\python.exe -m pytest tests/test_biz_state_compare_export.py tests/test_biz_state_compare_snapshot.py tests/test_biz_state_compare_memory.py tests/test_biz_state_compare_search.py tests/test_biz_state_compare.py tests/test_biz_state_compare_sql.py tests/test_biz_state_compare_recovery.py -q
+
+# 可选：百万行基准，耗时约数分钟，内存占用较大
+$env:NETX_BENCH_ROWS = '1000000'
+$env:NETX_BENCH_REF = 'afa8e06b2de64100f963ae30b3babfcb5d7009f4'
+.venv\Scripts\python.exe -m pytest tests/test_biz_state_compare_memory.py::test_memory_benchmark_against_git_baseline -q -s
+Remove-Item Env:NETX_BENCH_ROWS, Env:NETX_BENCH_REF
+
+# web 目录
+node --experimental-strip-types --test tests/cutoverDataCache.test.mjs
+node node_modules/vite/bin/vite.js build
+node node_modules/typescript/bin/tsc -b
+```
+
+浏览器脚本为 `web/tests/bizCompare.browser.mjs`。先启动本地 Vite 服务，再设置 `NETX_PLAYWRIGHT_MODULE` 为可用的 Playwright 包路径；`NETX_TEST_URL` 默认是 `http://127.0.0.1:5179`，浏览器默认使用已安装的 Chrome。所有 `/v1` 请求由脚本提供固定响应，不会访问实际设备或数据库。
+
+## 后续值得优化的部分
+
+1. **Python 大表分组。** 有界抽样和冗余复制优化已完成，百万行基准下降约 63% 的引擎新增分配峰值。分组与输入仍随原表规模增长；进一步降低内存可研究排序流式配对或分区执行，并保持重复键配对语义。
+2. **PostgreSQL 搜索与深分页。** JSON 字段包含搜索和精确总数统计仍可能扫描大量结果。需要用真实数据查看 `EXPLAIN (ANALYZE, BUFFERS)`，再决定常用字段索引、搜索文本索引或游标分页；本次没有盲目添加生产索引。
+3. **历史配置与原始数据保留。** 新运行的执行配置快照已完成。旧运行仍无法还原未记录的当时配置；精简结果还原依赖原批次行，后续可为批次清理增加引用保护或保留策略。
+4. **重复键的业务语义。** 引擎按采集顺序配对同键行。如果业务应按集合判断，CLI 行顺序改变可能产生多条“不一致”。当前保留原有语义；可优先完善身份键，或另行实现并测试 Python/SQL 一致的集合匹配。
+5. **原表回查的数量上限。** 每侧候选命中最多 2000 条，完整分组补齐每侧预算 10000 条，最终展示最多 200 对；这不是全表匹配总数。界面会提示截断，大范围检索需要收窄条件。
diff --git a/netx_api/biz_state/compare_engine.py b/netx_api/biz_state/compare_engine.py
index 3871cc4..5957d79 100644
--- a/netx_api/biz_state/compare_engine.py
+++ b/netx_api/biz_state/compare_engine.py
@@ -2,8 +2,10 @@
 
 from __future__ import annotations
 
-from collections import defaultdict, deque
-from typing import Any, Mapping, Sequence
+from collections import defaultdict
+from heapq import heappush, heapreplace
+from itertools import chain
+from typing import Any, Iterable, Mapping, Sequence
 
 from .compare_rules import field_rule_map, values_equal, explain_diff
 from .iface_normalize import (
@@ -28,36 +30,52 @@ def stratum_key(row: Mapping[str, Any] | None) -> str:
     return "|".join(parts) if parts else "_"
 
 
-def stratify_take(items: list[Any], limit: int, *, key_fn) -> list[Any]:
-    """Round-robin across strata so one neighbor/direction cannot consume the whole sample."""
-    lim = max(0, int(limit))
-    if lim <= 0 or not items:
+class _StratifiedSample:
+    """Keep the first N round-robin ranks in O(N) space, even with many strata."""
+
+    def __init__(self, limit: int) -> None:
+        self.limit = max(0, int(limit))
+        self.count = 0
+        self.strata: dict[str, tuple[int, int]] = {}
+        self.heap: list[tuple[int, int, int, Any]] = []
+
+    def add(self, item: Any, key: str) -> None:
+        serial = self.count
+        self.count += 1
+        if not self.limit:
+            return
+        state = self.strata.get(key)
+        if state is None:
+            # Later strata cannot beat the first item of N earlier strata.
+            if len(self.strata) >= self.limit:
+                return
+            index, round_n = len(self.strata), 0
+        else:
+            index, round_n = state
+        self.strata[key] = (index, round_n + 1)
+        entry = (-round_n, -index, serial, item)
+        if len(self.heap) < self.limit:
+            heappush(self.heap, entry)
+        elif entry[:2] > self.heap[0][:2]:
+            heapreplace(self.heap, entry)
+
+    def picked(self) -> list[Any]:
+        # The legacy helper preserves encounter order when no truncation occurs.
+        if self.count <= self.limit:
+            entries = sorted(self.heap, key=lambda e: e[2])
+        else:
+            entries = sorted(self.heap, key=lambda e: (-e[0], -e[1]))
+        return [entry[3] for entry in entries]
+
+
+def stratify_take(items: Iterable[Any], limit: int, *, key_fn) -> list[Any]:
+    """Round-robin across strata without retaining every candidate."""
+    sample = _StratifiedSample(limit)
+    if not sample.limit:
         return []
-    if len(items) <= lim:
-        return list(items)
-    buckets: dict[str, deque[Any]] = defaultdict(deque)
-    order: list[str] = []
-    for it in items:
-        sk = str(key_fn(it) or "_")
-        if sk not in buckets:
-            order.append(sk)
-        buckets[sk].append(it)
-    out: list[Any] = []
-    while len(out) < lim and buckets:
-        drained: list[str] = []
-        for sk in order:
-            q = buckets.get(sk)
-            if not q:
-                drained.append(sk)
-                continue
-            out.append(q.popleft())
-            if len(out) >= lim:
-                break
-        for sk in drained:
-            buckets.pop(sk, None)
-            if sk in order:
-                order = [x for x in order if x != sk]
-    return out
+    for item in items:
+        sample.add(item, str(key_fn(item) or "_"))
+    return sample.picked()
 
 
 def apply_port_map(
@@ -198,10 +216,10 @@ def compare_rows(
 
     before_norm = apply_iface_normalize_rows(
         before_rows, iface_fields=iface_list, rules=norm_rules
-    )
+    ) if norm_rules and iface_list else before_rows
     after_norm = apply_iface_normalize_rows(
         after_rows, iface_fields=iface_list, rules=norm_rules
-    )
+    ) if norm_rules and iface_list else after_rows
 
     ignore_ports = False
     # No map → optionally ignore port renames by dropping iface from match key.
@@ -216,9 +234,16 @@ def compare_rows(
             match_keys = list(key_fields)
         else:
             # Auto heuristic (legacy default)
-            before_c = [row_key(r, candidate) for r in before_norm]
-            after_c = [row_key(r, candidate) for r in after_norm]
-            if len(before_c) == len(set(before_c)) and len(after_c) == len(set(after_c)):
+            def unique_keys(rows: list[dict[str, Any]]) -> bool:
+                seen: set[tuple[str, ...]] = set()
+                for row in rows:
+                    key = row_key(row, candidate)
+                    if key in seen:
+                        return False
+                    seen.add(key)
+                return True
+
+            if unique_keys(before_norm) and unique_keys(after_norm):
                 match_keys = candidate
                 ignore_ports = True
             else:
@@ -226,15 +251,12 @@ def compare_rows(
     else:
         match_keys = list(key_fields)
 
-    before_mapped: list[dict[str, Any]] = [
-        apply_port_map(r, iface_fields=iface_list, port_map=pmap) for r in before_norm
-    ]
-
     before_groups: dict[tuple[str, ...], list[tuple[dict[str, Any], dict[str, Any]]]] = (
         defaultdict(list)
     )
     after_groups: dict[tuple[str, ...], list[dict[str, Any]]] = defaultdict(list)
-    for orig, mapped in zip(before_rows, before_mapped):
+    for orig, norm in zip(before_rows, before_norm):
+        mapped = apply_port_map(norm, iface_fields=iface_list, port_map=pmap) if iface_list else norm
         before_groups[row_key(mapped, match_keys)].append((orig, mapped))
     for r in after_norm:
         after_groups[row_key(r, match_keys)].append(r)
@@ -246,6 +268,7 @@ def compare_rows(
     multi_before_keys: list[tuple[str, ...]] = []
     multi_after_keys: list[tuple[str, ...]] = []
     unchanged_candidates: list[tuple[dict[str, Any], dict[str, Any], dict[str, Any]]] = []
+    sample = _StratifiedSample(limit_n) if include_unchanged and limit_n is not None else None
 
     def _key_obj(row: dict[str, Any]) -> dict[str, Any]:
         return {f: row.get(f, "") for f in key_fields}
@@ -294,18 +317,7 @@ def compare_rows(
             )
 
     # Stable key order: before encounter order, then after-only keys
-    seen_keys: set[tuple[str, ...]] = set()
-    ordered_keys: list[tuple[str, ...]] = []
-    for orig, mapped in zip(before_rows, before_mapped):
-        k = row_key(mapped, match_keys)
-        if k not in seen_keys:
-            seen_keys.add(k)
-            ordered_keys.append(k)
-    for r in after_norm:
-        k = row_key(r, match_keys)
-        if k not in seen_keys:
-            seen_keys.add(k)
-            ordered_keys.append(k)
+    ordered_keys = chain(before_groups, (k for k in after_groups if k not in before_groups))
     for k in ordered_keys:
         b_list = before_groups.get(k) or []
         a_list = after_groups.get(k) or []
@@ -377,17 +389,14 @@ def compare_rows(
             else:
                 unchanged += 1
                 if include_unchanged:
-                    unchanged_candidates.append((orig, mapped, after))
+                    if sample is not None:
+                        if sample.limit:
+                            sample.add((orig, mapped, after), stratum_key(mapped))
+                    else:
+                        unchanged_candidates.append((orig, mapped, after))
 
-    if include_unchanged and unchanged_candidates:
-        if limit_n is None:
-            picked = unchanged_candidates
-        else:
-            picked = stratify_take(
-                unchanged_candidates,
-                limit_n,
-                key_fn=lambda t: stratum_key(t[1]),
-            )
+    if include_unchanged:
+        picked = sample.picked() if sample is not None else unchanged_candidates
         for orig, mapped, after_row in picked:
             _append_unchanged_diff(orig, mapped, after_row)
 
diff --git a/netx_api/biz_state/compare_service.py b/netx_api/biz_state/compare_service.py
index e11b551..43e097b 100644
--- a/netx_api/biz_state/compare_service.py
+++ b/netx_api/biz_state/compare_service.py
@@ -8,13 +8,14 @@ import logging
 import threading
 import time
 import zipfile
+from copy import deepcopy
 from datetime import datetime
-from typing import Any, Callable
+from typing import Any, Callable, Iterable, Iterator
 from uuid import uuid4
 
 from fastapi import HTTPException
-from sqlalchemy import and_, or_
-from sqlalchemy.orm import Session
+from sqlalchemy import String, and_, cast, func, or_, tuple_
+from sqlalchemy.orm import Session, aliased
 
 from ..models import (
     BizCompareDiff,
@@ -29,7 +30,7 @@ from ..models import (
     BizStateTask,
 )
 from ..timeutil import utcnow_naive
-from .compare_engine import compare_rows, mapping_stats
+from .compare_engine import apply_port_map, compare_rows, mapping_stats, row_key
 from .compare_rules import (
     ROW_FILTER_PRESETS,
     apply_row_filters,
@@ -193,6 +194,7 @@ _FAIL_COMPACT_MIN = 50_000
 # Live search (kw): load at most this many matching rows per side, return ≤ this many pairs.
 _LIVE_SEARCH_LOAD_CAP = 2_000
 _LIVE_SEARCH_RESULT_CAP = 200
+_LIVE_SEARCH_GROUP_LOAD_CAP = 10_000
 # Success-row persist policy (see resolve_unchanged_policy)
 _STORE_UNCHANGED_MODES = frozenset({"auto", "always", "never", "sample", "keys"})
 _UNCHANGED_FULL_MAX = 20_000
@@ -269,7 +271,7 @@ def _strip_netx(row: Any) -> dict[str, Any]:
     return {k: v for k, v in row.items() if k != "_netx"}
 
 
-def _diff_search_text(d: dict[str, Any]) -> str:
+def _diff_search_text(d: dict[str, Any], *, limit: int | None = _SEARCH_TEXT_MAX) -> str:
     parts = [str(d.get("kind") or "")]
     for key in ("key", "before", "after", "mapped_before", "changes"):
         val = d.get(key)
@@ -278,7 +280,8 @@ def _diff_search_text(d: dict[str, Any]) -> str:
                 parts.append(json.dumps(val, ensure_ascii=False, default=str, separators=(",", ":")))
             except Exception:
                 parts.append(str(val))
-    return " ".join(parts)[:_SEARCH_TEXT_MAX]
+    value = " ".join(parts)
+    return value if limit is None else value[:limit]
 
 
 def _top_changed_fields(diffs: list[dict[str, Any]], *, limit: int = 8) -> list[dict[str, Any]]:
@@ -422,7 +425,10 @@ def _metric_rows_by_ids(db: Session, ids: list[str]) -> dict[str, dict[str, Any]
     return out
 
 
-def _hydrate_diff_rows(db: Session, items: list[dict[str, Any]]) -> list[dict[str, Any]]:
+def _hydrate_diff_rows(
+    db: Session, items: list[dict[str, Any]], *,
+    run: BizCompareRun | None = None, sheet: dict[str, Any] | None = None,
+) -> list[dict[str, Any]]:
     """Fill empty before/after from metric tables when row_ids are present."""
     need: list[str] = []
     for it in items:
@@ -433,16 +439,24 @@ def _hydrate_diff_rows(db: Session, items: list[dict[str, Any]]) -> list[dict[st
     if not need:
         return items
     by_id = _metric_rows_by_ids(db, need)
+    pmap, norm_rules = _run_transform_config(db, run) if run else ({}, [])
+    iface_fields = list((sheet or {}).get("iface_fields") or [])
     for it in items:
         brid = str(it.get("before_row_id") or "")
         arid = str(it.get("after_row_id") or "")
         if not it.get("before") and brid and brid in by_id:
             it["before"] = by_id[brid]
         if not it.get("after") and arid and arid in by_id:
-            it["after"] = by_id[arid]
-        # Success compact: no mapped_before stored — UI falls back to before
+            it["after"] = apply_iface_normalize_rows(
+                [by_id[arid]], iface_fields=iface_fields, rules=norm_rules,
+            )[0]
         if not it.get("mapped_before") and it.get("before"):
-            it["mapped_before"] = dict(it["before"])
+            normalized = apply_iface_normalize_rows(
+                [it["before"]], iface_fields=iface_fields, rules=norm_rules,
+            )[0]
+            it["mapped_before"] = apply_port_map(
+                normalized, iface_fields=iface_fields, port_map=pmap,
+            )
     return items
 
 
@@ -471,6 +485,7 @@ def _filter_inline_diffs(
     *,
     kind: str,
     kw: str,
+    field_q: dict[str, str] | None = None,
 ) -> list[dict[str, Any]]:
     kind_n = (kind or "diff").strip().lower()
     kw_n = (kw or "").strip().lower()
@@ -484,13 +499,62 @@ def _filter_inline_diffs(
         elif kind_n != "all" and dk != kind_n:
             continue
         if kw_n:
-            blob = _diff_search_text(d).lower()
+            blob = _diff_search_text(d, limit=None).lower()
             if kw_n not in blob:
                 continue
+        if field_q and not _diff_matches_search(d, kw="", field_q=field_q):
+            continue
         out.append(d)
     return out
 
