@dsh-cove/dsh-web-all 0.0.0-stage → 0.4.3

package/presets/code-vibe/skills/wish-programming/references/templates/requirements.md.tpl

@@ -0,0 +1,31 @@
+# <wish id> · 需求清单
+
+> 阶段 1 产出（1 个 pro agent，不拆）。每条需求/存量功能有编号，后续所有环节对照本文件。
+> 来源：`docs/wish/<id>/quote.md` 用户原话（最高优先级，逐字转录，不可改写）+ 用户口述补充。
+> **quote 版本：vN**（与 quote.md 头部版本一致。quote.md 更新到 vN+1 时，本节必须重新逐字转录并对齐版本；版本不齐 = stale，check_spec 检出。）
+
+## 0. 用户原话
+
+（从 quote.md **逐字**复制到这里，一字不改。本节与全文任何其他节冲突时以本节为准；冲突不得静默裁决，须回用户确认。）
+
+## 1. 需求清单（逐条编号，N 前缀）
+
+| 编号 | 需求（用户语言，禁用工程词汇） | 原话依据（引用原话哪一句） | 验收落点（哪个页面/交互，阶段 4 回填） |
+|---|---|---|---|
+| N1 | | | |
+
+## 2. 存量功能清单（逐条编号，O 前缀）
+
+> 升级/迁移/重设计/重构类愿望**必填**：盘点旧页面每一个用户可见功能点。全新建设类愿望写 N/A + 理由。
+
+| 编号 | 存量功能（旧页面路径 + 功能点，逐条） | 处置（保留到哪 / 合并到哪 / 删除——删除必须用户确认） | 新落点（阶段 4 回填） |
+|---|---|---|---|
+| O1 | | | |
+
+## 3. 澄清问题（≤3 个，问用户；无则写「无」）
+
+1. ……
+
+## 4. 语义不变量
+
+- （≥1 条；「存量功能零丢失」或显式豁免必须在此——豁免也要用户确认过）

package/presets/code-vibe/skills/wish-programming/references/workflow-template.md

@@ -0,0 +1,122 @@
+# Workflow 编排模板（并行构建 + 对抗验收 + 模型路由）
+
+> 给 conductor 的复用模板：把「建切片」和「独立验收」fan-out 给子代理。
+> 使用前提（2026-09-03 恢复 DSH 默认）：`workflow` 工具只在**用户显式点名 workflow** 时使用，不主动 fan-out。默认路径是 conductor 主上下文推进 + 浅层 subagent；本模板仅在用户点名 workflow 时按需套用。
+> 真源：DSH `workflow` 工具 + SKILL.md「Construction channel」+「模型路由与思考深度」。
+
+## 关键约束（先记住）
+
+- workflow 脚本**纯协调**：没有 filesystem / bash / 网络，只能 `agent()` / `pipeline()` / `parallel()` / `phase()` / `log()`。
+- 所以**闸门（journey.sh / verify.sh / replay.sh）必须由 conductor 在阶段之间用 bash 跑**，不能放进 workflow。
+- 每个 `agent()` 的 prompt 必须**自包含**（子代理看不到本对话）：只传**切片上下文块**（见下），不灌整份 contract/plan。验证者用 `agent()` 天然 fresh-context——**绝不用 fork 做验证**。
+- **模型路由（继承优先，同族启发）**：默认不传 `model`/`provider`。要换档就留在当前 provider、选同族兄弟（DeepSeek pro↔flash，Codex sol↔luna/terra），按任务轻重判断——不是死表。勿因目录热闹就跳到 Claude/GLM/Kimi；换族仅用户点名或任务明确需要。跨 provider 须同时传两者。不确定就继承。
+
+## 切片结构（conductor 读 plan.md 后作为 args 传入）
+
+```json
+{
+  "wish": "<id>",
+  "workspace": "/abs/path",
+  "slices": [
+    {
+      "id": "S1",
+      "entry": "…",
+      "oracle": "T-001, T-002",
+      "paths": ["src/a.ts"],
+      "boundary": "do not touch src/b/**; never edit the frozen oracle",
+      "context": "本切片相关契约事实：仅摘与本切片相关的入口/表/调用点（勿整份粘贴 contract.md）",
+      "deps": []
+    }
+  ]
+}
+```
+
+## 模板脚本（复制后按 wish 填 args）
+
+```js
+const { wish, workspace, slices } = args
+const contract = `${workspace}/docs/wish/${wish}/contract.md`
+const plan = `${workspace}/docs/wish/${wish}/plan.md`
+
+phase('build')
+const built = await pipeline(slices, async (prev, slice) => {
+  return agent(
