diff --git a/pipeline_bidding/bid_compose_capability.py b/pipeline_bidding/bid_compose_capability.py
index 4d9ca4e..7fdfe83 100644
--- a/pipeline_bidding/bid_compose_capability.py
+++ b/pipeline_bidding/bid_compose_capability.py
@@ -8,6 +8,7 @@ docx 生成用 python-docx；不可用时降级为 md（诚实降级，不假装
 import json
 import logging
 import os
+import re
 
 from .bid_common import (
     get_db, new_id, rec_to_dict, rows_to_dicts, to_float, to_int, record,
@@ -105,7 +106,6 @@ def _parse_md_image(line):
     """解析一行 markdown 图片语法，返回 (alt, url) 或 None。"""
     global _MD_IMG_RE
     if _MD_IMG_RE is None:
-        import re
         _MD_IMG_RE = re.compile(r'^\s*!\[([^\]]*)\]\(([^)\s]+)\)\s*$')
     m = _MD_IMG_RE.match(line or '')
     if not m:
@@ -113,6 +113,35 @@ def _parse_md_image(line):
     return m.group(1), m.group(2)
 
 
+_MD_TABLE_SEP_RE = re.compile(r'^\s*\|?[\s:|-]+\|[\s:|-]*$')
+
+
+def _is_table_sep(line):
+    """markdown 表头分隔行（|---|---| 或 |:--|--:|）。"""
+    s = (line or '').strip()
+    return bool(s) and '-' in s and set(s) <= set('|-: ')
+
+
+def _split_md_row(line):
+    """拆分一行 markdown 表格为单元格列表（去首尾竖线、按 | 切、strip）。"""
+    s = (line or '').strip()
+    if s.startswith('|'):
+        s = s[1:]
+    if s.endswith('|'):
+        s = s[:-1]
+    return [c.strip() for c in s.split('|')]
+
+
+def _clean_md_cell(text):
+    """清理单元格内联 markdown（**粗体**/`code`/图片占位）→ 纯文本。"""
+    t = text or ''
+    t = re.sub(r'!\[([^\]]*)\]\([^)]*\)', r'\1', t)   # 图片→alt
+    t = re.sub(r'\*\*([^*]+)\*\*', r'\1', t)          # 粗体
+    t = re.sub(r'`([^`]+)`', r'\1', t)                # 行内代码
+    t = re.sub(r'<br\s*/?>', ' ', t, flags=re.I)      # 换行标签
+    return t.strip()
+
+
 def _write_docx(path, title, subtitle, chapters):
     """用 python-docx 写标书。返回 (ok, page_hint_or_err)。
 
@@ -155,10 +184,45 @@ def _write_docx(path, title, subtitle, chapters):
         doc.add_page_break()
 
         img_ok = img_fail = 0
+        tbl_ok = 0
         for c in chapters:
             doc.add_heading("%s　%s" % (c.get("chapter_no", ""), c.get("title", "")), level=1)
-            for line in (c.get("content") or "").split("\n"):
-                t = line.rstrip()
+            _lines = (c.get("content") or "").split("\n")
+            _i = 0
+            _n = len(_lines)
+            while _i < _n:
+                t = _lines[_i].rstrip()
+                # ── markdown 表格块 → Word 表格（2026-09-16 用户报障：表格散成文字）──
+                # 识别：当前行含 '|' 且下一行是分隔行（|---|）。缓冲连续表格行整体渲染。
+                if '|' in t and _i + 1 < _n and _is_table_sep(_lines[_i + 1]):
+                    _rows = [_split_md_row(t)]
+                    _i += 2     # 跳过表头行 + 分隔行
+                    while _i < _n:
+                        _rt = _lines[_i].rstrip()
+                        if '|' not in _rt or not _rt.strip():
+                            break
+                        _rows.append(_split_md_row(_rt))
+                        _i += 1
+                    _ncol = max(len(r) for r in _rows)
+                    try:
+                        _tbl = doc.add_table(rows=0, cols=_ncol)
+                        _tbl.style = 'Table Grid'
+                        for _ri, _r in enumerate(_rows):
+                            _cells = _tbl.add_row().cells
+                            for _ci in range(_ncol):
+                                _txt = _clean_md_cell(_r[_ci] if _ci < len(_r) else '')
+                                _para = _cells[_ci].paragraphs[0]
+                                _run = _para.add_run(_txt)
+                                _run.font.size = Pt(9)