 
+def _diff_matches_search(d: dict[str, Any], *, kw: str, field_q: dict[str, str]) -> bool:
+    """Search a completed pair, never filter its two sides before pairing."""
+    if kw and kw.lower() not in _diff_search_text(d, limit=None).lower():
+        return False
+    return all(
+        any(
+            val.lower() in str((d.get(side) or {}).get(name, "")).lower()
+            for side in ("key", "before", "mapped_before", "after")
+        )
+        for name, val in field_q.items()
+    )
+
+
+def _search_stored_diffs(q: Any, *, metric_id: str, kw: str, field_q: dict[str, str]) -> Any:
+    """Filter persisted verdicts in SQL; join source rows for compact payloads."""
+    from ..models import BizStateMetricRow
+
+    sides = [BizCompareDiff.key_json, BizCompareDiff.before_json,
+             BizCompareDiff.mapped_before_json, BizCompareDiff.after_json]
+    sources = []
+    for row_id in (BizCompareDiff.before_row_id, BizCompareDiff.after_row_id):
+        model = BizStateLldpNeighbor if metric_id == "lldp_neighbor" else BizStateMetricRow
+        src = aliased(model)
+        q = q.outerjoin(src, src.id == row_id)
+        sources.append(src)
+    if kw:
+        expressions = [BizCompareDiff.search_text, *[cast(s, String) for s in sides]]
+        for src in sources:
+            if metric_id == "lldp_neighbor":
+                expressions.extend(getattr(src, f) for f in
+                                   ("local_if", "remote_sys", "remote_if", "remote_ip", "protocol"))
+            else:
+                expressions.append(cast(src.data_json, String))
+        q = q.filter(or_(*[func.lower(e).contains(kw.lower(), autoescape=True) for e in expressions]))
+    for name, val in field_q.items():
+        expressions = [s[name].as_string() for s in sides]
+        for src in sources:
+            if metric_id == "lldp_neighbor":
+                if name in ("local_if", "remote_sys", "remote_if", "remote_ip", "protocol"):
+                    expressions.append(getattr(src, name))
+            else:
+                expressions.append(src.data_json[name].as_string())
+        q = q.filter(or_(*[func.lower(cast(e, String)).contains(val.lower(), autoescape=True)
+                           for e in expressions]))
+    return q
+
+
 def _sheet_meta_from_summary(summary: dict[str, Any], run: BizCompareRun, tpl: Any) -> list[dict[str, Any]]:
     sheets = list(summary.get("sheets") or [])
     if sheets:
@@ -2193,6 +2257,18 @@ def _create_running_run(
 ) -> BizCompareRun:
     first_metric = str(sheets_cfg[0].get("metric_id") or "")
     pending_sheets = [_pending_sheet_meta(s) for s in sheets_cfg]
+    mapping = db.get(BizPortMapping, job.mapping_id) if job.mapping_id else None
+    config_snapshot = deepcopy({
+        "version": 1,
+        "template": _template_out(tpl),
+        "sheets": sheets_cfg,
+        "port_map": _port_map_dict(db, str(job.mapping_id or "")),
+        "store_unchanged": normalize_store_unchanged(getattr(job, "store_unchanged", None)),
+        "job_name": job.name or "",
+        "before_task_id": job.before_task_id or "",
+        "after_task_id": job.after_task_id or "",
+        "mapping_name": (mapping.name if mapping else "") or "",
+    })
     run = BizCompareRun(
         id=uuid4().hex,
         job_id=job.id,
@@ -2203,6 +2279,7 @@ def _create_running_run(
         metric_id=first_metric,
         status="running",
         summary_json={
+            "config_snapshot": config_snapshot,
             "progress": {
                 "phase": "queued",
                 "sheet_index": 0,
@@ -2232,6 +2309,26 @@ def _create_running_run(
     return run
 
 
+def _run_config_snapshot(run: BizCompareRun) -> dict[str, Any] | None:
+    snapshot = (run.summary_json or {}).get("config_snapshot")
+    if isinstance(snapshot, dict) and snapshot.get("version") == 1:
+        return snapshot
+    return None
+
+
+def _run_transform_config(
+    db: Session, run: BizCompareRun,
+) -> tuple[dict[str, str], list[dict[str, str]]]:
+    snapshot = _run_config_snapshot(run)
+    if snapshot is not None:
+        return dict(snapshot.get("port_map") or {}), normalize_iface_rules(
+            (snapshot.get("template") or {}).get("iface_normalize_rules") or [],
+        )
+    # Legacy runs did not record these settings; retain their fallback behavior.
+    tpl = db.get(BizCompareTemplate, run.template_id) if run.template_id else None
+    return _port_map_dict(db, str(run.mapping_id or "")), template_iface_normalize(tpl)
+
+
 def _set_run_progress(
     db: Session,
     run: BizCompareRun,
@@ -2309,14 +2406,15 @@ def _execute_compare_into_run(db: Session, run_id: str) -> dict[str, Any]:
     run = db.get(BizCompareRun, run_id)
     if not run:
         raise HTTPException(status_code=404, detail="run_not_found")
+    snapshot = _run_config_snapshot(run)
     j = db.get(BizCompareJob, run.job_id)
     if not j:
         run.status = "failed"
         run.message = "job_not_found"
         db.commit()
         raise HTTPException(status_code=404, detail="job_not_found")
-    tpl = db.get(BizCompareTemplate, run.template_id)
-    if not tpl:
+    tpl = db.get(BizCompareTemplate, run.template_id) if snapshot is None else None
+    if not tpl and snapshot is None:
         run.status = "failed"
         run.message = "template_not_found"
         db.commit()
@@ -2324,25 +2422,25 @@ def _execute_compare_into_run(db: Session, run_id: str) -> dict[str, Any]:
 
     before_batch_id = str(run.before_batch_id or "")
     after_batch_id = str(run.after_batch_id or "")
-    sheets_cfg = _filter_enabled_sheets(
-        template_metrics(tpl), getattr(j, "enabled_sheet_ids", None)
-    )
+    if snapshot is not None:
+        sheets_cfg = deepcopy(snapshot.get("sheets") or [])
+    else:
+        sheets_cfg = _filter_enabled_sheets(template_metrics(tpl), getattr(j, "enabled_sheet_ids", None))
     if not sheets_cfg:
         run.status = "failed"
         run.message = "no_enabled_sheets"
         db.commit()
         raise HTTPException(status_code=400, detail="no_enabled_sheets")
-    sheets_cfg = _order_sheets_small_first(
-        db,
-        sheets_cfg,
-        before_batch_id=before_batch_id,
-        after_batch_id=after_batch_id,
-    )
+    if snapshot is None:
+        sheets_cfg = _order_sheets_small_first(
+            db, sheets_cfg, before_batch_id=before_batch_id, after_batch_id=after_batch_id,
+        )
 
     started_mono = time.monotonic()
-    store_mode = normalize_store_unchanged(getattr(j, "store_unchanged", None))
-    pmap = _port_map_dict(db, run.mapping_id)
-    norm_rules = template_iface_normalize(tpl)
+    store_mode = normalize_store_unchanged(
+        snapshot.get("store_unchanged") if snapshot is not None else getattr(j, "store_unchanged", None),
+    )
+    pmap, norm_rules = _run_transform_config(db, run)
     unchanged_listed_total = 0
     unchanged_truncated_any = False
     unchanged_compact_any = False
@@ -2636,6 +2734,8 @@ def _execute_compare_into_run(db: Session, run_id: str) -> dict[str, Any]:
             },
             "sheets": sheet_metas,
         }
+        if snapshot is not None:
+            summary_payload["config_snapshot"] = snapshot
         run = db.get(BizCompareRun, run_id) or run
         if str(run.status or "") == "cancelled":
             return get_run(db, run.id)
@@ -2842,7 +2942,7 @@ def _csv_cell(v: Any) -> str:
     return s
 
 
-def _sheet_csv(sheet: dict[str, Any]) -> str:
+def _sheet_csv_lines(sheet: dict[str, Any], diffs: Iterable[dict[str, Any]]) -> Iterator[str]:
     keys = list(sheet.get("key_fields") or [])
     key_set = set(keys)
     compare = [f for f in list(sheet.get("compare_fields") or []) if f not in key_set]
@@ -2861,8 +2961,8 @@ def _sheet_csv(sheet: dict[str, Any]) -> str:
             headers.append(f"{f}__post")
         else:
             headers.append(f)
-    lines = [",".join(_csv_cell(h) for h in headers)]
-    for d in list(sheet.get("diffs") or []):
+    yield ",".join(_csv_cell(h) for h in headers)
+    for d in diffs:
         kind = str(d.get("kind") or "")
         pre = dict(d.get("mapped_before") or d.get("before") or {})
         post = dict(d.get("after") or {})
@@ -2887,8 +2987,11 @@ def _sheet_csv(sheet: dict[str, Any]) -> str:
                     row.append(pre.get(f, ""))
                 else:
                     row.append(post.get(f, pre.get(f, "")))
-        lines.append(",".join(_csv_cell(x) for x in row))
-    return "\ufeff" + "\n".join(lines) + "\n"
+        yield ",".join(_csv_cell(x) for x in row)
+
+
+def _sheet_csv(sheet: dict[str, Any]) -> str:
+    return "\ufeff" + "\n".join(_sheet_csv_lines(sheet, sheet.get("diffs") or [])) + "\n"
 
 
 def _enrich_summary(summary: dict[str, Any], sheets: list[dict[str, Any]]) -> dict[str, Any]:
@@ -3001,8 +3104,13 @@ def get_run(db: Session, run_id: str) -> dict[str, Any]:
     r = db.get(BizCompareRun, run_id)
     if not r:
         raise HTTPException(status_code=404, detail="run_not_found")
-    tpl = db.get(BizCompareTemplate, r.template_id) if r.template_id else None
     summary = dict(r.summary_json or {})
+    snapshot = _run_config_snapshot(r)
+    tpl = db.get(BizCompareTemplate, r.template_id) if r.template_id and snapshot is None else None
+    template_out = (
+        deepcopy(snapshot.get("template")) if snapshot is not None
+        else (_template_out(tpl) if tpl else None)
+    )
     raw_sheets = _sheet_meta_from_summary(summary, r, tpl)
     # Never return full diffs in run detail (million-row safe)
     sheets = [
@@ -3023,22 +3131,31 @@ def get_run(db: Session, run_id: str) -> dict[str, Any]:
     ]
     enriched = _enrich_summary(summary, raw_sheets)
     stored = "rows" if _run_has_diff_rows(db, run_id) else "inline"
-    job = db.get(BizCompareJob, r.job_id) if r.job_id else None
-    mapping = db.get(BizPortMapping, r.mapping_id) if r.mapping_id else None
+    job = db.get(BizCompareJob, r.job_id) if r.job_id and snapshot is None else None
+    mapping = db.get(BizPortMapping, r.mapping_id) if r.mapping_id and snapshot is None else None
     before_side = _compare_side(
-        db, r.before_batch_id, fallback_task_id=(job.before_task_id if job else "")
+        db, r.before_batch_id, fallback_task_id=(
+            snapshot.get("before_task_id", "") if snapshot is not None else (job.before_task_id if job else "")
+        ),
     )
     after_side = _compare_side(
-        db, r.after_batch_id, fallback_task_id=(job.after_task_id if job else "")
+        db, r.after_batch_id, fallback_task_id=(
+            snapshot.get("after_task_id", "") if snapshot is not None else (job.after_task_id if job else "")
+        ),
     )
     return {
         "id": r.id,
         "job_id": r.job_id,
-        "job_name": (job.name if job else "") or "",
+        "job_name": (
+            snapshot.get("job_name", "") if snapshot is not None else ((job.name if job else "") or "")
+        ),
         "template_id": r.template_id,
-        "template_name": (tpl.name if tpl else "") or "",
+        "template_name": (template_out or {}).get("name") or "",
         "mapping_id": r.mapping_id,
-        "mapping_name": (mapping.name if mapping else "") or "",
+        "mapping_name": (
+            snapshot.get("mapping_name", "") if snapshot is not None else ((mapping.name if mapping else "") or "")
+        ),
+        "config_snapshot_version": snapshot["version"] if snapshot is not None else None,
         "before_batch_id": r.before_batch_id,
         "after_batch_id": r.after_batch_id,
         "before": before_side,
@@ -3052,7 +3169,7 @@ def get_run(db: Session, run_id: str) -> dict[str, Any]:
         "mapping_stats": r.mapping_stats_json or {},
         "message": r.message,
         "created_at": r.created_at.isoformat() + "Z" if r.created_at else None,
-        "template": _template_out(tpl) if tpl else None,
+        "template": template_out,
     }
 
 
@@ -3152,13 +3269,17 @@ def _field_qf_sql(field_q: dict[str, str], *, prefix: str = "qf") -> tuple[str,
     for i, (name, val) in enumerate(field_q.items()):
         sf = _safe_field(name)
         key = f"{prefix}_{i}"
-        params[key] = f"%{val.lower()}%"
+        params[key] = _literal_search_pattern(val)
         parts.append(
             f"lower(trim(both from coalesce(data_json->>'{sf}', ''))) LIKE :{key}"
         )
     return "(" + " AND ".join(parts) + ")", params
 
 
+def _literal_search_pattern(value: str) -> str:
+    return "%" + value.lower().replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_") + "%"
+
+
 def _load_metric_rows_for_search(
     db: Session,
     *,
@@ -3188,7 +3309,9 @@ def _load_metric_rows_for_search(
     lim = max(1, min(int(cap), _LIVE_SEARCH_LOAD_CAP))
     filters = [f for f in (row_filters or []) if isinstance(f, dict)]
 
-    if _dialect_is_postgres(db):
+    if mid != "lldp_neighbor" and _dialect_is_postgres(db) and (
+        not filters or _filters_sql_compatible(filters)
+    ):
         filter_sql, filter_params = ("TRUE", {})
         if filters and _filters_sql_compatible(filters):
             filter_sql, filter_params = compile_row_filters_sql(filters)
@@ -3202,7 +3325,7 @@ def _load_metric_rows_for_search(
             **qf_params,
         }
         if needle:
-            params["kw"] = f"%{needle.lower()}%"
+            params["kw"] = _literal_search_pattern(needle)
             search_parts.append(_kw_match_sql(key_fields))
         search_sql = " AND ".join(f"({p})" for p in search_parts if p and p != "TRUE")
         if not search_sql:
@@ -3246,20 +3369,21 @@ def _load_metric_rows_for_search(
         return out, truncated
 
     # Non-PG / fallback: scan with early stop (OK for tests / small sheets)
-    q = (
-        db.query(BizStateMetricRow)
-        .filter(
-            BizStateMetricRow.batch_id == bid,
-            BizStateMetricRow.metric_id == mid,
-        )
-        .order_by(BizStateMetricRow.seq.asc(), BizStateMetricRow.id.asc())
-    )
+    model = BizStateLldpNeighbor if mid == "lldp_neighbor" else BizStateMetricRow
+    q = db.query(model).filter(model.batch_id == bid)
+    if mid == "lldp_neighbor":
+        q = q.order_by(model.id.asc())
+    else:
+        q = q.filter(model.metric_id == mid).order_by(model.seq.asc(), model.id.asc())
     out = []
     truncated = False
     needle_l = needle.lower()
     key_set = [str(k).strip() for k in key_fields if str(k).strip()]
     for r in q.yield_per(500):
-        data = dict(r.data_json or {})
+        data = (
+            {f: getattr(r, f) for f in ("local_if", "remote_sys", "remote_if", "remote_ip", "protocol")}
+            if mid == "lldp_neighbor" else dict(r.data_json or {})
+        )
         row = {
             **data,
             "_netx": {
@@ -3297,14 +3421,75 @@ def _load_metric_rows_for_search(
             if not hit:
                 db.expunge(r)
                 continue
-        out.append(row)
         db.expunge(r)
         if len(out) >= lim:
             truncated = True
             break
+        out.append(row)
     return out, truncated
 