+    `You are an implementer in a wish-programming flow. The block below is the ONLY spec you may act on. ` +
+    `Slice: ${slice.id} (${slice.entry}). Touched paths (hard allowlist): ${slice.paths.join(', ')}. ` +
+    `Boundary: ${slice.boundary}. Relevant contract facts: ${slice.context}. ` +
+    `Do not open other slices or the full contract/plan unless the block points to a section. ` +
+    `Run the slice's oracle (${slice.oracle}) and append one evidence row to ${workspace}/.sdd/acceptance/${wish}.tsv ` +
+    `with the v2 columns (test_id, kind, result, evidence_path, command, exit, output, ts). ` +
+    `Report the LIVE exit code + evidence path + output snapshot path.`,
+    { label: `build-${slice.id}`, phase: 'build' }  // omit model → inherit conductor provider/model
+  )
+})
+
+phase('verify')
+const verified = await pipeline(slices, async (prev, slice) => {
+  return agent(
+    `You are an INDEPENDENT reviewer in a FRESH context (never a fork): you did not implement ` +
+    `anything and never will edit any file. Do not trust the implementer's self-report. ` +
+    `Read ${contract} and the diff of ${slice.paths.join(', ')}. ` +
+    `Falsify slice ${slice.id} against the frozen oracle ${slice.oracle} by RUNNING the oracle ` +
+    `yourself — your evidence is the real command output, not anyone's claim. ` +
+    `Reject smoke/health/build-only evidence. Return a BOUNDED verdict: at most 3 concrete, ` +
+    `evidence-backed gaps — do not chase findings into over-engineering.`,
+    { label: `verify-${slice.id}`, phase: 'verify', schema: {
+      type: 'object',
+      properties: {
+        verdict: { type: 'string', enum: ['pass', 'fail'] },
+        reasons: { type: 'array', items: { type: 'string' } }
+      },
+      required: ['verdict']
+    } }
+  )
+})
+
+return { built, verified }
+```
+
+## 模型路由速查（软启发）
+
+| 切片性质 | 建议 |
+|---|---|
+| 默认 / 不确定 | 省略 `model`/`provider`（继承） |
+| 机械实施 / 证据收集 | 可同族快档；拿不准就继承 |
+| 语义敏感 / 评审 / 裁决 | 同族强档，或直接继承 |
+| 换到无关模型族 | 仅用户点名或任务明确需要该族 |
+
+同族示例：DeepSeek pro↔flash；Codex sol↔luna/terra。强档推理、快档窄 I/O；开放设计不要丢快档。
+
+## 并行 vs 串行
+
+- **并行**：`deps` 为空 且 `paths` 不重叠的切片。
+- **串行**：有依赖、改同一文件、或竞争同一数据库/端口/账号（见 workspace 并行规则）。
+- 一切片 = 一次 `agent()`；禁止把多片塞进一个 prompt。
+
+## 双相编排模式（2026-08-28 实测自「网关管理后台对抗验收」会话，pro+high 下 7 次 workflow 完整跑通）
+
+对任何「多切片建设/走查/修复」任务，用固定双相结构，并把结果结构化：
+
+```js
+// Phase 1 · 并行：无共享状态的切片全部 fan-out（走查/调研/修复）
+// Phase 2 · 串行：需要浏览器/同一端口/同一账号的操作一个接一个
+const findingSchema = {
+  type: 'object',
+  properties: {
+    slice: { type: 'string' },
+    verdict: { type: 'string', enum: ['pass', 'partial', 'fail'] },   // 三态判定