+                                if _ri == 0:
+                                    _run.bold = True
+                        tbl_ok += 1
+                    except Exception as e:
+                        doc.add_paragraph("［表格渲染失败：%s］" % str(e)[:80])
+                        for _r in _rows:
+                            doc.add_paragraph(" | ".join(_clean_md_cell(x) for x in _r))
+                    continue
+                _i += 1
                 if not t:
                     continue
                 img = _parse_md_image(t)
@@ -191,12 +255,25 @@ def _write_docx(path, title, subtitle, chapters):
                     doc.add_heading(t[2:], level=2)
                 elif t.lstrip().startswith(("- ", "* ")):
                     doc.add_paragraph(t.lstrip()[2:], style="List Bullet")
+                elif _is_table_sep(t):
+                    continue    # 游离分隔行（无表头）跳过
+                elif t.strip().startswith('|') and t.strip().endswith('|'):
+                    # 单行竖线内容但无分隔行跟随 → 按表格行渲染为单行表，避免散文字
+                    _cells1 = [_clean_md_cell(x) for x in _split_md_row(t)]
+                    try:
+                        _t1 = doc.add_table(rows=1, cols=max(1, len(_cells1)))
+                        _t1.style = 'Table Grid'
+                        for _ci, _cv in enumerate(_cells1):
+                            _t1.rows[0].cells[_ci].paragraphs[0].add_run(_cv).font.size = Pt(9)
+                        tbl_ok += 1
+                    except Exception:
+                        doc.add_paragraph(t)
                 else:
                     doc.add_paragraph(t)
             doc.add_page_break()
         doc.save(path)
-        if img_ok or img_fail:
-            logger.info("compose: docx 配图嵌入 成功%d 失败%d", img_ok, img_fail)
+        if img_ok or img_fail or tbl_ok:
+            logger.info("compose: docx 配图嵌入 成功%d 失败%d；表格%d个", img_ok, img_fail, tbl_ok)
         return True, ""
     except Exception as e:
         return False, "%s: %s" % (type(e).__name__, str(e)[:200])
diff --git a/pipeline_bidding/bid_review_capability.py b/pipeline_bidding/bid_review_capability.py
index f45898d..c3e296b 100644
--- a/pipeline_bidding/bid_review_capability.py
+++ b/pipeline_bidding/bid_review_capability.py
@@ -79,6 +79,20 @@ async def review_chapter(chapter_id, score, max_score="", comments="", detail=""
                        "生成真实图片，产物 URL 以 ![图N 标题](URL) 独占一行嵌入；"
                        "删除 mermaid/ASCII 字符画后重写提交。" % "；".join(sorted(set(_dg))))
             comments = (dg_note + "\n" + (comments or "")).strip()
+
+        # ── 实质性响应门禁（2026-09-16 用户报障根治）：指南体/表单大面积待填必退 ──
+        # submit_chapter 已拦一道；此处兜底存量正文/绕行写入，章节评审级即给反馈
+        # （不必等整书评分），意见自动附带。确定性检测，不采信评分器满分。
+        try:
+            from pipeline_service.substance_gate import find_substance_problems, gate_message
+            _sg = find_substance_problems(c.get("content") or "")
+            _sg_note = gate_message(_sg, scene="本章") if _sg else ""
+        except ImportError:
+            _sg_note = ""
+        if _sg_note:
+            passed = False
+            comments = (_sg_note + "\n" + (comments or "")).strip()
+        _substance_fail = bool(_sg_note)   # 实质门禁不通过 → 禁止 material_missing 豁免绕过
         if not passed and not (comments or "").strip():
             return False, ("评审不通过必须给出改进意见（comments），"
                            "写清缺哪些评分维度、应补什么内容、对照哪条得分规则")
@@ -121,6 +135,14 @@ async def review_chapter(chapter_id, score, max_score="", comments="", detail=""
         # 未回传 → 章节通过（真实分数+材料缺失披露注记，整书评分如实反映），
         # 同时发 material_supply 信息流待办通知人类补料，不冻结任何流转。