 
+def _load_search_key_groups(
+    db: Session, *, batch_id: str, metric_id: str, keys: set[tuple[str, ...]],
+    match_keys: list[str], iface_fields: list[str], row_filters: list[dict[str, Any]],
+    norm_rules: list[dict[str, str]], port_map: dict[str, str],
+) -> tuple[list[dict[str, Any]], set[tuple[str, ...]]]:
+    """Complete candidate key groups so a one-sided search cannot change verdicts.
+
+    Interface transforms run in Python. Other identity fields narrow the SQL
+    query. If the load budget is exceeded, omit entire groups rather than compare
+    partial groups and manufacture additions/deletions.
+    """
+    from ..models import BizStateMetricRow
+
+    model = BizStateLldpNeighbor if metric_id == "lldp_neighbor" else BizStateMetricRow
+    invariant = [f for f in match_keys if f not in iface_fields]
+    positions = [match_keys.index(f) for f in invariant]
+    coarse = sorted({tuple(k[i] for i in positions) for k in keys})
+    out: list[dict[str, Any]] = []
+    complete: set[tuple[str, ...]] = set()
+    # Bound memory even when a candidate identifies a very large duplicate group.
+    budget = _LIVE_SEARCH_GROUP_LOAD_CAP
+    for start in range(0, len(coarse), 100):
+        group = coarse[start:start + 100]
+        group_set = set(group)
+        target = {k for k in keys if tuple(k[i] for i in positions) in group_set}
+        q = db.query(model).filter(model.batch_id == batch_id)
+        if metric_id == "lldp_neighbor":
+            q = q.order_by(model.id.asc())
+        else:
+            q = q.filter(model.metric_id == metric_id).order_by(model.seq.asc(), model.id.asc())
+        if invariant:
+            exprs = [func.trim(func.coalesce(
+                cast(getattr(model, f) if metric_id == "lldp_neighbor"
+                     else model.data_json[f].as_string(), String), "")) for f in invariant]
+            q = q.filter(tuple_(*exprs).in_(group))
+        rows = q.limit(budget + 1).all()
+        if len(rows) > budget:
+            for r in rows:
+                db.expunge(r)
+            # No partial verdicts. Later groups may still fit the remaining budget.
+            continue
+        budget -= len(rows)
+        complete.update(target)
+        for r in rows:
+            data = (
+                {f: getattr(r, f) for f in ("local_if", "remote_sys", "remote_if", "remote_ip", "protocol")}
+                if metric_id == "lldp_neighbor" else dict(r.data_json or {})
+            )
+            row = {**data, "_netx": {"row_id": r.id}}
+            db.expunge(r)
+            if row_filters and not all(row_matches_filter(row, f) for f in row_filters):
+                continue
+            normalized = apply_iface_normalize_rows(
+                [row], iface_fields=iface_fields, rules=norm_rules,
+            )[0]
+            mapped = apply_port_map(normalized, iface_fields=iface_fields, port_map=port_map)
+            if row_key(mapped, match_keys) in target:
+                out.append(row)
+    return out, complete
+
+
 def _live_search_sheet_diffs(
     db: Session,
     run: BizCompareRun,
@@ -3316,7 +3501,7 @@ def _live_search_sheet_diffs(
     page: int,
     page_size: int,
 ) -> dict[str, Any]:
-    """Search before/after metric tables, zip-compare, filter by kind tab."""
+    """Find candidate keys, complete both sides, then compare and filter pairs."""
     mid_src = str(sheet.get("metric_id") or "").strip()
     sid = sheet_key(sheet)
     key_fields = list(sheet.get("key_fields") or [])
@@ -3328,7 +3513,7 @@ def _live_search_sheet_diffs(
     )
     row_filters = list(sheet.get("row_filters") or [])
     # Older runs may lack row_filters on sheet meta — fall back to template
-    if not row_filters and run.template_id:
+    if "row_filters" not in sheet and _run_config_snapshot(run) is None and run.template_id:
         tpl = db.get(BizCompareTemplate, run.template_id)
         if tpl:
             for s in template_metrics(tpl):
@@ -3363,9 +3548,7 @@ def _live_search_sheet_diffs(
     ignore_ports = sheet.get("ignore_port_changes")
     if ignore_ports is not None:
         ignore_ports = bool(ignore_ports)
-    pmap = _port_map_dict(db, str(run.mapping_id or ""))
-    tpl = db.get(BizCompareTemplate, run.template_id) if run.template_id else None
-    norm_rules = template_iface_normalize(tpl)
+    pmap, norm_rules = _run_transform_config(db, run)
 
     before_rows, trunc_b = _load_metric_rows_for_search(
         db,
@@ -3385,6 +3568,31 @@ def _live_search_sheet_diffs(
         kw=free_kw,
         field_q=merged_q,
     )
+    match_keys = list((sheet.get("summary") or {}).get("match_key_fields") or key_fields)
+    if ignore_ports is True and not pmap:
+        match_keys = [f for f in key_fields if f not in iface_fields] or key_fields
+    keys: set[tuple[str, ...]] = set()
+    for rows, mapping in ((before_rows, pmap), (after_rows, {})):
+        for row in apply_iface_normalize_rows(rows, iface_fields=iface_fields, rules=norm_rules):
+            keys.add(row_key(apply_port_map(row, iface_fields=iface_fields, port_map=mapping), match_keys))
+    before_rows, complete_b = _load_search_key_groups(
+        db, batch_id=str(run.before_batch_id or ""), metric_id=mid_src, keys=keys,
+        match_keys=match_keys, iface_fields=iface_fields, row_filters=row_filters,
+        norm_rules=norm_rules, port_map=pmap,
+    )
+    after_rows, complete_a = _load_search_key_groups(
+        db, batch_id=str(run.after_batch_id or ""), metric_id=mid_src, keys=keys,
+        match_keys=match_keys, iface_fields=iface_fields, row_filters=row_filters,
+        norm_rules=norm_rules, port_map={},
+    )
+    complete = complete_b & complete_a
+    def keep_complete(rows: list[dict[str, Any]], mapping: dict[str, str]) -> list[dict[str, Any]]:
+        normalized = apply_iface_normalize_rows(rows, iface_fields=iface_fields, rules=norm_rules)
+        return [orig for orig, norm in zip(rows, normalized) if row_key(
+            apply_port_map(norm, iface_fields=iface_fields, port_map=mapping), match_keys,
+        ) in complete]
+    before_rows = keep_complete(before_rows, pmap)
+    after_rows = keep_complete(after_rows, {})
     result = compare_rows(
         before_rows=before_rows,
         after_rows=after_rows,
@@ -3394,17 +3602,19 @@ def _live_search_sheet_diffs(
         port_map=pmap,
         field_rules=field_rules,
         iface_normalize_rules=norm_rules,
-        ignore_port_changes=ignore_ports,
+        ignore_port_changes=(match_keys != key_fields) if not pmap else False,
         include_unchanged=True,
         unchanged_limit=None,
         compact_unchanged=False,
     )
     kind_n = (kind or "diff").strip().lower()
     filtered = [
-        d for d in list(result.get("diffs") or []) if _kind_allows(kind_n, str(d.get("kind") or ""))
+        d for d in list(result.get("diffs") or [])
+        if _kind_allows(kind_n, str(d.get("kind") or ""))
+        and _diff_matches_search(d, kw=free_kw, field_q=merged_q)
     ]
     # Cap pairs returned to keep UI snappy
-    truncated = bool(trunc_b or trunc_a or len(filtered) > _LIVE_SEARCH_RESULT_CAP)
+    truncated = bool(trunc_b or trunc_a or complete != keys or len(filtered) > _LIVE_SEARCH_RESULT_CAP)
     filtered = filtered[:_LIVE_SEARCH_RESULT_CAP]
     total = len(filtered)
     start = (page - 1) * page_size
@@ -3452,10 +3662,17 @@ def list_run_diffs(
     sheets = _sheet_meta_from_summary(summary, r, tpl)
     asked = (metric_id or "").strip()
     sheet = _lookup_sheet(sheets, asked) if asked else (sheets[0] if sheets else None)
+    if asked and sheet is None:
+        raise HTTPException(status_code=404, detail="sheet_not_found")
     mid = sheet_key(sheet) if sheet else (asked or str(r.metric_id or ""))
 
-    # Unified search: kw and/or field filters → live source lookup; kind tab only filters.
-    if (kw_n or field_q) and sheet:
+    # Only reconstruct missing success rows. Persisted verdicts must stay stable
+    # while searching and avoid loading/recomparing source tables on every page.
+    sheet_summary = (sheet or {}).get("summary") or {}
+    success_missing = int(sheet_summary.get("unchanged") or 0) > int(
+        sheet_summary.get("unchanged_listed") or 0
+    )
+    if (kw_n or field_q) and sheet and kind_n in ("all", "unchanged") and success_missing:
         return _live_search_sheet_diffs(
             db,
             r,
@@ -3480,6 +3697,9 @@ def list_run_diffs(
             q = q.filter(BizCompareDiff.kind == "changed")
         elif kind_n != "all":
             q = q.filter(BizCompareDiff.kind == kind_n)
+        if kw_n or field_q:
+            q = _search_stored_diffs(q, metric_id=str((sheet or {}).get("metric_id") or mid),
+                                     kw=kw_n, field_q=field_q)
         total = q.count()
         rows = (
             q.order_by(BizCompareDiff.seq.asc(), BizCompareDiff.id.asc())
@@ -3487,7 +3707,7 @@ def list_run_diffs(
             .limit(size_n)
             .all()
         )
-        items = _hydrate_diff_rows(db, [_diff_row_out(x) for x in rows])
+        items = _hydrate_diff_rows(db, [_diff_row_out(x) for x in rows], run=r, sheet=sheet)
         return {
             "total": total,
             "page": page_n,
@@ -3505,7 +3725,7 @@ def list_run_diffs(
     inline = list((sheet or {}).get("diffs") or [])
     if not inline and mid == r.metric_id:
         inline = list(r.diffs_json or [])
-    filtered = _filter_inline_diffs(inline, kind=kind_n, kw="")
+    filtered = _filter_inline_diffs(inline, kind=kind_n, kw=kw_n, field_q=field_q)
     total = len(filtered)
     start = (page_n - 1) * size_n
     page_items = filtered[start : start + size_n]
@@ -3520,44 +3740,53 @@ def list_run_diffs(
     }
 
 
-def _iter_sheet_diffs(db: Session, run_id: str, metric_id: str) -> list[dict[str, Any]]:
-    """Load all diffs for one sheet (export). Prefer row table; fall back to inline.
+def _iter_sheet_diffs(db: Session, run_id: str, metric_id: str) -> Iterator[dict[str, Any]]:
+    """Stream one sheet using a stable (seq, id) cursor; fall back to inline.
 
     Compact success rows are hydrated in chunks from metric tables.
     """
+    r = db.get(BizCompareRun, run_id)
+    if not r:
+        return
+    tpl = db.get(BizCompareTemplate, r.template_id) if r.template_id else None
+    sheets = _sheet_meta_from_summary(dict(r.summary_json or {}), r, tpl)
+    sheet = next((s for s in sheets if sheet_key(s) == metric_id), None)
     if _run_has_diff_rows(db, run_id):
-        out: list[dict[str, Any]] = []
-        offset = 0
+        cursor: tuple[int, str] | None = None
         while True:
-            rows = (
-                db.query(BizCompareDiff)
-                .filter(BizCompareDiff.run_id == run_id, BizCompareDiff.metric_id == metric_id)
+            q = db.query(BizCompareDiff).filter(
+                BizCompareDiff.run_id == run_id, BizCompareDiff.metric_id == metric_id,
+            )
+            if cursor is not None:
+                seq, diff_id = cursor
+                q = q.filter(or_(BizCompareDiff.seq > seq, and_(
+                    BizCompareDiff.seq == seq, BizCompareDiff.id > diff_id,
+                )))
+            rows = (q
                 .order_by(BizCompareDiff.seq.asc(), BizCompareDiff.id.asc())
-                .offset(offset)
                 .limit(_DIFF_CHUNK)
                 .all()
             )
             if not rows:
                 break
-            chunk = _hydrate_diff_rows(db, [_diff_row_out(x) for x in rows])
-            out.extend(chunk)
-            offset += len(rows)
+            chunk = _hydrate_diff_rows(db, [_diff_row_out(x) for x in rows], run=r, sheet=sheet)
+            cursor = (rows[-1].seq, rows[-1].id)
+            yield from chunk
             if len(rows) < _DIFF_CHUNK:
                 break
-        return out
-    r = db.get(BizCompareRun, run_id)
-    if not r:
-        return []
+            # Release the previous chunk before loading the next one.
+            del rows, chunk
+        return
     summary = dict(r.summary_json or {})
     sheets = list(summary.get("sheets") or [])
     for sh in sheets:
         if sheet_key(sh) == metric_id or (
             str(sh.get("metric_id") or "") == metric_id and sheet_key(sh) == metric_id
         ):
-            return list(sh.get("diffs") or [])
+            yield from sh.get("diffs") or []
+            return
     if metric_id == r.metric_id:
-        return list(r.diffs_json or [])
-    return []
+        yield from r.diffs_json or []
 