+    gaps: { type: 'array', items: { type: 'string' } },
+    evidence: { type: 'string' }                                       // 证据路径/命令输出快照
+  },
+  required: ['slice', 'verdict']
+}
+```
+
+启动前先 **probe 子代理可用性**（一个最小 `subagent` 探针，确认 spawn/工具/恢复链路正常再大规模派发）；子代理失败时 conductor **按失败原因重启该切片**，而不是回退串行。修复类任务用 fixSchema（agent + fixed 列表）让结果可机读。
+
+## 之后（conductor，不在 workflow 里跑）
+
+1. `replay.sh` 逐条红绿回放（基线红 → 改后绿）；
+2. `verify.sh --ui` 关版闸（证据成立才过；unit/integration/health 会被 verify **live 重跑**）；
+3. `journey.sh --approve --by <人>`（产品级人批注）→ `journey.sh --transition acceptance-passed`；
+4. 失败的切片：conductor 按 journey.sh 路由 repairing（必须 `--reason`），repair 预算 5 轮，耗尽转 `stuck`。

package/presets/code-vibe/skills/wish-programming/scripts/MANIFEST.sha256

@@ -0,0 +1,11 @@
+ff03eaf6f082ddf9e5c55ece2dfab9dbe8076208bbaf3841a95b568ddaa9256a  check_spec.py
+bf82defda0364e9c354927eac4122d0141903d3cde98e72189d8a9dbcd347bb1  check_whitelist.py
+b105c73226d3adceddf6dc04a1146f2f9208d321ae39c2449936e41ab201a149  coverage_check.py
+5f33af86b68e11d5cd29c76959ab1444e52dfa4bfbc5d3a0c20c290ec0817852  gate-lock.sh
+5822b898beb212bbc5fd6e9c5fc51eaa660bbdaa144be16f5921844e74ebd58d  journey.sh
+99ccacb30ec30a5394ffc27e2560efc57f01ec9a0ac5cbfd3e201508e2837ea3  project.sh
+c86fd0eb7b69199466d80b0968960f88bb38c7fc31b4b9f0fe3ef67f955765fb  replay.sh
+8e49467b1584ffd1e88d4ecb00d10a7d07eeef8f001fca5a5fbe24b28274ce8d  scan_drift.py
+b92784a68a33e944f59a1b49c12228384dcd4ceca163f46fea7fa7beb186c004  verify_preset_budgets.py
+4ce8883889b38d89321262f5c875e32ba15ba187e163d5b49cb709079e909acd  verify.sh
+82943a481aefcb128d21d481aa4a46eb10f3fcee69978fd155c63008566fbeeb  wish.sh

package/presets/code-vibe/skills/wish-programming/scripts/check_spec.py

@@ -0,0 +1,416 @@
+#!/usr/bin/env python3
+"""Product-aware spec-structure gate — v2 五阶段工件结构门（机械下限，不是质量）。
+
+只检查「产生真实质量后果」的工件结构：原话 / 需求清单 / 规范单 / 存量落点 / 截图。
+评审文档与状态机仪式不再检查。
+
+Phases:
+  --phase design   (阶段 1/2 门):  quote.md（原话，v1 保留）
+                    + requirements.md（需求清单 N 编号 + 存量功能清单 + 语义不变量）
+                    + design-spec.md（页面树 + 组件白名单 + 共享件清单）
+                    + product.md 条件式（存在才查「用户原话」节，不存在跳过）
+                    + confirmation-pack.md 禁用词汇表检查（若存在）
+  --phase plan     (技术架构类门):  arch.md 存在才查其结构 + review-arch.md 交锋段格式与原话对照；
+                    页面/前端类愿望无 arch.md → 跳过
+  --phase build    (默认):         上述全部 + plan.md 纵向切片；contract.md 条件式（存在才查事实映射）
+  --phase close    (acceptance-passed 门): 回填存在性检查（仓库感知）
+
+Usage:
+  python3 check_spec.py <workspace> <id> [--phase design|plan|build|close]
+Exit: 0 ok, 1 fail, 2 usage.