         _mat_missing = str(material_missing).strip().lower() in ("1", "true", "yes", "y")
+        # 守卫（2026-09-16）：实质门禁不通过的章节禁止用 material_missing 豁免绕过——
+        # 「指南体/表单大面积待填」不是实体材料缺失，是写作本身没写实质内容；
+        # 若放行会把我上一轮加的占位符豁免变成质量门禁的后门（实测满分放行的帮凶）。
+        if _mat_missing and _substance_fail:
+            _mat_missing = False
+            comments = ("【material_missing 豁免被机制拒绝】本章未通过实质性响应门禁"
+                        "（是指南体/表单待填，非实体材料缺失），不适用材料缺失豁免通道，"
+                        "必须按实质门禁意见重写正文。\n" + (comments or "")).strip()
         if _mat_missing:
             await sor.sqlExe(
                 "UPDATE " + TABLE + " SET status=${st}$, review_score=${sc}$, max_score=${ms}$, "
diff --git a/pipeline_bidding/bid_score_capability.py b/pipeline_bidding/bid_score_capability.py
index cc394e0..078817b 100644
--- a/pipeline_bidding/bid_score_capability.py
+++ b/pipeline_bidding/bid_score_capability.py
@@ -200,6 +200,30 @@ async def finalize_bid_score(review_id, comments="", who=None, agent_id=None):
         ratio = (total / maxsum) if maxsum > 0 else 0.0
         passed = (ratio >= th["bid_pass_ratio"]) and not veto_fail
 
+        # ── 实质性响应确定性门禁（2026-09-16 用户报障根治，最关键的一道）──
+        # 实测：指南体标书（投标函全是【待填】、实施方案写「本章覆盖评分子项①②③」）
+        # 被 LLM 评分器判 100/100 满分放行——评分报告自己都列了「无效标」却 passed。
+        # 根治：质量闸门不靠 LLM 自觉，机制层用 substance_gate 确定性扫全书章节正文，
+        # 命中指南体/表单大面积待填 → 强制 passed=False，命中章节退回重写。
+        # 与 diagram_gate 同一设计哲学（花钱/交付闸门必须确定性校验兜底）。
+        substance_chs = []          # [(chapter_id, chapter_no, title, problems)]
+        try:
+            from pipeline_service.substance_gate import find_substance_problems
+            _chrecs = await sor.sqlExe(
+                "SELECT id, chapter_no, title, content FROM bid_chapters "
+                "WHERE project_id=${p}$ ORDER BY order_no, chapter_no", {"p": pid})
+            await sor.sqlExe("COMMIT", {})
+            for _c in (_chrecs or []):
+                _cd = rec_to_dict(_c)
+                _probs = find_substance_problems(_cd.get("content") or "")
+                if _probs:
+                    substance_chs.append((_cd.get("id"), _cd.get("chapter_no"),
+                                          _cd.get("title"), _probs))
+        except ImportError:
+            pass    # 宿主未装 pipeline-service 时跳过（逃逸阀）
+        if substance_chs:
+            passed = False
+
         weak = [s for s in scores
                 if to_float(s.get("max_score"), 0.0) > 0
                 and to_float(s.get("score"), 0.0) / to_float(s.get("max_score"), 1.0) < th["bid_pass_ratio"]]
@@ -207,6 +231,13 @@ async def finalize_bid_score(review_id, comments="", who=None, agent_id=None):
                          % (total, maxsum, ratio * 100, th["bid_pass_ratio"] * 100)]
         if veto_fail:
             summary_lines.append("⛔ 否决项未响应：" + "、".join(veto_fail[:5]))
+        if substance_chs:
+            summary_lines.append("⛔ 实质性响应门禁不通过：%d 个章节是「编写指南体/表单大面积待填」"
+                                 "（不是投标文件正文本身），机制确定性拦截（不采信评分满分）："
+                                 % len(substance_chs))
+            for _cid, _cno, _ct, _probs in substance_chs[:12]:
+                summary_lines.append("  - 第%s章《%s》：%s"
+                                     % (_cno, (_ct or "")[:24], "；".join(_probs)[:200]))
         for s in weak[:20]:
             summary_lines.append("- %s %.1f/%.1f → %s"
                                  % (s.get("item_name"), to_float(s.get("score")),
@@ -260,10 +291,43 @@ async def finalize_bid_score(review_id, comments="", who=None, agent_id=None):
 
         # 把改进意见落到问题章节：approved → rejected（reconciler 会派重写任务）