 
 def export_run_zip(db: Session, run_id: str) -> bytes:
@@ -3586,14 +3815,17 @@ def export_run_zip(db: Session, run_id: str) -> bytes:
                 f"+{card.get('added')}/-{card.get('removed')}/~{card.get('changed')}/= {card.get('unchanged')}"
             )
         zf.writestr("manifest.txt", "\n".join(manifest) + "\n")
+        run = db.get(BizCompareRun, run_id)
+        snapshot = _run_config_snapshot(run) if run else None
+        if snapshot is not None:
+            zf.writestr("config_snapshot.json", json.dumps(snapshot, ensure_ascii=False, indent=2))
         for sheet in list(detail.get("sheets") or []):
             sid = sheet_key(sheet) or "sheet"
             safe = "".join(ch if ch.isalnum() or ch in "-_." else "_" for ch in sid)[:80] or "sheet"
-            sheet_full = {
-                **sheet,
-                "diffs": _iter_sheet_diffs(db, run_id, sid),
-            }
-            zf.writestr(f"tables/{safe}.csv", _sheet_csv(sheet_full))
+            with zf.open(f"tables/{safe}.csv", "w", force_zip64=True) as csv_file:
+                csv_file.write(b"\xef\xbb\xbf")
+                for line in _sheet_csv_lines(sheet, _iter_sheet_diffs(db, run_id, sid)):
+                    csv_file.write((line + "\n").encode("utf-8"))
         sum_lines = ["metric_id,mode,before,after,added,removed,changed,unchanged,diff_count,pass_rate"]
         for card in list(s.get("sheet_cards") or []):
             sum_lines.append(
diff --git a/tests/test_biz_state_compare_export.py b/tests/test_biz_state_compare_export.py
new file mode 100644
index 0000000..d176c40
--- /dev/null
+++ b/tests/test_biz_state_compare_export.py
@@ -0,0 +1,87 @@
+"""Export must page lazily, keep tied sequence rows, and produce valid UTF-8 CSV."""
+
+from __future__ import annotations
+
+import csv
+import io
+import unittest
+import zipfile
+from unittest.mock import patch
+
+from sqlalchemy import create_engine, event
+from sqlalchemy.orm import Session
+
+from netx_api.biz_state.compare_service import _iter_sheet_diffs, export_run_zip
+from netx_api.db import Base
+from netx_api.models import BizCompareDiff, BizCompareRun
+
+
+class CompareExportTests(unittest.TestCase):
+    def setUp(self):
+        self.engine = create_engine("sqlite+pysqlite:///:memory:")
+        Base.metadata.create_all(self.engine)
+        self.db = Session(self.engine)
+        self.sheet = {"sheet_id": "sheet", "metric_id": "test", "key_fields": ["id"],
+                      "compare_fields": ["note"], "display_fields": ["id", "note"],
+                      "summary": {"unchanged": 7}, "status": "done"}
+        self.run = BizCompareRun(id="run", metric_id="test", status="success",
+                                 summary_json={"sheets": [self.sheet]})
+        self.db.add(self.run)
+        for seq, ids in ((0, "cba"), (1, "ed"), (2, "gf")):
+            for rid in ids:
+                note = '中文, "quoted"\nsecond line'
+                self.db.add(BizCompareDiff(id=rid, run_id="run", metric_id="sheet", seq=seq,
+                                           kind="unchanged", key_json={"id": rid},
+                                           before_json={"id": rid, "note": note},
+                                           after_json={"id": rid, "note": note}))
+        self.db.commit()
+
+    def tearDown(self):
+        self.db.close()
+        self.engine.dispose()
+
+    def test_cursor_handles_tied_sequence_numbers_across_chunks(self):
+        statements = []
+        def record(_conn, _cursor, statement, _parameters, _context, _executemany):
+            if "ORDER BY biz_compare_diff.seq" in statement:
+                statements.append(statement)
+        event.listen(self.engine, "before_cursor_execute", record)
+        with patch("netx_api.biz_state.compare_service._DIFF_CHUNK", 2):
+            iterator = _iter_sheet_diffs(self.db, "run", "sheet")
+            self.assertEqual(statements, [])
+            self.assertEqual(next(iterator)["key"]["id"], "a")
+            self.assertEqual(next(iterator)["key"]["id"], "b")
+            self.assertEqual(len(statements), 1, "do not load the next chunk early")
+            self.assertEqual([d["key"]["id"] for d in iterator], list("cdefg"))
+        self.assertEqual(len(statements), 4)
+        self.assertTrue(all("biz_compare_diff.seq >" in q and "biz_compare_diff.id >" in q
+                            for q in statements[1:]))
+
+    def test_export_csv_escapes_unicode_quotes_commas_and_newlines(self):
+        with patch("netx_api.biz_state.compare_service._DIFF_CHUNK", 2):
+            archive_bytes = export_run_zip(self.db, "run")
+        with zipfile.ZipFile(io.BytesIO(archive_bytes)) as archive:
+            payload = archive.read("tables/sheet.csv")
+            self.assertTrue(payload.startswith(b"\xef\xbb\xbf"))
+            rows = list(csv.DictReader(io.StringIO(payload.decode("utf-8-sig"))))
+            self.assertEqual([r["id"] for r in rows], list("abcdefg"))
+            self.assertTrue(all(r["note__pre"] == '中文, "quoted"\nsecond line' for r in rows))
+            self.assertTrue(all(r["note__pre"] == r["note__post"] for r in rows))
+
+    def test_other_sheets_do_not_leak_into_export(self):
+        self.db.add(BizCompareDiff(id="other", run_id="run", metric_id="another", seq=0,
+                                   kind="added", key_json={"id": "other"}, after_json={"note": "other"}))
+        self.db.commit()
+        with patch("netx_api.biz_state.compare_service._DIFF_CHUNK", 2):
+            self.assertEqual([d["key"]["id"] for d in _iter_sheet_diffs(self.db, "run", "sheet")], list("abcdefg"))
+
+    def test_legacy_inline_and_empty_exports(self):
+        self.db.query(BizCompareDiff).delete()
+        inline = {"kind": "added", "key": {"id": "legacy"}, "after": {"note": "value"}}
+        self.run.summary_json = {"sheets": [{**self.sheet, "diffs": [inline]}]}
+        self.db.commit()
+        self.assertEqual(list(_iter_sheet_diffs(self.db, "run", "sheet")), [inline])
+        self.run.summary_json = {"sheets": [self.sheet]}
+        self.db.commit()
+        with zipfile.ZipFile(io.BytesIO(export_run_zip(self.db, "run"))) as archive:
+            self.assertEqual(archive.read("tables/sheet.csv").decode("utf-8-sig"), "kind,id,note__pre,note__post\n")
diff --git a/tests/test_biz_state_compare_memory.py b/tests/test_biz_state_compare_memory.py
new file mode 100644
index 0000000..7850798
--- /dev/null
+++ b/tests/test_biz_state_compare_memory.py
@@ -0,0 +1,103 @@
+"""Bounded sampling regressions and an opt-in comparison memory benchmark."""
+
+from __future__ import annotations
+
+import gc
+import hashlib
+import json
+import os
+import random
+import subprocess
+import time
+import tracemalloc
+import types
+from collections import defaultdict, deque
+from copy import deepcopy
+
+import pytest
+
+from netx_api.biz_state.compare_engine import _StratifiedSample, compare_rows, stratify_take
+
+
+def reference_sample(items, limit):
+    if len(items) <= limit:
+        return list(items)
+    buckets = defaultdict(deque)
+    for item in items:
+        buckets[item[0]].append(item)
+    out = []
+    while buckets and len(out) < limit:
+        for key in list(buckets):
+            out.append(buckets[key].popleft())
+            if not buckets[key]:
+                del buckets[key]
+            if len(out) >= limit:
+                break
+    return out
+
+
+def test_streaming_sample_preserves_round_robin_and_small_input_order():
+    rng = random.Random(42)
+    for _ in range(100):
+        items = [(str(rng.randrange(20)), i) for i in range(rng.randrange(100))]
+        for limit in (0, 1, 5, 20, 100):
+            expected = reference_sample(items, limit)
+            assert stratify_take(iter(items), limit, key_fn=lambda item: item[0]) == expected
+
+
+def test_sample_storage_is_bounded_for_both_many_rows_and_many_strata():
+    sample = _StratifiedSample(50)
+    for i in range(20_000):
+        sample.add(i, str(i) if i >= 10_000 else "one-large-peer")
+    assert len(sample.heap) == 50
+    assert len(sample.strata) == 50
+    assert sample.picked() == [0, *range(10_000, 10_049)]
+
+
+@pytest.mark.parametrize("iface_fields,rules,pmap", [
+    ([], [], {}),
+    (["interface"], [], {}),
+    (["interface"], [{"from": "GE", "to": "gei"}], {"gei-1": "gei-2"}),
+])
+def test_compare_does_not_mutate_inputs_when_reusing_rows(iface_fields, rules, pmap):
+    before = [{"id": "1", "interface": "GE-1", "state": "up"}]
+    after = [{"id": "1", "interface": "GE-2", "state": "up"}]
+    original = deepcopy((before, after))
+    compare_rows(before_rows=before, after_rows=after, key_fields=["id"],
+                 iface_fields=iface_fields, compare_fields=["state"], port_map=pmap,
+                 iface_normalize_rules=rules, include_unchanged=True, unchanged_limit=1)
+    assert (before, after) == original
+
+
+def test_memory_benchmark_against_git_baseline():
+    """Run explicitly with NETX_BENCH_ROWS=1000000; not part of routine CI."""
+    n = int(os.environ.get("NETX_BENCH_ROWS", "0"))
+    if not n:
+        pytest.skip("opt-in large synthetic benchmark")
+    baseline_ref = os.environ.get("NETX_BENCH_REF", "HEAD")
+    source = subprocess.run(["git", "show", f"{baseline_ref}:netx_api/biz_state/compare_engine.py"],
+                            check=True, capture_output=True, text=True, encoding="utf-8").stdout
+    baseline = types.ModuleType("netx_api.biz_state._benchmark_baseline")
+    baseline.__package__ = "netx_api.biz_state"
+    exec(compile(source, "baseline_compare_engine.py", "exec"), baseline.__dict__)
+    before = [{"id": str(i), "neighbor": str(i % 50), "state": "up", "afi": "vpnv4"} for i in range(n)]
+    after = [dict(row) for row in before]
+    records = []
+    digests = []
+    for name, compare in (("baseline", baseline.compare_rows), ("optimized", compare_rows)):
+        gc.collect()
+        tracemalloc.start()
+        started = time.perf_counter()
+        result = compare(before_rows=before, after_rows=after, key_fields=["id"], iface_fields=[],
+                         compare_fields=["state"], include_unchanged=True, unchanged_limit=5000,
+                         compact_unchanged=True)
+        elapsed = time.perf_counter() - started
+        _, peak = tracemalloc.get_traced_memory()
+        tracemalloc.stop()
+        digests.append(hashlib.sha256(json.dumps(result, sort_keys=True).encode()).hexdigest())
+        records.append({"engine": name, "rows_per_side": n, "peak_engine_mib": round(peak / 2**20, 2),
+                        "seconds_with_tracemalloc": round(elapsed, 3)})
+        del result
+    assert digests[0] == digests[1], "large-sheet results and sample order must match"
+    print("\n" + json.dumps({"baseline_ref": baseline_ref, "input_allocations_included": False,
+                              "results": records}, indent=2))
diff --git a/tests/test_biz_state_compare_search.py b/tests/test_biz_state_compare_search.py
new file mode 100644
index 0000000..0b30c8f
--- /dev/null
+++ b/tests/test_biz_state_compare_search.py
@@ -0,0 +1,186 @@
+"""Regression tests: searching a pair must not change its comparison verdict."""
+
+from __future__ import annotations
+
+import unittest
+from unittest.mock import patch
+
+from fastapi import HTTPException
+from sqlalchemy import create_engine
+from sqlalchemy.dialects import postgresql
+from sqlalchemy.orm import Session
+
+from netx_api.biz_state.compare_engine import compare_rows
+from netx_api.biz_state.compare_service import (
+    _load_metric_rows_for_search,
+    _literal_search_pattern,
+    _persist_sheet_diffs,
+    _search_stored_diffs,
+    list_run_diffs,
+)
+from netx_api.db import Base
+from netx_api.models import (
+    BizCompareDiff, BizCompareRun, BizCompareTemplate, BizPortMappingRow,
+    BizStateLldpNeighbor, BizStateMetricRow,
+)
+
+
+class CompareSearchTests(unittest.TestCase):
+    def setUp(self) -> None:
+        self.engine = create_engine("sqlite+pysqlite:///:memory:")
+        Base.metadata.create_all(self.engine, tables=[m.__table__ for m in (
+            BizCompareDiff, BizCompareRun, BizCompareTemplate, BizPortMappingRow,
+            BizStateLldpNeighbor, BizStateMetricRow,
+        )])
+        self.db = Session(self.engine)
+
+    def tearDown(self) -> None:
+        self.db.close()
+        self.engine.dispose()
+
+    def seed(self, before, after, *, metric="test", keys=None, fields=None,
+             iface=None, mapping=None, sample=False, compact=False, filters=None):
+        keys = keys or ["id"]
+        iface = iface or []
+        fields = fields or ["state"]
+        for side, rows in (("before", before), ("after", after)):
+            for seq, row in enumerate(rows):
+                rid = f"{side}-{seq}"
+                if metric == "lldp_neighbor":
+                    self.db.add(BizStateLldpNeighbor(id=rid, batch_id=side, **row))
+                else:
+                    self.db.add(BizStateMetricRow(id=rid, batch_id=side, metric_id=metric,
+                                                seq=seq, data_json=dict(row)))
+                row["_netx"] = {"row_id": rid}
+        if mapping:
+            for b, a in mapping.items():
+                self.db.add(BizPortMappingRow(mapping_id="map", before_if=b, after_if=a))
+        result = compare_rows(before_rows=before, after_rows=after, key_fields=keys,
+                              compare_fields=fields, iface_fields=iface, port_map=mapping or {},
+                              include_unchanged=True, compact_unchanged=compact)
+        sheet = {"sheet_id": "sheet", "metric_id": metric, "key_fields": keys,
+                 "compare_fields": fields, "iface_fields": iface,
+                 "summary": {**result["summary"], "unchanged_listed": 0 if sample
+                             else result["summary"]["unchanged"]},
+                 "row_filters": filters or []}
+        self.db.add(BizCompareRun(id="run", status="success", before_batch_id="before",
+                                 after_batch_id="after", mapping_id="map" if mapping else "",
+                                 summary_json={"sheets": [sheet]}))
+        self.db.commit()
+        _persist_sheet_diffs(self.db, run_id="run", metric_id="sheet",
+                             diffs=[d for d in result["diffs"]
+                                    if not sample or d["kind"] != "unchanged"])
+        return result
+
+    def search(self, **kwargs):
+        return list_run_diffs(self.db, "run", metric_id="sheet", **kwargs)
+
+    def test_search_before_value_preserves_changed_verdict(self):
+        self.seed([{"id": "1", "state": "up"}], [{"id": "1", "state": "down"}])
+        with patch("netx_api.biz_state.compare_service.compare_rows") as compare:
+            result = self.search(kw="up", kind="changed")
+            compare.assert_not_called()
+        self.assertEqual(result["total"], 1)
+        self.assertEqual(result["items"][0]["after"]["state"], "down")
+        self.assertEqual(self.search(kw="up", kind="removed")["total"], 0)
+
+    def test_field_filter_matches_after_value_on_compact_diff(self):
+        self.seed([{"id": "1", "state": "up"}], [{"id": "1", "state": "down"}], compact=True)
+        result = self.search(qf={"state": "down"}, kind="changed")
+        self.assertEqual(result["total"], 1)
+        self.assertEqual(result["items"][0]["before"]["state"], "up")
+
+    def test_sample_lookup_completes_pair_matching_only_one_side(self):
+        self.seed([{"id": "1", "state": "up", "age": "old"}],
+                  [{"id": "1", "state": "up", "age": "new"}], sample=True)
+        result = self.search(kw="old", kind="unchanged")
+        self.assertEqual(result["source"], "live")
+        self.assertEqual(result["total"], 1)
+        self.assertEqual(result["items"][0]["after"]["age"], "new")
+
+    def test_sample_lookup_does_not_turn_changed_into_removed(self):
+        self.seed([{"id": "1", "state": "up"}, {"id": "2", "state": "up"}],
+                  [{"id": "1", "state": "down"}, {"id": "2", "state": "up"}], sample=True)
+        result = self.search(qf={"id": "1", "state": "up"}, kind="all")
+        self.assertEqual([d["kind"] for d in result["items"]], ["changed"])
+
+    def test_sample_lookup_keeps_duplicate_group_pairing(self):
+        self.seed([{"id": "1", "state": "up", "note": "first"},
+                   {"id": "1", "state": "down", "note": "second"}],
+                  [{"id": "1", "state": "up", "note": "other"},
+                   {"id": "1", "state": "down", "note": "other"}], sample=True)
+        result = self.search(kw="second", kind="all")
+        self.assertEqual([d["kind"] for d in result["items"]], ["unchanged"])
+        self.assertEqual(result["items"][0]["after"]["state"], "down")
+
+    def test_sample_search_is_not_limited_by_persisted_search_text_size(self):
+        self.seed([{"id": "1", "state": "up", "note": "x" * 5000 + "needle"}],
+                  [{"id": "1", "state": "up", "note": "other"}], sample=True)
+        self.assertEqual(self.search(kw="needle", kind="unchanged")["total"], 1)
+
+    def test_sample_lookup_applies_port_mapping_before_matching(self):
+        self.seed([{"id": "1", "interface": "old", "state": "up"}],
+                  [{"id": "1", "interface": "new", "state": "up"}],
+                  keys=["id", "interface"], iface=["interface"], mapping={"old": "new"}, sample=True)
+        result = self.search(qf={"interface": "old"}, kind="unchanged")
+        self.assertEqual(result["total"], 1)
+        self.assertEqual(result["items"][0]["after"]["interface"], "new")