+"""
+import re
+import sys
+from pathlib import Path
+
+EMPTYISH = {"", "-", "—", "n/a", "N/A", "TODO", "TBD", "…", "...", "Verified/Unverified"}
+
+
+def clean(s):
+    return s.strip().strip("`").strip()
+
+
+def cells(line):
+    return [clean(c) for c in line.strip().strip("|").split("|")]
+
+
+def is_delim(row):
+    parts = [c for c in row if c]
+    return bool(parts) and all(re.fullmatch(r":?-{3,}:?", c) for c in parts)
+
+
+def table_after(text, heading):
+    lines = text.splitlines()
+    start = None
+    for i, l in enumerate(lines):
+        if l.startswith("#") and heading in l:
+            start = i + 1
+            break
+    if start is None:
+        return []
+    while start < len(lines) and ("|" not in lines[start] or lines[start].lstrip().startswith("#")):
+        if lines[start].lstrip().startswith("#"):
+            return []
+        start += 1
+    if start >= len(lines):
+        return []
+    headers = cells(lines[start])
+    start += 1
+    if start < len(lines) and is_delim(cells(lines[start])):
+        start += 1
+    rows = []
+    while start < len(lines) and "|" in lines[start]:
+        vals = cells(lines[start])
+        if len(vals) == len(headers):
+            rows.append({headers[i]: vals[i] for i in range(len(headers))})
+        start += 1
+    return rows
+
+
+def section_lines(text, heading):
+    """All non-blank body lines between the heading and the next heading."""
+    lines = text.splitlines()
+    out = []
+    started = False
+    for l in lines:
+        if l.startswith("#") and heading in l:
+            started = True
+            continue
+        if started:
+            if l.lstrip().startswith("#"):
+                break
+            if l.strip():
+                out.append(l.strip())
+    return out
+
+
+def section_exists(text, heading):
+    return any(l.startswith("#") and heading in l for l in text.splitlines())
+
+
+def filled(v):
+    v = clean(v)
+    return bool(v) and v not in EMPTYISH
+
+
+def review_pass(path):
+    if not path.is_file():
+        return False
+    lines = [l.strip() for l in path.read_text(encoding="utf-8").splitlines() if l.strip()]
+    if not lines:
+        return False
+    return bool(re.match(r"^结论\s*[:：]\s*pass$", lines[-1], re.IGNORECASE))
+
+
+def review_structured(path, name):
+    """评审文件必须含交锋段格式：advocate 结论 / skeptic 结论 / 裁决（引用双方），末行 结论：pass。"""
+    if not review_pass(path):
+        return False, f"{name} missing or not ending in 结论：pass"
+    text = path.read_text(encoding="utf-8")
+    for marker in ("advocate 结论", "skeptic 结论", "裁决"):
+        if not re.search(rf"^#{{1,4}}\s*.*{marker}|^[#]{{1,4}}\s*{marker}", text, re.MULTILINE) and marker not in text:
+            return False, f"{name} lacks 交锋段 marker: {marker}"
+    # 裁决段必须非空（引用双方结论）
+    m = re.search(r"裁决[^\n]*\n((?:[ \t]*\S[^\n]*\n)+)", text)
+    body = m.group(1).strip() if m else ""
+    if not body or len(body.splitlines()) < 2:
+        return False, f"{name} 裁决 section empty (must quote both verdicts)"
+    return True, ""
+
+
+FORBIDDEN_WORDS = ["scope", "adminonly", "治理动作", "路由", "schema", "命名空间", "endpoint", "限流配额", "gateway 管理"]
+
+
+def check_confirmation_pack(path):
+    """确认包禁用词汇表：禁用词出现行必须同段含「白话」解释。"""
+    if not path.is_file():
+        return []  # 包不存在则跳过（旧愿望或非 UI 愿望）
+    errs = []
+    for i, line in enumerate(path.read_text(encoding="utf-8").splitlines(), 1):
+        low = line.lower()
+        for w in FORBIDDEN_WORDS:
+            if w in low and "白话" not in low:
+                errs.append(f"confirmation-pack.md:{i} contains forbidden word {w!r} without 白话 explanation")
+    return errs
+
+
+def table_header(text, heading):
+    lines = text.splitlines()
+    start = None
+    for i, l in enumerate(lines):
+        if l.startswith("#") and heading in l:
+            start = i + 1
+            break
+    if start is None:
+        return []
+    while start < len(lines) and ("|" not in lines[start] or lines[start].lstrip().startswith("#")):