         touched = 0
+        # ① 实质性门禁命中章节强制退回（评分器可能给它们满分，但机制判定是指南体/待填）
+        for _cid, _cno, _ct, _probs in substance_chs:
+            if not _cid:
+                continue
+            _crec = await sor.sqlExe(
+                "SELECT id, status, revise_count FROM bid_chapters WHERE id=${c}$", {"c": _cid})
+            await sor.sqlExe("COMMIT", {})
+            if not _crec:
+                continue
+            _c = rec_to_dict(_crec[0])
+            if _c.get("status") == CH_REJECTED:
+                continue    # 已退回，避免 revise_count 重复累加
+            _imp = ("【实质性响应门禁 · 第%s轮 · 机制确定性拦截】本章是「编写指南/作业指导书」"
+                    "或表单大面积未填实，不是投标文件正文本身。检出：%s\n"
+                    "必须重写为正文：① 删除全部「编制说明/编制依据/填写口径/自查表/人工填写清单」"
+                    "等讲『怎么写』的元内容；② 投标函/承诺/报价按招标格式**直接填真实内容**"
+                    "（投标人/报价/日期等能从项目或知识库取到的填实，实体材料缺失用规范占位符"
+                    "【材料待补：具体材料名】，禁止【待填-xxx】自造占位）；③ 技术/实施方案写实质"
+                    "措施（架构/流程/量化指标/责任人/工期），不写「覆盖评分子项①②③」式说明。"
+                    % (review.get("round"), "；".join(_probs)[:600]))
+            await sor.sqlExe(
+                "UPDATE bid_chapters SET status=${st}$, review_comment=${cm}$, "
+                "revise_count=${rc}$, updated_at=NOW() WHERE id=${c}$",
+                {"st": CH_REJECTED, "cm": _imp[:60000],
+                 "rc": to_int(_c.get("revise_count"), 0) + 1, "c": _cid})
+            await sor.sqlExe("COMMIT", {})
+            await record(sor, pid, "bid_chapters", _cid, "substance_reject",
+                         from_state=_c.get("status"), to_state=CH_REJECTED,
+                         who=who, agent_id=agent_id,
+                         detail="实质性门禁退回：%s" % ("；".join(_probs))[:300])
+            touched += 1
+        # ② 评分不达标章节退回（原有逻辑）
+        _substance_cids = set(_cid for _cid, _, _, _ in substance_chs if _cid)
         for s in weak:
             cid = s.get("target_chapter_id") or ""
-            if not cid:
-                continue
+            if not cid or cid in _substance_cids:
+                continue    # 已被 substance 门禁退回，避免 revise_count 重复累加
             crec = await sor.sqlExe(
                 "SELECT id, status, revise_count, title FROM bid_chapters WHERE id=${c}$",
                 {"c": cid})
diff --git a/pipeline_bidding/bid_write_capability.py b/pipeline_bidding/bid_write_capability.py
index 08d484f..725b17d 100644
--- a/pipeline_bidding/bid_write_capability.py
+++ b/pipeline_bidding/bid_write_capability.py
@@ -169,6 +169,18 @@ async def submit_chapter(chapter_id, note="", who=None, agent_id=None):
         d = rec_to_dict(recs[0])
         if len((d.get("content") or "").strip()) < MIN_CONTENT_LEN:
             return False, "章节正文为空或过短，请先 write_chapter 写入正文"
+        # ── 实质性响应硬门禁（2026-09-16 用户报障根治）──
+        # 拦「编写指南体/表单大面积待填」——交付的必须是投标文件正文本身，不是
+        # 「教人怎么写标书」的元内容（编制说明/填写口径/自查表/【待填-xxx】占位）。
+        # 确定性检测（substance_gate），不靠评审 LLM 自觉（实测指南体曾拿 100/100 满分）。
+        # 挂在 submit_chapter（定稿提交唯一关口）：append 分批写中途不误判，提交时查全文。
+        try:
+            from pipeline_service.substance_gate import find_substance_problems, gate_message
+            _sg = find_substance_problems(d.get("content") or "")
+            if _sg:
+                return False, gate_message(_sg, scene="章节「%s」" % (d.get("title") or chapter_id))
+        except ImportError:
+            pass  # 宿主未装 pipeline-service 时放行（逃逸阀，与产线零接触分层一致）
         if d.get("status") not in (CH_WRITING, CH_PENDING, CH_REJECTED):
             return False, "当前状态 %s 不可提交（需 writing/pending/rejected）" % d.get("status")
         await sor.sqlExe(