+
+    def test_lldp_source_search(self):
+        self.seed([{"local_if": "gei-1", "remote_sys": "peer", "remote_if": "port", "remote_ip": "1"}],
+                  [{"local_if": "gei-1", "remote_sys": "peer", "remote_if": "port", "remote_ip": "1"}],
+                  metric="lldp_neighbor", keys=["local_if", "remote_sys"], fields=["remote_ip"],
+                  iface=["local_if"], sample=True)
+        self.assertEqual(self.search(kw="peer", kind="unchanged")["total"], 1)
+
+    def test_lldp_compact_diff_field_search(self):
+        self.seed([{"local_if": "gei-1", "remote_sys": "peer", "remote_ip": "1"}],
+                  [{"local_if": "gei-1", "remote_sys": "peer", "remote_ip": "2"}],
+                  metric="lldp_neighbor", keys=["local_if", "remote_sys"], fields=["remote_ip"],
+                  iface=["local_if"], compact=True)
+        self.assertEqual(self.search(qf={"remote_ip": "2"}, kind="changed")["total"], 1)
+
+    def test_exact_load_cap_is_not_truncated(self):
+        self.seed([{"id": "1", "state": "up"}], [{"id": "1", "state": "up"}])
+        rows, truncated = _load_metric_rows_for_search(self.db, batch_id="before", metric_id="test",
+                                                       row_filters=[], key_fields=["id"], kw="up", cap=1)
+        self.assertEqual(len(rows), 1)
+        self.assertFalse(truncated)
+
+    def test_incomplete_duplicate_groups_are_omitted(self):
+        self.seed([{"id": "1", "state": "up"}, {"id": "1", "state": "down"}],
+                  [{"id": "1", "state": "up"}, {"id": "1", "state": "down"}], sample=True)
+        with patch("netx_api.biz_state.compare_service._LIVE_SEARCH_GROUP_LOAD_CAP", 1):
+            result = self.search(kw="up", kind="all")
+        self.assertEqual(result["items"], [])
+        self.assertTrue(result["truncated"])
+
+    def test_postgres_search_pattern_treats_wildcards_literally(self):
+        self.assertEqual(_literal_search_pattern("a_10%"), "%a\\_10\\%%")
+
+    def test_postgres_unsupported_filter_uses_python(self):
+        self.seed([{"id": "1", "state": "up", "age": "H"}],
+                  [{"id": "1", "state": "up", "age": "H"}])
+        with patch("netx_api.biz_state.compare_sql._dialect_is_postgres", return_value=True):
+            rows, _ = _load_metric_rows_for_search(self.db, batch_id="before", metric_id="test",
+                         row_filters=[{"field": "age", "op": "age_timer"}], key_fields=["id"], kw="up")
+        self.assertEqual(rows, [])
+
+    def test_unknown_sheet_is_rejected(self):
+        self.seed([{"id": "1", "state": "up"}], [{"id": "1", "state": "up"}])
+        with self.assertRaises(HTTPException) as error:
+            list_run_diffs(self.db, "run", metric_id="missing")
+        self.assertEqual(error.exception.detail, "sheet_not_found")
+
+    def test_stored_search_escapes_like_wildcards(self):
+        self.seed([{"id": "x_1", "state": "up"}, {"id": "xy1", "state": "up"}],
+                  [{"id": "x_1", "state": "down"}, {"id": "xy1", "state": "down"}])
+        self.assertEqual(self.search(qf={"id": "x_1"}, kind="changed")["total"], 1)
+        query = _search_stored_diffs(self.db.query(BizCompareDiff), metric_id="test",
+                                    kw="up", field_q={"id": "x_1"})
+        compiled = query.statement.compile(dialect=postgresql.dialect())
+        self.assertIn("LEFT OUTER JOIN", str(compiled))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_biz_state_compare_snapshot.py b/tests/test_biz_state_compare_snapshot.py
new file mode 100644
index 0000000..3736fa8
--- /dev/null
+++ b/tests/test_biz_state_compare_snapshot.py
@@ -0,0 +1,170 @@
+"""Historical comparisons must use the settings captured when they were queued."""
+
+from __future__ import annotations
+
+import csv
+import io
+import json
+import unittest
+import zipfile
+from copy import deepcopy
+from unittest.mock import patch
+
+from sqlalchemy import create_engine
+from sqlalchemy.orm import sessionmaker
+
+from netx_api.biz_state.compare_service import (
+    _create_running_run, _execute_compare_into_run, cancel_compare_run,
+    export_run_zip, get_run, list_run_diffs, template_metrics,
+)
+from netx_api.db import Base
+from netx_api.models import (
+    BizCompareDiff, BizCompareJob, BizCompareTemplate, BizPortMapping,
+    BizPortMappingRow, BizStateMetricRow,
+)
+
+
+class CompareSnapshotTests(unittest.TestCase):
+    def setUp(self):
+        self.engine = create_engine("sqlite+pysqlite:///:memory:")
+        Base.metadata.create_all(self.engine)
+        self.factory = sessionmaker(bind=self.engine, expire_on_commit=False)
+        self.db = self.factory()
+        # Progress uses a separate session in production; keep every test write local.
+        self.session_patch = patch("netx_api.db.SessionLocal", self.factory)
+        self.session_patch.start()
+        self.sheet = {
+            "sheet_id": "sheet", "metric_id": "test", "title": "Original sheet",
+            "key_fields": ["id", "interface"], "iface_fields": ["interface"],
+            "compare_fields": ["state"], "display_fields": ["id", "interface", "state"],
+            "row_filters": [], "ignore_port_changes": False,
+        }
+        self.tpl = BizCompareTemplate(
+            id="tpl", name="Original template", metrics_json=[deepcopy(self.sheet)],
+            iface_normalize_json=[{"from": "GE", "to": "gei"}],
+        )
+        self.job = BizCompareJob(id="job", name="Original job", template_id="tpl",
+                                 mapping_id="map", store_unchanged="sample")
+        self.mapping = BizPortMapping(id="map", name="Original mapping")
+        self.pair = BizPortMappingRow(mapping_id="map", before_if="gei-1", after_if="gei-2")
+        self.db.add_all([self.tpl, self.job, self.mapping, self.pair])
+        for side, iface, note in (("before", "GE-1.100", "old"), ("after", "GE-2.100", "new")):
+            self.db.add(BizStateMetricRow(id=side, batch_id=side, metric_id="test", seq=0,
+                                         data_json={"id": "1", "interface": iface,
+                                                    "state": "up", "note": note}))
+        self.db.commit()
+        self.run = _create_running_run(
+            self.db, job=self.job, tpl=self.tpl, before_batch_id="before",
+            after_batch_id="after", sheets_cfg=template_metrics(self.tpl),
+        )
+
+    def tearDown(self):
+        self.session_patch.stop()
+        self.db.close()
+        self.engine.dispose()
+
+    def edit_current_settings(self):
+        changed = {**self.sheet, "title": "Edited sheet", "compare_fields": ["note"],
+                   "row_filters": [{"field": "state", "op": "eq", "value": "down"}]}
+        self.tpl.name = "Edited template"
+        self.tpl.metrics_json = [changed]
+        self.tpl.iface_normalize_json = [{"from": "GE", "to": "other"}]
+        self.pair.after_if = "gei-99"
+        self.job.name = "Edited job"
+        self.job.store_unchanged = "never"
+        self.job.enabled_sheet_ids = ["different-sheet"]
+        self.mapping.name = "Edited mapping"
+        self.db.commit()
+
+    def execute(self):
+        return _execute_compare_into_run(self.db, self.run.id)
+
+    def test_queued_run_uses_original_sheet_rules_mapping_and_storage(self):
+        original = deepcopy(self.run.summary_json["config_snapshot"])
+        self.edit_current_settings()
+        detail = self.execute()
+        self.assertEqual(detail["status"], "success")
+        self.assertEqual(detail["summary"]["unchanged"], 1)
+        self.assertEqual(detail["summary"]["unchanged_listed"], 1)
+        self.assertEqual(detail["summary"]["store_unchanged"], "sample")
+        self.assertEqual(detail["sheets"][0]["compare_fields"], ["state"])
+        self.assertEqual(self.run.summary_json["config_snapshot"], original)
+
+    def test_compact_hydration_restores_original_normalized_values(self):
+        self.execute()
+        self.edit_current_settings()
+        diff = list_run_diffs(self.db, self.run.id, metric_id="sheet", kind="unchanged")["items"][0]
+        self.assertEqual(diff["before"]["interface"], "GE-1.100")
+        self.assertEqual(diff["mapped_before"]["interface"], "gei-2.100")
+        self.assertEqual(diff["after"]["interface"], "gei-2.100")
+
+    def test_live_search_does_not_inherit_new_filters_or_transforms(self):
+        with patch("netx_api.biz_state.compare_service._UNCHANGED_SAMPLE_MAX", 0):
+            self.execute()
+        self.edit_current_settings()
+        result = list_run_diffs(self.db, self.run.id, metric_id="sheet", kind="unchanged", kw="old")
+        self.assertEqual(result["source"], "live")
+        self.assertEqual(result["total"], 1)
+        self.assertEqual(result["items"][0]["mapped_before"]["interface"], "gei-2.100")
+
+    def test_snapshot_can_be_read_after_mapping_and_template_are_deleted(self):
+        self.execute()
+        self.db.query(BizPortMappingRow).delete()
+        self.db.delete(self.mapping)
+        self.db.delete(self.tpl)
+        self.db.commit()
+        detail = get_run(self.db, self.run.id)
+        self.assertEqual(detail["template_name"], "Original template")
+        self.assertEqual(detail["mapping_name"], "Original mapping")
+        diff = list_run_diffs(self.db, self.run.id, metric_id="sheet", kind="unchanged")["items"][0]
+        self.assertEqual(diff["mapped_before"]["interface"], "gei-2.100")
+
+    def test_queued_run_can_execute_after_its_template_and_mapping_are_deleted(self):
+        self.db.query(BizPortMappingRow).delete()
+        self.db.delete(self.mapping)
+        self.db.delete(self.tpl)
+        self.db.commit()
+        detail = self.execute()
+        self.assertEqual(detail["status"], "success")
+        self.assertEqual(detail["summary"]["unchanged"], 1)
+
+    def test_detail_returns_original_names_and_independent_template_copy(self):
+        self.edit_current_settings()
+        detail = get_run(self.db, self.run.id)
+        self.assertEqual(detail["job_name"], "Original job")
+        self.assertEqual(detail["config_snapshot_version"], 1)
+        self.assertEqual(detail["template"]["iface_normalize_rules"], [{"from": "GE", "to": "gei"}])
+        self.assertNotIn("config_snapshot", detail["summary"])
+        detail["template"]["iface_normalize_rules"].clear()
+        self.assertTrue(self.run.summary_json["config_snapshot"]["template"]["iface_normalize_rules"])
+
+    def test_export_contains_original_config_and_hydrated_csv(self):
+        self.execute()
+        self.edit_current_settings()
+        with zipfile.ZipFile(io.BytesIO(export_run_zip(self.db, self.run.id))) as archive:
+            config = json.loads(archive.read("config_snapshot.json"))
+            self.assertEqual(config["port_map"], {"gei-1": "gei-2"})
+            self.assertEqual(config["template"]["name"], "Original template")
+            rows = list(csv.DictReader(io.StringIO(archive.read("tables/sheet.csv").decode("utf-8-sig"))))
+            self.assertEqual(rows[0]["interface"], "gei-2.100")
+
+    def test_failed_and_cancelled_runs_keep_snapshot(self):
+        original = deepcopy(self.run.summary_json["config_snapshot"])
+        with patch("netx_api.biz_state.compare_service._run_sheet", side_effect=RuntimeError("test-failure")):
+            with self.assertRaisesRegex(RuntimeError, "test-failure"):
+                self.execute()
+        self.assertEqual(self.run.summary_json["config_snapshot"], original)
+        self.run.status = "running"
+        self.db.commit()
+        cancel_compare_run(self.db, self.run.id)
+        self.assertEqual(self.run.summary_json["config_snapshot"], original)
+
+    def test_legacy_runs_without_snapshot_remain_readable(self):
+        self.execute()
+        self.run.summary_json = {k: v for k, v in self.run.summary_json.items() if k != "config_snapshot"}
+        self.db.commit()
+        detail = get_run(self.db, self.run.id)
+        self.assertEqual(detail["status"], "success")
+        self.assertIsNone(detail["config_snapshot_version"])
+        self.assertEqual(list_run_diffs(self.db, self.run.id, metric_id="sheet", kind="unchanged")["total"], 1)
+        self.assertEqual(self.db.query(BizCompareDiff).count(), 1)
diff --git a/web/src/i18n/en.ts b/web/src/i18n/en.ts
index 1061040..7c247dd 100644
--- a/web/src/i18n/en.ts
+++ b/web/src/i18n/en.ts
@@ -474,6 +474,10 @@ const en = {
     sideAfter: "After",
     sidesVs: "vs",
     sidesTitle: "Compare sides",
+    configSnapshot: "Execution settings saved",
+    configSnapshotHint: "Template rules and port mapping are fixed for this run. Later edits do not affect historical lookup. The downloaded report includes the execution settings.",
+    configLegacy: "Legacy run: settings not saved",
+    configLegacyHint: "This run has no configuration snapshot. Success lookup and compact row reconstruction use the current settings, which may differ from the original execution.",
     pick: "Select…",
     sameAsBefore: "Same as before / select…",
     autoHint: "auto: new after-task batches trigger compare",
@@ -532,8 +536,14 @@ const en = {
     ranWithDuration: "Compare finished ({{s}}s)",
     unchangedNotStored: "Success rows were counted but not stored. Set “Store success rows” to sample and re-run for spot-check.",
     unchangedSampleHint: "{{total}} success rows total; browsing a stratified sample of {{listed}}. Search any route for live source lookup. Pass rate uses all {{total}}.",
-    liveSearchHint: "Field filters look up source batches live; tabs only filter kind. E.g. direction=out, network=1.1.1.1.",
-    liveSearchTruncatedHint: "Search results truncated — narrow the field filters.",
+    liveSearchHint: "Pairs are completed from source batches before filtering. Pass rate still uses the full sheet.",
+    liveSearchTruncatedHint: "Search limit reached; some entries are omitted. Narrow the filters. Incomplete key groups were skipped.",
+    clearSearch: "Clear filters",
+    resultLoading: "Loading comparison details…",
+    resultRange: "{{from}}–{{to}} of {{total}}",
+    resultLoadFailed: "Could not load comparison details",
+    retry: "Retry",
+    searchEmpty: "No results match these filters. Clear filters or select another result type.",
     clearKeyFilters: "Clear fields",
     storeUnchanged: "Store success rows",
     storeUnchangedAuto: "Auto (full if small / sample 5k + hydrate if large)",
@@ -567,8 +577,6 @@ const en = {
     pickBatchRun: "Select compare run",
     pickRun: "Select run…",
     runCount: "{{n}} runs",
-    resultEmpty: "No matching diff rows",
-    resultFilterPh: "Free text, or direction:out network:1.1.1.1",
     kindAll: "All",
     kindDiff: "Fail",
     kindAdded: "Added",
@@ -712,7 +720,6 @@ const en = {
     newBaseline: "New baseline batch",
     needProjectFields: "Enter a name and pick old/new portrait tasks",
     needExpectPorts: "Select at least one expected object for this batch",
-    needBaselineFirst: "Save an old baseline batch first to list expect objects",
     needBothBaselines: "Save both old and new baseline batches before evaluating",
     errOldBaselineRequired: "Old baseline batch is required",
     errNewBaselineRequired: "New baseline batch is required",
diff --git a/web/src/i18n/zh.ts b/web/src/i18n/zh.ts
index 1c9b0fc..2eb360f 100644
--- a/web/src/i18n/zh.ts
+++ b/web/src/i18n/zh.ts
@@ -473,6 +473,10 @@ const zh = {
     sideAfter: "操作后",
     sidesVs: "比对",
     sidesTitle: "比对双方",
+    configSnapshot: "执行配置已保存",
+    configSnapshotHint: "此记录的模板规则和端口映射已固定；后续编辑不影响历史回查，下载报告包含执行配置。",
+    configLegacy: "旧记录：未保存执行配置",
+    configLegacyHint: "此记录创建时未保存配置快照；成功行回查和精简明细还原使用当前配置，修改配置后可能与当时结果不同。",
     pick: "选择…",
     sameAsBefore: "同操作前 / 选择…",
     autoHint: "auto：操作后任务新批次到达会自动跑比对",
@@ -531,8 +535,14 @@ const zh = {
     ranWithDuration: "比对完成（耗时 {{s}} 秒）",
     unchangedNotStored: "成功行仅统计数量未落库。可在任务配置将「成功行保存」改为抽样后重新比对（抽查用）。",
     unchangedSampleHint: "成功共 {{total}} 条，明细抽样 {{listed}} 条（分层抽查）。要查任意路由请输入筛选条件——将按原表即时判定。通过率按全部 {{total}} 计。",
-    liveSearchHint: "按字段回查原表即时判定；页签只过滤种类。例：direction=out，network=1.1.1.1。",
-    liveSearchTruncatedHint: "搜索结果已截断，请再收窄字段条件。",
+    liveSearchHint: "从原始批次补齐同一业务的前后数据后筛选；通过率仍按全表统计。",
+    liveSearchTruncatedHint: "搜索已达上限，部分业务未展示。请收窄条件；不完整的业务分组已跳过。",
+    clearSearch: "清空筛选",
+    resultLoading: "正在加载对比明细…",
+    resultRange: "{{from}}–{{to}} / {{total}} 条",
+    resultLoadFailed: "对比明细加载失败",
+    retry: "重试",
+    searchEmpty: "当前筛选条件下没有匹配结果，可清空筛选或切换结果类型。",
     clearKeyFilters: "清空字段",
     storeUnchanged: "成功行保存",
     storeUnchangedAuto: "自动（小表全量 / 大表抽样 5000+原表补全）",
@@ -778,7 +788,6 @@ const zh = {
     metaBaseline: "基线",
     baselineOk: "已设",
     baselineMissing: "缺",
-    needBaselineFirst: "先设基线",
     enableHighfreq: "开启高频",
     highfreqReady: "高频已就绪",
     collectNow: "立即采集",
@@ -830,10 +839,6 @@ const zh = {
     currentMissingBanner: "缺当前批次：{{metrics}}",
     highfreqReadyInactive: "高频已绑定（窗口未开，未立即采集）",
     collectPartial: "部分失败（成功 {{ok}} / 失败 {{fail}}）",
-    confirmProjectDone: "确认完成项目并暂停全部高频？",
-    projectMarkedDone: "已完成，高频已停",
-    confirmDeleteProject: "确认删除项目？将停高频并清批次记录。",
-    projectDeleted: "已删除",
     errMetricNeedsBindings: "监控项「{{metric}}」需要参数绑定（{{params}}），割接高频暂不支持，请去掉该项",
     errNoProfileForMetric: "监控项「{{metric}}」在当前设备厂商下无可用采集 profile",
     acceptTitle: "本批验收",
diff --git a/web/src/index.css b/web/src/index.css
index 2046460..96e00ff 100644
--- a/web/src/index.css
+++ b/web/src/index.css
@@ -11931,6 +11931,22 @@ body:has(.bs-cmp-board:fullscreen) {
   opacity: 0.75;
 }
 