+        if lines[start].lstrip().startswith("#"):
+            return []
+        start += 1
+    if start >= len(lines):
+        return []
+    return cells(lines[start])
+
+
+def main():
+    if len(sys.argv) not in (3, 5):
+        print("usage: check_spec.py <workspace> <id> [--phase design|plan|build|close]", file=sys.stderr)
+        return 2
+    root = Path(sys.argv[1]).resolve()
+    wid = sys.argv[2]
+    phase = "build"
+    if len(sys.argv) == 5:
+        if sys.argv[3] != "--phase":
+            print("usage: check_spec.py <workspace> <id> [--phase design|plan|build|close]", file=sys.stderr)
+            return 2
+        phase = sys.argv[4]
+        if phase not in {"design", "plan", "build", "close"}:
+            print(f"check_spec: invalid phase {phase!r}", file=sys.stderr)
+            return 2
+
+    wish_dir = root / "docs" / "wish" / wid
+    product_md = wish_dir / "product.md"
+    arch_md = wish_dir / "arch.md"
+    review_arch = wish_dir / "review-arch.md"
+    contract_md = wish_dir / "contract.md"
+    plan_md = wish_dir / "plan.md"
+
+    errs = 0
+    need_design = phase in {"design", "build"}
+    need_plan = phase in {"plan", "build"}
+
+    # ---- design gate: quote.md + requirements.md + design-spec.md ------------
+    if need_design:
+        # ---- quote.md 用户原话工件（最高优先级，v1 保留） ---------------------
+        quote_md = wish_dir / "quote.md"
+        if not quote_md.is_file():
+            print("ERROR: missing quote.md — 用户原话逐字转录（最高优先级工件）; wish.sh --new --quote 生成，或手工补录", file=sys.stderr)
+            errs += 1
+        else:
+            qlines = [l.strip() for l in quote_md.read_text(encoding="utf-8").splitlines() if l.strip()]
+            if not qlines or "用户原话" not in qlines[0]:
+                print("ERROR: quote.md must start with heading 用户原话（最高优先级 · 逐字转录 · 不可改写）", file=sys.stderr)
+                errs += 1
+            elif any("待补录" in l for l in qlines):
+                print("ERROR: quote.md 仍是占位 — 必须逐字补录用户原话（原话最高优先级，禁止以转述代替）", file=sys.stderr)
+                errs += 1
+            # 原话指纹（v3 对抗修复）：journey --confirm 时锁定 quote.md sha256；
+            # 内容被改动（含只改一字）但未重新确认 → 拦截。存量 wish 无指纹文件 → 降级不查。
+            qhash_f = root / ".sdd" / "acceptance" / (wid + ".quote-hash")
+            if qhash_f.is_file():
+                import hashlib
+                locked = qhash_f.read_text(encoding="utf-8", errors="ignore").splitlines()[0].split("=", 1)[-1].strip()
+                cur = hashlib.sha256(quote_md.read_bytes()).hexdigest()
+                if not locked or cur != locked:
+                    print("ERROR: quote.md 内容与确认时指纹不一致（原话被改动后未重新 --confirm；quote 更新须版本+1、下游重对齐、并重新确认）", file=sys.stderr)
+                    errs += 1
+
+        # ---- requirements.md（阶段 1 产出） ----------------------------------
+        req_md = wish_dir / "requirements.md"
+        if not req_md.is_file():
+            print(f"ERROR: missing {req_md} (copy references/templates/requirements.md.tpl)", file=sys.stderr)
+            errs += 1
+        else:
+            rtext = req_md.read_text(encoding="utf-8")
+            n_ids = [r for r in table_after(rtext, "1. 需求清单") if re.match(r"^N\d+", r.get("编号", ""))]
+            if len(n_ids) < 1:
+                print("ERROR: requirements.md 需求清单 table needs ≥1 N-numbered row (N1/N2...)", file=sys.stderr)
+                errs += 1
+            if not (section_exists(rtext, "存量功能清单") or "N/A" in rtext):
+                print("ERROR: requirements.md missing 存量功能清单 section (全新建设类愿望须注明 N/A)", file=sys.stderr)
+                errs += 1
+            if not section_exists(rtext, "语义不变量"):
+                print("ERROR: requirements.md missing section 语义不变量", file=sys.stderr)
+                errs += 1
+            elif not section_lines(rtext, "语义不变量"):
+                print("ERROR: requirements.md 语义不变量 section empty (≥1 条；存量功能零丢失或显式豁免必须在此)", file=sys.stderr)
+                errs += 1
+