+.bs-cmp-board__config {
+  flex: 0 1 auto;
+  min-width: 0;
+  padding: 4px 7px;
+  border: 1px solid var(--bs-cmp-border, rgba(148, 163, 184, 0.22));
+  border-radius: 6px;
+  color: var(--nm-muted, #94a3b8);
+  font-size: 11px;
+  line-height: 1.4;
+  overflow-wrap: anywhere;
+}
+
+.bs-cmp-board__config.is-legacy {
+  color: #fbbf24;
+}
+
 .bs-cmp-board__actions {
   margin-left: auto;
   flex: 0 0 auto;
@@ -12497,7 +12513,7 @@ body:has(.bs-cmp-board:fullscreen) {
   flex-wrap: wrap;
   align-items: center;
   gap: 4px;
-  flex: 1 1 280px;
+  flex: 1 1 100%;
   min-width: 0;
   padding: 2px;
   border-radius: 8px;
@@ -12615,13 +12631,13 @@ body:has(.bs-cmp-board:fullscreen) {
 
 .bs-cmp-filter-bar {
   display: flex;
-  flex-wrap: nowrap;
+  flex-wrap: wrap;
   gap: 8px;
   align-items: center;
-  flex: 1 1 200px;
-  min-width: 160px;
-  max-width: 320px;
-  margin-left: auto;
+  flex: 1 1 100%;
+  min-width: 0;
+  max-width: none;
+  margin-left: 0;
   min-height: 32px;
 }
 
@@ -12630,6 +12646,8 @@ body:has(.bs-cmp-board:fullscreen) {
 .bs-cmp-filter-bar input {
   min-width: 0;
   width: 100%;
+  flex: 1 1 220px;
+  max-width: 440px;
 }
 
 .bs-cmp-filter-bar .bs-sheet-count {
@@ -12717,11 +12735,57 @@ body:has(.bs-cmp-board:fullscreen) {
 }
 
 .bs-cmp-result-table.is-loading {
-  opacity: 0.55;
+  opacity: 1;
   /* Keep header/filters clickable; only mute row hit-testing via opacity cue */
   pointer-events: auto;
 }
 
+.bs-cmp-result-state {
+  min-height: 180px;
+  display: flex;
+  flex-direction: column;
+  align-items: center;
+  justify-content: center;
+  gap: 12px;
+  padding: 24px;
+  text-align: center;
+  white-space: normal;
+  overflow-wrap: anywhere;
+  color: var(--bs-cmp-muted);
+}
+
+.bs-cmp-result-state strong {
+  color: var(--bs-cmp-ink);
+  font-size: 14px;
+}
+
+.bs-cmp-result-state.is-error strong {
+  color: var(--danger, #fca5a5);
+}
+
+.bs-cmp-strip__kind:focus-visible,
+.bs-cmp-strip__toggle:focus-visible,
+.bs-cmp-nav__item:focus-visible {
+  outline: 2px solid var(--accent, #60a5fa);
+  outline-offset: 2px;
+}
+
+@media (max-width: 600px) {
+  .bs-cmp-strip__sheet {
+    flex-wrap: wrap;
+  }
+
+  .bs-cmp-filter-bar input {
+    flex-basis: 100%;
+    max-width: none;
+  }
+
+  .bs-cmp-key-filter {
+    flex: 1 1 120px;
+    max-width: none;
+  }
+}
+
 .bs-cmp-board.is-fullscreen .bs-cmp-result-table {
   max-height: none;
   min-height: 0;
@@ -12752,6 +12816,7 @@ body:has(.bs-cmp-board:fullscreen) {
 @media (max-width: 960px) {
   .bs-cmp-board__body {
     grid-template-columns: 1fr;
+    grid-template-rows: auto minmax(0, 1fr);
   }
 
   .bs-cmp-board__body.is-nav-collapsed {
@@ -12760,6 +12825,7 @@ body:has(.bs-cmp-board:fullscreen) {
 
   .bs-cmp-nav {
     max-height: 180px;
+    align-self: start;
   }
 
   .bs-cmp-nav.is-collapsed {
@@ -12769,6 +12835,8 @@ body:has(.bs-cmp-board:fullscreen) {
   .bs-cmp-nav__list {
     flex-direction: row;
     flex-wrap: nowrap;
+    flex: 0 0 auto;
+    align-items: flex-start;
     overflow-x: auto;
     overflow-y: hidden;
   }
@@ -12779,8 +12847,9 @@ body:has(.bs-cmp-board:fullscreen) {
 
   .bs-cmp-nav__item {
     flex: 0 0 auto;
-    min-width: 180px;
-    grid-template-columns: 8px minmax(0, 1fr);
+    width: auto;
+    min-width: 240px;
+    grid-template-columns: 8px minmax(0, 1fr) auto;
   }
 }
 
diff --git a/web/src/pages/network/BizComparePage.tsx b/web/src/pages/network/BizComparePage.tsx
index 22f2213..57c5079 100644
--- a/web/src/pages/network/BizComparePage.tsx
+++ b/web/src/pages/network/BizComparePage.tsx
@@ -845,7 +845,7 @@ function TplRowFiltersEditor({
                       Array.isArray(filt.value) ? filt.value.join(",") : String(filt.value ?? "")
                     }
                     placeholder={t("bizCompare.filterValue")}
-                    isDisabled={["empty", "not_empty", "age_timer"].includes(filt.op || "")}
+                    disabled={["empty", "not_empty", "age_timer"].includes(filt.op || "")}
                     onChange={(e) => {
                       const next = groups.map((g) => g.map((x) => ({ ...x })));
                       const op = next[gi][fi]?.op || "eq";
@@ -893,10 +893,6 @@ function TplRowFiltersEditor({
   );
 }
 
-function metricLabel(id: string) {
-  return id;
-}
-
 export type BizComparePageMode = "jobs" | "templates" | "all";
 
 export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePageMode }) {
@@ -988,6 +984,12 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
   const [diffsLoading, setDiffsLoading] = useState(false);
   const [diffsSource, setDiffsSource] = useState<DiffListSource>("");
   const [diffsTruncated, setDiffsTruncated] = useState(false);
+  const [diffsError, setDiffsError] = useState("");
+  const [diffsRetry, setDiffsRetry] = useState(0);
+  const jobRequestRef = useRef(0);
+  const runRequestRef = useRef(0);
+  const listsRequestRef = useRef(0);
+  const diffFiltersRef = useRef("");
   const boardRef = useRef<HTMLDivElement | null>(null);
   const tableScrollRef = useRef<HTMLDivElement | null>(null);
   const tableScrollPosRef = useRef({ top: 0, left: 0 });
@@ -996,6 +998,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
   const tplImportRef = useRef<HTMLInputElement | null>(null);
 
   const refresh = useCallback(async (opts?: { force?: boolean }) => {
+    const request = ++listsRequestRef.current;
     type Bundle = {
       taskRes: Awaited<ReturnType<typeof bizStateListTasks>>;
       tpl: Awaited<ReturnType<typeof bizCompareListTemplates>>;
@@ -1014,6 +1017,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
       return { taskRes, tpl, maps, j, met };
     };
     const apply = (b: Bundle) => {
+      if (request !== listsRequestRef.current) return;
       setTasks((b.taskRes.items || []) as TaskOpt[]);
       setTemplates((b.tpl.items || []) as Template[]);
       setMappings((b.maps.items || []) as Mapping[]);
@@ -1038,8 +1042,12 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, [refresh]);
 
+  useEffect(() => () => { listsRequestRef.current += 1; }, []);
+
   // Keep local tab + dismiss overlays when route mode flips (component may be reused).
   useEffect(() => {
+    jobRequestRef.current += 1;
+    runRequestRef.current += 1;
     if (pageMode === "templates") setPageTab("templates");
     else if (pageMode === "jobs") setPageTab("jobs");
     setTplOpen(false);
@@ -1053,39 +1061,55 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
       void document.exitFullscreen().catch(() => undefined);
     }
     setBoardFs(false);
+    return () => {
+      jobRequestRef.current += 1;
+      runRequestRef.current += 1;
+    };
   }, [pageMode]);
 
   useEffect(() => {
+    let cancelled = false;
+    setBeforeBatches([]);
     void (async () => {
-      if (!beforeTaskId) {
-        setBeforeBatches([]);
-        return;
+      if (!beforeTaskId) return;
+      try {
+        const b = await cutoverCachedGet(`bizCompare:batches:${beforeTaskId}`,
+          () => bizStateListBatches(beforeTaskId), { ttlMs: 5000 });
+        if (!cancelled) setBeforeBatches((b.items || []) as BatchOpt[]);
+      } catch (e) {
+        if (!cancelled) showError(formatErr(e));
       }
-      const b = await bizStateListBatches(beforeTaskId);
-      setBeforeBatches((b.items || []) as BatchOpt[]);
     })();
-  }, [beforeTaskId]);
+    return () => { cancelled = true; };
+  }, [beforeTaskId, showError]);
 
   useEffect(() => {
+    let cancelled = false;
+    setAfterBatches([]);
     void (async () => {
-      if (!afterTaskId) {
-        setAfterBatches([]);
-        return;
+      if (!afterTaskId) return;
+      try {
+        const b = await cutoverCachedGet(`bizCompare:batches:${afterTaskId}`,
+          () => bizStateListBatches(afterTaskId), { ttlMs: 5000 });
+        if (!cancelled) setAfterBatches((b.items || []) as BatchOpt[]);
+      } catch (e) {
+        if (!cancelled) showError(formatErr(e));
       }
-      const b = await bizStateListBatches(afterTaskId);
-      setAfterBatches((b.items || []) as BatchOpt[]);
     })();
-  }, [afterTaskId]);
+    return () => { cancelled = true; };
+  }, [afterTaskId, showError]);
+
+  const templatesById = useMemo(() => new Map(templates.map((x) => [x.id, x])), [templates]);
 
   const filteredJobs = useMemo(() => {
     const kw = debouncedListKw.trim().toLowerCase();
     if (!kw) return jobs;
     return jobs.filter((j) => {
-      const tpl = templates.find((x) => x.id === j.template_id);
+      const tpl = templatesById.get(j.template_id);
       const mids = (tpl?.metric_ids || templateSheets(tpl).map((s) => s.metric_id)).join(" ");
       return `${j.name} ${j.mode} ${j.status} ${tpl?.name || ""} ${mids}`.toLowerCase().includes(kw);
     });
-  }, [jobs, templates, debouncedListKw]);
+  }, [jobs, templatesById, debouncedListKw]);
 
   const filteredTemplates = useMemo(() => {
     const kw = debouncedListKw.trim().toLowerCase();
@@ -1135,10 +1159,14 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
       const next = currentlyOn.includes(sid)
         ? currentlyOn.filter((id) => id !== sid)
         : [...currentlyOn, sid];
+      if (!next.length) {
+        showError(t("bizCompare.needSheets"));
+        return;
+      }
       // Empty list means "all on" (new template sheets auto-included)
       setEnabledSheetIds(next.length === all.length ? [] : next);
     },
-    [jobSheetAllIds, enabledSheetIds],
+    [jobSheetAllIds, enabledSheetIds, showError, t],
   );
 
   const setJobTemplateAndSheets = useCallback(
@@ -1195,28 +1223,18 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
     () => runSheets.find((s) => sheetIdentity(s) === resultSheetId) || runSheets[0] || null,
     [runSheets, resultSheetId],
   );
+  const activeRunSheetId = sheetIdentity(activeRunSheet);
+  const activeRunSheetStatus = String(activeRunSheet?.status || "");
 
   // Clear field filters when switching sheet
   useEffect(() => {
     setResultKeyFilters({});
   }, [resultSheetId]);
 
-  // Reset page when sheet / filter / page size changes
-  useEffect(() => {
-    setResultPage(1);
-  }, [
-    resultSheetId,
-    kindFilter,
-    debouncedResultKw,
-    debouncedKeyFiltersJson,
-    resultPageSize,
-    runDetail?.id,
-  ]);
-
   useEffect(() => {
     const runId = String(runDetail?.id || "");
-    const mid = resultSheetId || sheetIdentity(activeRunSheet) || "";
-    const sheetSt = String(activeRunSheet?.status || "");
+    const mid = resultSheetId || activeRunSheetId || "";
+    const sheetSt = activeRunSheetStatus;
     // Block only while *this* sheet is still in flight — done sheets are readable mid-run
     const sheetStillRunning = ["pending", "running", "queued"].includes(sheetSt);
     if (!runId || !mid || jobDetailTab !== "result" || sheetStillRunning) {
@@ -1224,11 +1242,30 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
       setResultTotal(0);
       setDiffsSource("");
       setDiffsTruncated(false);
+      setDiffsLoading(false);
+      setDiffsError("");
       return;
     }
+    const filterKey = JSON.stringify([runId, mid, kindFilter, debouncedResultKw,
+      debouncedKeyFiltersJson, resultPageSize]);
+    if (diffFiltersRef.current !== filterKey) {
+      diffFiltersRef.current = filterKey;
+      if (resultPage !== 1) {
+        setPagedDiffs([]);
+        setResultTotal(0);
+        setDiffsLoading(true);
+        setResultPage(1);
+        return;
+      }
+    }
     let cancelled = false;
     void (async () => {
       setDiffsLoading(true);
+      setDiffsError("");
+      setPagedDiffs([]);
+      setResultTotal(0);
+      setDiffsSource("");
+      setDiffsTruncated(false);
       try {
         const res = await bizCompareListRunDiffs({
           runId,
@@ -1253,8 +1290,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
         if (resultPage > pages) setResultPage(pages);
       } catch (e) {
         if (!cancelled) {
-          // Keep previous rows to avoid strip/table jump; only clear on hard empty run
-          showError(formatErr(e));
+          setDiffsError(formatErr(e));
         }
       } finally {
         if (!cancelled) setDiffsLoading(false);
@@ -1267,15 +1303,16 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
     runDetail?.id,
     runDetail?.status,
     resultSheetId,
-    activeRunSheet?.metric_id,
-    activeRunSheet?.status,
+    activeRunSheetId,
+    activeRunSheetStatus,
     kindFilter,
     debouncedResultKw,
     debouncedKeyFiltersJson,
+    debouncedKeyFilters,
     resultPage,
     resultPageSize,
     jobDetailTab,
-    showError,
+    diffsRetry,
   ]);
 
   const resultColumns = useMemo(() => {
@@ -1991,9 +2028,12 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
   };
 
   const openJob = async (id: string) => {
+    const request = ++jobRequestRef.current;
+    runRequestRef.current += 1;
     setJobId(id);
     setJobDetailTab("config");
     setRunDetail(null);
+    setRuns([]);
     setResultSheetId("");
     setKindFilter("diff");
     setResultKw("");
@@ -2001,20 +2041,24 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
     if (job) resetJobForm(job);
     try {
       const r = await bizCompareListRuns(id, 50);
+      if (request !== jobRequestRef.current) return;
       const items = r.items || [];
       setRuns(items);
       if (items.length) {
         const latest = await bizCompareGetRun(String((items as any[])[0].id));
+        if (request !== jobRequestRef.current) return;
         setRunDetail(latest);
         // Land on batch list so stuck/cancelled runs are visible and actionable
         setJobDetailTab("runs");
       }
     } catch (e) {
-      showError(formatErr(e));
+      if (request === jobRequestRef.current) showError(formatErr(e));
     }
   };
 
   const closeJob = () => {
+    jobRequestRef.current += 1;
+    runRequestRef.current += 1;
     if (document.fullscreenElement === boardRef.current) {
       void document.exitFullscreen().catch(() => undefined);
     }
@@ -2061,20 +2105,26 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
       setJobDetailTab("runs");
       return;
     }
+    const request = jobRequestRef.current;
     setBusy(true);
     try {
       await bizCompareUpdateJob(jobId, jobConfigBody());
       // Async enqueue — returns immediately with status=running; poll below.
       const run = await bizCompareRunJob(jobId);
+      if (request !== jobRequestRef.current) return;
+      runRequestRef.current += 1;
       setRunDetail(run);
       setJobDetailTab("runs");
       showOk(t("bizCompare.runStarted"));
       const r = await bizCompareListRuns(jobId, 50);
+      if (request !== jobRequestRef.current) return;
       setRuns(r.items || []);
       await refresh({ force: true });
     } catch (e) {
-      showError(formatErr(e));
-      setJobDetailTab("runs");
+      if (request === jobRequestRef.current) {
+        showError(formatErr(e));
+        setJobDetailTab("runs");
+      }
     } finally {
       setBusy(false);
     }
@@ -2082,20 +2132,25 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
   const runProgress = (runDetail?.summary?.progress || {}) as RunProgressInfo;
   const runEngine = String(runProgress.engine || "").toLowerCase();
 
-  // Poll active compare runs so the modal can be closed and reopened safely.
+  // One poller; schedule the next tick after requests settle to prevent overlap.
   useEffect(() => {
-    if (!runIsActive || !runDetail?.id) return;
+    if (!jobId || (!runIsActive && !(jobHasActiveRun && jobDetailTab === "runs"))) return;
     let cancelled = false;
     let notified = false;
+    let timer: number | undefined;
+    const jobRequest = jobRequestRef.current;
+    const runId = String(runDetail?.id || "");
     const tick = async () => {
+      const runRequest = runRequestRef.current;
       try {
-        const d = await bizCompareGetRun(String(runDetail.id));
-        if (cancelled) return;
-        setRunDetail(d);
-        if (jobId) {
-          const r = await bizCompareListRuns(jobId);
-          if (!cancelled) setRuns(r.items || []);
-        }
+        const [r, d] = await Promise.all([
+          bizCompareListRuns(jobId, 50),
+          runIsActive && runId ? bizCompareGetRun(runId) : Promise.resolve(null),
+        ]);
+        if (cancelled || jobRequest !== jobRequestRef.current) return;
+        setRuns(r.items || []);
+        if (!d || runRequest !== runRequestRef.current) return;
+        setRunDetail((prev: typeof runDetail) => String(prev?.id || "") === runId ? d : prev);
         const st = String(d.status || "");
         if (!notified && st === "success") {
           notified = true;
@@ -2111,62 +2166,45 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
         }
       } catch (e) {
         if (!cancelled) showError(formatErr(e));
+      } finally {
+        if (!cancelled) timer = window.setTimeout(() => void tick(), 2000);
       }
     };
     void tick();
-    const id = window.setInterval(() => void tick(), 2000);
     return () => {
       cancelled = true;
-      window.clearInterval(id);
+      window.clearTimeout(timer);
     };
-  }, [runDetail?.id, runIsActive, jobId, showOk, showError, t]);
-
-  // Keep the runs list fresh while any batch on this job is active
-  useEffect(() => {
-    if (!jobId || !jobHasActiveRun || jobDetailTab !== "runs") return;
-    let cancelled = false;
-    const tick = async () => {
-      try {
-        const r = await bizCompareListRuns(jobId, 50);
-        if (!cancelled) setRuns(r.items || []);
-        const curId = String(runDetail?.id || "");
-        if (curId) {
-          const d = await bizCompareGetRun(curId);
-          if (!cancelled) setRunDetail(d);
-        }
-      } catch {
-        /* ignore list poll errors */
-      }
-    };
-    void tick();
-    const id = window.setInterval(() => void tick(), 2500);
-    return () => {
-      cancelled = true;
-      window.clearInterval(id);
-    };
-  }, [jobId, jobHasActiveRun, jobDetailTab, runDetail?.id]);
+  }, [runDetail?.id, runIsActive, jobId, jobHasActiveRun, jobDetailTab, showOk, showError, t]);
 
   const loadRun = async (runId: string) => {
+    const request = ++runRequestRef.current;
+    const jobRequest = jobRequestRef.current;
     try {
       const d = await bizCompareGetRun(runId);
+      if (request !== runRequestRef.current || jobRequest !== jobRequestRef.current) return;
       setRunDetail(d);
       setKindFilter("diff");
       setResultKw("");
+      setResultKeyFilters({});
       setJobDetailTab("result");
     } catch (e) {
-      showError(formatErr(e));
+      if (request === runRequestRef.current && jobRequest === jobRequestRef.current) showError(formatErr(e));
     }
   };
 
   const cancelRun = async (runId: string) => {
     if (!runId) return;
     if (!window.confirm(t("bizCompare.confirmCancelRun"))) return;
+    const request = jobRequestRef.current;
     setBusy(true);
     try {
       const d = await bizCompareCancelRun(runId);
-      if (String(runDetail?.id || "") === runId) setRunDetail(d);
+      if (request !== jobRequestRef.current) return;
+      setRunDetail((prev: typeof runDetail) => String(prev?.id || "") === runId ? d : prev);
       if (jobId) {
         const r = await bizCompareListRuns(jobId, 50);
+        if (request !== jobRequestRef.current) return;
         setRuns(r.items || []);
       }
       showOk(t("bizCompare.runCancelled"));
@@ -2180,19 +2218,24 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
   const removeRun = async (runId: string) => {
     if (!runId) return;
     if (!window.confirm(t("bizCompare.confirmDeleteRun"))) return;
+    const request = jobRequestRef.current;
+    const runRequest = runRequestRef.current;
     setBusy(true);
     try {
       const wasCurrent = String(runDetail?.id || "") === runId;
       await bizCompareDeleteRun(runId);
+      if (request !== jobRequestRef.current) return;
       let nextRuns: typeof runs = [];
       if (jobId) {
         const r = await bizCompareListRuns(jobId, 50);
+        if (request !== jobRequestRef.current) return;
         nextRuns = r.items || [];
       } else {
         nextRuns = (runs || []).filter((r) => String(r.id) !== runId);
       }
       setRuns(nextRuns);
-      if (wasCurrent) {
+      if (wasCurrent && runRequest === runRequestRef.current) {
+        runRequestRef.current += 1;
         setRunDetail(null);
         setResultSheetId("");
         setJobDetailTab("runs");
@@ -2542,7 +2585,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
               </thead>
               <tbody>
                 {filteredJobs.map((j) => {
-                  const tpl = templates.find((x) => x.id === j.template_id);
+                  const tpl = templatesById.get(j.template_id);
                   const n = templateSheets(tpl).length;
                   return (
                     <tr key={j.id}>
@@ -2882,7 +2925,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                                         : ""
                                     }
                                     placeholder={cmpMode === "percent" ? "%" : ""}
-                                    isDisabled={!isCompare || !needsTol}
+                                    disabled={!isCompare || !needsTol}
                                     onChange={(e) => {
                                       const raw = e.target.value.trim();
                                       const tol = raw === "" ? undefined : Number(raw);
@@ -3113,7 +3156,12 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
               className={jobDetailTab === "runs" ? "is-active" : undefined}
               onPress={() => {
                 setJobDetailTab("runs");
-                if (jobId) void bizCompareListRuns(jobId, 50).then((r) => setRuns(r.items || []));
+                const request = jobRequestRef.current;
+                if (jobId) void bizCompareListRuns(jobId, 50).then((r) => {
+                  if (request === jobRequestRef.current) setRuns(r.items || []);
+                }).catch((e) => {
+                  if (request === jobRequestRef.current) showError(formatErr(e));
+                });
               }}
             >
               {t("bizCompare.tabRuns")}
@@ -3354,6 +3402,18 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                       })()}
                     </div>
                   ) : null}
+                  {runDetail ? (
+                    <span
+                      className={`bs-cmp-board__config${runDetail.config_snapshot_version ? "" : " is-legacy"}`}
+                      title={t(runDetail.config_snapshot_version
+                        ? "bizCompare.configSnapshotHint"
+                        : "bizCompare.configLegacyHint")}
+                    >
+                      {t(runDetail.config_snapshot_version
+                        ? "bizCompare.configSnapshot"
+                        : "bizCompare.configLegacy")}
+                    </span>
+                  ) : null}
                   {!boardFs ? (
                     <>
                       <select
@@ -3757,6 +3817,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                             className={`bs-cmp-strip__kind bs-cmp-strip__kind--${cls}${
                               kindFilter === id ? " is-active" : ""
                             }`}
+                            aria-pressed={kindFilter === id}
                             onClick={() => setKindFilter(id as KindFilter)}
                           >
                             {id === "diff"
@@ -3782,16 +3843,22 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                       <div className="bs-cmp-filter-bar">
                         <Input
                           value={resultKw}
+                          aria-label={t("bizCompare.resultFilterPh")}
                           placeholder={t("bizCompare.resultFilterPh")}
                           onChange={(e) => setResultKw(e.target.value)}
                         />
-                        <span className="muted bs-sheet-count">
-                          {diffsLoading ? "…" : `${pagedDiffs.length}/${resultTotal}`}
+                        <span className="muted bs-sheet-count" role="status" aria-live="polite">
+                          {diffsLoading ? "…" : t("bizCompare.resultRange", {
+                            from: resultTotal ? (resultPage - 1) * resultPageSize + 1 : 0,
+                            to: Math.min(resultPage * resultPageSize, resultTotal),
+                            total: resultTotal,
+                          })}
                         </span>
                         {resultSearchKeyFields.length ? (
                           <button
                             type="button"
                             className={`bs-cmp-strip__toggle${keyFiltersVisible ? " is-active" : ""}`}
+                            aria-expanded={keyFiltersVisible}
                             onClick={() => setKeyFiltersOpen((v) => !v)}
                           >
                             {t("bizCompare.keyFiltersToggle")}
@@ -3809,6 +3876,14 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                         >
                           {t("bizCompare.displayColsToggle")}
                         </button>
+                      {(resultKw.trim() || activeKeyFilterCount > 0) ? (
+                        <Button size="sm" variant="ghost" onPress={() => {
+                          setResultKw("");
+                          setResultKeyFilters({});
+                        }}>
+                          {t("bizCompare.clearSearch")}
+                        </Button>
+                      ) : null}
                       </div>
                       {resultSearchKeyFields.length && keyFiltersVisible ? (
                         <div className="bs-cmp-key-filters">
@@ -3856,10 +3931,10 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                       ) : null}
                       {!isLiveSearch &&
                       kindFilter === "unchanged" &&
-                      (runDetail?.summary?.unchanged_truncated ||
+                      (activeRunSheet?.summary?.unchanged_truncated ||
                         (Number(
-                          (activeRunSheet?.summary as any)?.unchanged ||
-                            runDetail?.summary?.unchanged ||
+                          (activeRunSheet?.summary as any)?.unchanged ??
+                            runDetail?.summary?.unchanged ??
                             0,
                         ) >
                           Number(
@@ -3899,6 +3974,7 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                       className={`pt-list-table-wrap bs-sheet-table bs-cmp-result-table${
                         diffsLoading ? " is-loading" : ""
                       }`}
+                      aria-busy={diffsLoading}
                     >
                       <table className="data-table pt-list-table bs-cmp-diff-table">
                         <thead>
@@ -4004,6 +4080,22 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                           })()}
                         </thead>
                         <tbody>
+                          {diffsLoading || diffsError ? (
+                            <tr>
+                              <td colSpan={resultEmptyColSpan}>
+                                <div className={`bs-cmp-result-state${diffsError ? " is-error" : ""}`}
+                                  role={diffsError ? "alert" : "status"}>
+                                  <strong>{t(diffsError ? "bizCompare.resultLoadFailed" : "bizCompare.resultLoading")}</strong>
+                                  {diffsError ? <>
+                                    <span>{diffsError}</span>
+                                    <Button size="sm" variant="secondary" onPress={() => setDiffsRetry((n) => n + 1)}>
+                                      {t("bizCompare.retry")}
+                                    </Button>
+                                  </> : null}
+                                </div>
+                              </td>
+                            </tr>
+                          ) : null}
                           {pagedDiffs.map((d, i) => {
                             const pre = pickSideRow(
                               d.mapped_before as Record<string, unknown> | null | undefined,
@@ -4081,13 +4173,13 @@ export function BizComparePage({ pageMode = "all" }: { pageMode?: BizComparePage
                               </tr>
                             );
                           })}
-                          {runDetail && !diffsLoading && !pagedDiffs.length ? (
+                          {runDetail && !diffsLoading && !diffsError && !pagedDiffs.length ? (
                             <tr>
                               <td colSpan={resultEmptyColSpan}>
                                 <div className="pt-list-empty">
-                                  {kindFilter === "unchanged"
-                                    ? Number(runDetail?.summary?.unchanged || 0) > 0 &&
-                                      !Number(runDetail?.summary?.unchanged_listed || 0)
+                                  {hasResultSearch ? t("bizCompare.searchEmpty") : kindFilter === "unchanged"
+                                    ? Number(activeRunSheet?.summary?.unchanged || 0) > 0 &&
+                                      !Number(activeRunSheet?.summary?.unchanged_listed || 0)
                                       ? t("bizCompare.unchangedNotStored")
                                       : t("bizCompare.resultEmpty")
                                     : activeSheetPending
diff --git a/web/src/pages/network/cutoverDataCache.ts b/web/src/pages/network/cutoverDataCache.ts
index cc0748b..ff5e03b 100644
--- a/web/src/pages/network/cutoverDataCache.ts
+++ b/web/src/pages/network/cutoverDataCache.ts
@@ -7,6 +7,7 @@
 type Entry = { at: number; data: unknown };
 
 const store = new Map<string, Entry>();
+const inFlight = new Map<string, Promise<unknown>>();
 const DEFAULT_TTL_MS = 45_000;
 
 export async function cutoverCachedGet<T>(
@@ -21,9 +22,17 @@ export async function cutoverCachedGet<T>(
       return hit.data as T;
     }
   }
-  const data = await fetcher();
-  store.set(key, { at: Date.now(), data });
-  return data;
+  const existing = inFlight.get(key);
+  if (!opts?.force && existing) return existing as Promise<T>;
+  const request = Promise.resolve().then(fetcher).then((data) => {
+    // An invalidation or forced refresh may supersede this request.
+    if (inFlight.get(key) === request) store.set(key, { at: Date.now(), data });
+    return data;
+  }).finally(() => {
+    if (inFlight.get(key) === request) inFlight.delete(key);
+  });
+  inFlight.set(key, request);
+  return request;
 }
 