+        # ---- design-spec.md（阶段 2 产出） -----------------------------------
+        design_spec = wish_dir / "design-spec.md"
+        if not design_spec.is_file():
+            print(f"ERROR: missing {design_spec} (copy references/templates/design-spec.md.tpl)", file=sys.stderr)
+            errs += 1
+        else:
+            dtext = design_spec.read_text(encoding="utf-8")
+            for sec in ("页面树", "组件白名单", "共享件清单"):
+                if not section_exists(dtext, sec):
+                    print(f"ERROR: design-spec.md missing section {sec}", file=sys.stderr)
+                    errs += 1
+
+        # ---- product.md 条件式（存在才查「用户原话」节；不存在跳过不报错） ----
+        if product_md.is_file():
+            ptext = product_md.read_text(encoding="utf-8")
+            if not section_exists(ptext, "用户原话"):
+                print("ERROR: product.md missing section 用户原话 (从 quote.md 逐字复制，最高优先级)", file=sys.stderr)
+                errs += 1
+            elif not section_lines(ptext, "用户原话"):
+                print("ERROR: product.md 用户原话 section empty (必须逐字转录 quote.md)", file=sys.stderr)
+                errs += 1
+
+        # ---- confirmation-pack 禁用词汇表 + 原话回显段（v1 保留） -------------
+        pack = wish_dir / "confirmation-pack.md"
+        if pack.is_file():
+            ptext2 = pack.read_text(encoding="utf-8")
+            if not section_exists(ptext2, "你的原话"):
+                print("ERROR: confirmation-pack.md missing 你的原话回显段 (§0 原样回显 quote.md，原话最高优先级)", file=sys.stderr)
+                errs += 1
+        for e in check_confirmation_pack(pack):
+            print(f"ERROR: {e}", file=sys.stderr)
+            errs += 1
+
+        # ---- quote 版本一致性（v2 新增：stale 检出） -------------------------
+        qver = None
+        if quote_md.is_file():
+            mq = re.search(r"quote 版本\s*[：:]\s*v(\d+)", quote_md.read_text(encoding="utf-8"))
+            qver = mq.group(1) if mq else None
+        if qver:
+            for label, path, sec in (
+                ("requirements.md", req_md, "需求清单"),
+                ("design-spec.md", design_spec, "需求覆盖矩阵"),
+            ):
+                if not path.is_file():
+                    continue
+                t = path.read_text(encoding="utf-8")
+                mv = re.search(r"quote 版本\s*[：:]\s*v(\d+)", t)
+                if mv is None:
+                    print(f"ERROR: {label} missing quote 版本 line (quote= v{qver}；新 wish 必须声明并一致)", file=sys.stderr)
+                    errs += 1
+                elif mv.group(1) != qver:
+                    print(f"ERROR: {label} quote 版本 v{mv.group(1)} != quote.md v{qver} — stale，须重新对齐原话", file=sys.stderr)
+                    errs += 1
+            if pack.is_file():
+                pt = pack.read_text(encoding="utf-8")
+                mv = re.search(r"quote 版本\s*[：:]\s*v(\d+)", pt)
+                if mv is None:
+                    print(f"ERROR: confirmation-pack.md missing quote 版本 line (quote= v{qver})", file=sys.stderr)
+                    errs += 1
+                elif mv.group(1) != qver:
+                    print(f"ERROR: confirmation-pack.md quote 版本 v{mv.group(1)} != quote.md v{qver} — stale，本包作废需重做", file=sys.stderr)
+                    errs += 1
+
+        # ---- design-spec 备选方案节（v2 新增；旧 wish 无版本行时仅警告） ------
+        if design_spec.is_file():
+            dtext2 = design_spec.read_text(encoding="utf-8")
+            if section_exists(dtext2, "备选方案"):
+                opts_lines = section_lines(dtext2, "备选方案")
+                if not opts_lines:
+                    print("ERROR: design-spec.md 备选方案 section empty (≥1 个被否方案+理由)", file=sys.stderr)
+                    errs += 1
+            else:
+                if qver:
+                    print("ERROR: design-spec.md missing section 备选方案 (Alternatives considered；缺失 = 决策会被重审)", file=sys.stderr)
+                    errs += 1
+                else:
+                    print("WARN: design-spec.md missing 备选方案 section（存量 wish 不强制；下次进入本 wish 时补齐）", file=sys.stderr)