 /** Return stale immediately (if any), refresh in background and notify. */
@@ -34,27 +43,28 @@ export async function cutoverCachedGetSWR<T>(
 ): Promise<T> {
   const hit = store.get(key);
   if (hit) {
-    void fetcher()
+    void cutoverCachedGet(key, fetcher, { ttlMs: 0 })
       .then((data) => {
-        store.set(key, { at: Date.now(), data });
-        onFresh?.(data);
+        if (store.get(key)?.data === data) onFresh?.(data);
       })
       .catch(() => {
         /* keep stale */
       });
     return hit.data as T;
   }
-  const data = await fetcher();
-  store.set(key, { at: Date.now(), data });
-  return data;
+  return cutoverCachedGet(key, fetcher);
 }
 
 export function invalidateCutoverCache(prefix = ""): void {
   if (!prefix) {
     store.clear();
+    inFlight.clear();
     return;
   }
   for (const k of [...store.keys()]) {
     if (k.startsWith(prefix)) store.delete(k);
   }
+  for (const k of [...inFlight.keys()]) {
+    if (k.startsWith(prefix)) inFlight.delete(k);
+  }
 }
diff --git a/web/tests/bizCompare.browser.mjs b/web/tests/bizCompare.browser.mjs
new file mode 100644
index 0000000..6d3f62d
--- /dev/null
+++ b/web/tests/bizCompare.browser.mjs
@@ -0,0 +1,133 @@
+// Run against a local Vite server. All API responses are fixtures; no device or DB is touched.
+// NETX_PLAYWRIGHT_MODULE can point at a bundled Playwright package.
+import assert from "node:assert/strict";
+import { createRequire } from "node:module";
+import { mkdir } from "node:fs/promises";
+import path from "node:path";
+
+const require = createRequire(import.meta.url);
+const { chromium } = require(process.env.NETX_PLAYWRIGHT_MODULE || "playwright");
+const base = process.env.NETX_TEST_URL || "http://127.0.0.1:5179";
+const output = path.resolve(process.env.NETX_TEST_OUTPUT || "../docs/reviews/assets");
+await mkdir(output, { recursive: true });
+const browser = await chromium.launch({ headless: true, channel: process.env.NETX_TEST_BROWSER || "chrome" });
+const page = await browser.newPage({ viewport: { width: 1440, height: 1000 } });
+const errors = [];
+page.on("pageerror", (error) => errors.push(error.message));
+const fields = ["neighbor", "vrf", "remote_as", "state"];
+const sheet = {
+  sheet_id: "peer", title: "BGP 邻居", metric_id: "bgp_peer",
+  key_fields: ["neighbor", "vrf"], iface_fields: [], compare_fields: ["remote_as", "state"],
+  status: "done", summary: { changed: 1, removed: 1, added: 1, unchanged: 2, unchanged_listed: 2 },
+};
+const side = (label, ne_name, time) => ({ label, ne_name, started_at: time, ne_ip: "10.0.0.1" });
+const run = {
+  id: "run-1", status: "success", job_id: "job-1", config_snapshot_version: 1, created_at: "2026-10-10T10:08:00Z",
+  before: side("before", "PE-BEFORE", "2026-10-10T09:00:00Z"),
+  after: side("after", "PE-AFTER", "2026-10-10T10:00:00Z"),
+  sheets: [sheet],
+  summary: { ...sheet.summary, before_count: 4, after_count: 4, pass_rate: 50, duration_ms: 1250,
+    sheet_cards: [{ ...sheet, ...sheet.summary, before_count: 4, after_count: 4, pass_rate: 50 }] },
+};
+const job = { id: "job-1", name: "割接前后 · 业务验收", template_id: "tpl", mode: "manual", status: "ready",
+  before_task_id: "t1", after_task_id: "t2", before_batch_id: "b", after_batch_id: "a" };
+const row = (neighbor, state, remote_as = "64512") => ({ neighbor, vrf: "core", state, remote_as });
+const diffs = [
+  { kind: "changed", key: { neighbor: "10.2.1.1", vrf: "core" },
+    before: row("10.2.1.1", "Established"), after: row("10.2.1.1", "Idle"),
+    changes: { state: { before: "Established", after: "Idle" } } },
+  { kind: "removed", key: { neighbor: "10.2.1.2", vrf: "core" }, before: row("10.2.1.2", "Established"), after: null, changes: {} },
+  { kind: "added", key: { neighbor: "10.2.1.3", vrf: "core" }, before: null, after: row("10.2.1.3", "Established"), changes: {} },
+  { kind: "unchanged", key: { neighbor: "10.2.1.4", vrf: "core" }, before: row("10.2.1.4", "Established"), after: row("10.2.1.4", "Established"), changes: {} },
+];
+let failDiffs = false;
+let active = false;
+let legacy = false;
+let pollCounts = { runs: 0, detail: 0 };
+await page.route("**/v1/**", async (route) => {
+  const url = new URL(route.request().url());
+  const p = url.pathname;
+  let data = {};
+  if (p === "/v1/auth/me") data = { user: { id: "test", username: "test", role: "admin", is_active: true } };
+  else if (p.endsWith("/compare/jobs")) data = { items: [job, { ...job, id: "job-slow", name: "慢任务" }] };
+  else if (p.endsWith("/compare/templates")) data = { items: [{ id: "tpl", name: "割接验收模板", metrics: [sheet] }] };
+  else if (p.endsWith("/compare/metrics")) data = { items: [{ metric_id: "bgp_peer", fields: fields.map((name) => ({ name, display_name: name, dtype: "string", is_key: sheet.key_fields.includes(name) })) }] };
+  else if (p.endsWith("/tasks")) data = { items: [{ id: "t1", ne_name: "PE-BEFORE", ne_ip: "10.0.0.1" }, { id: "t2", ne_name: "PE-AFTER", ne_ip: "10.0.0.2" }] };
+  else if (p.endsWith("/batches")) data = { items: [{ id: p.includes("t1") ? "b" : "a", status: "success", row_count: 4 }] };
+  else if (p.endsWith("/runs") && p.includes("/jobs/")) {
+    pollCounts.runs++;
+    if (p.includes("job-slow")) await new Promise((resolve) => setTimeout(resolve, 900));
+    data = { items: [{ ...run, status: active ? "running" : "success" }] };
+  } else if (p.endsWith("/diffs")) {
+    if (failDiffs) { await route.fulfill({ status: 503, contentType: "application/json", body: JSON.stringify({ detail: "测试：连接暂时中断" }) }); return; }
+    const kind = url.searchParams.get("kind") || "diff";
+    const kw = url.searchParams.get("kw") || "";
+    const items = diffs.filter((d) => (kind === "all" || (kind === "diff" ? ["removed", "changed"].includes(d.kind) : d.kind === kind)) && JSON.stringify(d).includes(kw));
+    data = { items, total: items.length, page: 1, page_size: 100, source: "stored", truncated: false };
+  } else if (p.endsWith("/runs/run-1")) {
+    pollCounts.detail++;
+    data = { ...run, config_snapshot_version: legacy ? null : 1, status: active ? "running" : "success" };
+  } else if (p.endsWith("/mappings")) data = { items: [] };
+  else data = { items: [] };
+  await route.fulfill({ status: 200, contentType: "application/json", body: JSON.stringify(data) });
+});
+
+try {
+  await page.goto(`${base}/network/cutover/biz-compare`);
+  await page.getByRole("row").filter({ hasText: "割接前后 · 业务验收" }).getByRole("button", { name: "详情", exact: true }).click();
+  await page.getByRole("button", { name: "比对结果", exact: true }).click();
+  await page.locator(".bs-cmp-diff-table tbody tr.bs-cmp-row").first().waitFor();
+  assert.equal(await page.locator(".bs-cmp-diff-table tbody tr.bs-cmp-row").count(), 2);
+  assert.equal(await page.locator(".bs-cmp-board__config").textContent(), "执行配置已保存");
+  await page.screenshot({ path: path.join(output, "biz-compare-desktop.png"), fullPage: true });
+
+  const search = page.locator(".bs-cmp-filter-bar input");
+  failDiffs = true;
+  await search.fill("故障");
+  await page.getByText("对比明细加载失败", { exact: true }).waitFor();
+  assert.equal(await page.locator(".bs-cmp-row").count(), 0);
+  await page.screenshot({ path: path.join(output, "biz-compare-error.png"), fullPage: true });
+  failDiffs = false;
+  await page.getByRole("button", { name: "重试", exact: true }).click();
+  await page.getByText("当前筛选条件下没有匹配结果，可清空筛选或切换结果类型。", { exact: true }).waitFor();
+  await page.getByRole("button", { name: "清空筛选", exact: true }).click();
+  await page.locator(".bs-cmp-row").first().waitFor();
+
+  await page.setViewportSize({ width: 768, height: 1000 });
+  await page.screenshot({ path: path.join(output, "biz-compare-narrow.png"), fullPage: true });
+  const fits = await page.locator(".bs-cmp-strip").evaluate((element) => element.scrollWidth <= element.clientWidth + 1);
+  assert.ok(fits, "filter toolbar must fit its available width");
+  await page.setViewportSize({ width: 390, height: 1000 });
+  assert.ok(await page.locator(".bs-cmp-strip").evaluate((element) => element.scrollWidth <= element.clientWidth + 1));
+  assert.ok(await page.locator(".bs-cmp-board__toolbar").evaluate((element) => element.scrollWidth <= element.clientWidth + 1));
+  await page.setViewportSize({ width: 768, height: 1000 });
+
+  await page.getByRole("button", { name: "关闭", exact: true }).click();
+  await page.getByRole("row").filter({ hasText: "慢任务" }).getByRole("button", { name: "详情", exact: true }).click();
+  await page.getByRole("button", { name: "关闭", exact: true }).click();
+  await page.getByRole("row").filter({ hasText: "割接前后 · 业务验收" }).getByRole("button", { name: "详情", exact: true }).click();
+  await page.getByRole("button", { name: "比对结果", exact: true }).click();
+  await page.locator(".bs-cmp-row").first().waitFor();
+  await page.waitForTimeout(1100);
+  assert.equal(await page.locator(".bs-cmp-row").count(), 2, "slow old job must not change the current tab");
+
+  legacy = true;
+  await page.reload();
+  await page.getByRole("row").filter({ hasText: "割接前后 · 业务验收" }).getByRole("button", { name: "详情", exact: true }).click();
+  await page.getByRole("button", { name: "比对结果", exact: true }).click();
+  await page.getByText("旧记录：未保存执行配置", { exact: true }).waitFor();
+  await page.screenshot({ path: path.join(output, "biz-compare-legacy.png"), fullPage: true });
+
+  active = true;
+  await page.reload();
+  await page.getByRole("row").filter({ hasText: "割接前后 · 业务验收" }).getByRole("button", { name: "详情", exact: true }).click();
+  await page.locator(".bs-cmp-runs").waitFor();
+  pollCounts = { runs: 0, detail: 0 };
+  await page.waitForTimeout(5200);
+  assert.ok(pollCounts.runs <= 3 && pollCounts.detail <= 3, JSON.stringify(pollCounts));
+  assert.ok(pollCounts.runs >= 2 && pollCounts.detail >= 2, "active run should keep polling");
+  assert.deepEqual(errors, []);
+  console.log(JSON.stringify({ passed: ["result layout", "saved execution settings", "legacy settings warning", "error clears stale rows", "retry", "clear filters", "768px toolbar", "old job response ignored", "single poller", "no runtime errors"], pollCounts, output }));
+} finally {
+  await browser.close();
+}
diff --git a/web/tests/cutoverDataCache.test.mjs b/web/tests/cutoverDataCache.test.mjs
new file mode 100644
index 0000000..8c12694
--- /dev/null
+++ b/web/tests/cutoverDataCache.test.mjs
@@ -0,0 +1,72 @@
+import assert from "node:assert/strict";
+import { beforeEach, test } from "node:test";
+import { cutoverCachedGet, cutoverCachedGetSWR, invalidateCutoverCache } from "../src/pages/network/cutoverDataCache.ts";
+
+const deferred = () => {
+  let resolve, reject;
+  const promise = new Promise((ok, fail) => { resolve = ok; reject = fail; });
+  return { promise, resolve, reject };
+};
+
+beforeEach(() => invalidateCutoverCache());
+
+test("concurrent readers share one fetch", async () => {
+  const pending = deferred();
+  let calls = 0;
+  const fetcher = () => { calls++; return pending.promise; };
+  const first = cutoverCachedGet("batches", fetcher);
+  const second = cutoverCachedGet("batches", fetcher);
+  pending.resolve("fresh");
+  assert.deepEqual(await Promise.all([first, second]), ["fresh", "fresh"]);
+  assert.equal(calls, 1);
+});
+
+test("invalidated pending requests cannot repopulate the cache", async () => {
+  const pending = deferred();
+  const old = cutoverCachedGet("batches", () => pending.promise);
+  invalidateCutoverCache("batches");
+  pending.resolve("stale");
+  await old;
+  assert.equal(await cutoverCachedGet("batches", async () => "fresh"), "fresh");
+});
+
+test("forced refresh supersedes a slow older request", async () => {
+  const pending = deferred();
+  const old = cutoverCachedGet("lists", () => pending.promise);
+  assert.equal(await cutoverCachedGet("lists", async () => "fresh", { force: true }), "fresh");
+  pending.resolve("old");
+  await old;
+  assert.equal(await cutoverCachedGet("lists", async () => "unexpected"), "fresh");
+});
+
+test("a failed fetch can be retried", async () => {
+  await assert.rejects(cutoverCachedGet("lists", async () => { throw new Error("offline"); }));
+  assert.equal(await cutoverCachedGet("lists", async () => "recovered"), "recovered");
+});
+
+test("SWR readers share background refresh and each receives the fresh value", async () => {
+  await cutoverCachedGet("lists", async () => "cached");
+  const pending = deferred();
+  let calls = 0;
+  const notifications = [];
+  const fetcher = () => { calls++; return pending.promise; };
+  const first = await cutoverCachedGetSWR("lists", fetcher, (x) => notifications.push(x));
+  const second = await cutoverCachedGetSWR("lists", fetcher, (x) => notifications.push(x));
+  assert.equal(first, "cached");
+  assert.equal(second, "cached");
+  pending.resolve("fresh");
+  await new Promise((resolve) => setImmediate(resolve));
+  assert.equal(calls, 1);
+  assert.deepEqual(notifications, ["fresh", "fresh"]);
+});
+
+test("invalidated SWR requests do not notify with stale data", async () => {
+  await cutoverCachedGet("lists", async () => "cached");
+  const pending = deferred();
+  const notifications = [];
+  await cutoverCachedGetSWR("lists", () => pending.promise, (x) => notifications.push(x));
+  invalidateCutoverCache("lists");
+  pending.resolve("stale");
+  await new Promise((resolve) => setImmediate(resolve));
+  assert.deepEqual(notifications, []);
+});