+
+    # ---- plan gate: arch.md 条件式（技术架构类愿望才需要） --------------------
+    if need_plan:
+        if arch_md.is_file():
+            text = arch_md.read_text(encoding="utf-8")
+            headers = table_header(text, "方案选项")
+            option_cols = [h for h in headers if h.startswith("方案")]
+            opts = table_after(text, "方案选项")
+            if len(option_cols) < 2:
+                print("ERROR: arch.md 方案选项 table needs ≥2 option columns (方案 A / 方案 B ...)", file=sys.stderr)
+                errs += 1
+            if not opts:
+                print("ERROR: arch.md 方案选项 table has no filled row (权衡维度)", file=sys.stderr)
+                errs += 1
+            if not section_exists(text, "方向对齐"):
+                print("ERROR: arch.md missing section 方向对齐 (逐条标 符合/偏离/缺口)", file=sys.stderr)
+                errs += 1
+            else:
+                body = "\n".join(section_lines(text, "方向对齐"))
+                if not any(k in body for k in ("符合", "偏离", "缺口")):
+                    print("ERROR: arch.md 方向对齐 section lacks 符合/偏离/缺口 annotations", file=sys.stderr)
+                    errs += 1
+
+            # ---- architecture review（仅 arch.md 存在时要求，交锋段格式） ----
+            ok, msg = review_structured(review_arch, "review-arch.md (arch-reviewer)")
+            if not ok:
+                print(f"ERROR: {msg}", file=sys.stderr)
+                errs += 1
+            elif review_arch.is_file() and "原话对照" not in review_arch.read_text(encoding="utf-8"):
+                print("ERROR: review-arch.md (arch-reviewer) lacks 原话对照 (rubric 第 0 条：对照 quote.md)", file=sys.stderr)
+                errs += 1
+        # else: 页面/前端类愿望无 arch.md → 跳过（不报错）
+
+    # ---- build gate: plan.md 必查 + contract.md 条件式 ----------------------
+    if phase == "build":
+        if not plan_md.is_file():
+            print(f"ERROR: missing {plan_md}", file=sys.stderr)
+            errs += 1
+        else:
+            plan = plan_md.read_text(encoding="utf-8")
+            srows = table_after(plan, "纵向切片")
+            if not srows:
+                print("ERROR: plan.md 纵向切片 table missing", file=sys.stderr)
+                errs += 1
+            else:
+                good = [r for r in srows
+                        if filled(r.get("切片 ID", "")) and filled(r.get("入口", ""))
+                        and filled(r.get("触及路径", ""))
+                        and any(re.search(r"T-\d+", str(v)) for v in r.values())]
+                if not good:
+                    print("ERROR: 纵向切片 has no complete row (切片ID+入口+T-xxx oracle+触及路径)", file=sys.stderr)
+                    errs += 1
+        if contract_md.is_file():
+            contract = contract_md.read_text(encoding="utf-8")
+            rows = table_after(contract, "事实映射")
+            if not rows:
+                print("ERROR: contract.md 事实映射 table missing", file=sys.stderr)
+                errs += 1
+            else:
+                filled_rows = [r for r in rows
+                               if filled(r.get("入口", "")) and filled(r.get("代码调用点", "")) and filled(r.get("证据级", ""))]
+                if not filled_rows:
+                    print("ERROR: 事实映射 has no filled row (入口+代码调用点+证据级)", file=sys.stderr)
+                    errs += 1
+                else:
+                    for r in filled_rows:
+                        ev = r.get("证据级", "")
+                        if ev not in {"Verified", "Unverified", "Assumption"}:
+                            print(f"ERROR: 事实映射 证据级 must be Verified|Unverified|Assumption, got {ev!r}", file=sys.stderr)
+                            errs += 1
+            m = re.search(r"主结果[^\n:：]*[：:]\s*(.+)", contract)
+            if not m or not filled(m.group(1)):
+                print("ERROR: contract.md 主结果 not filled", file=sys.stderr)
+                errs += 1
+
+    # ---- close gate: backfill existence checks (repo-aware) ------------------
+    if phase == "close":
+        repo_readme = root / "docs" / "README.md"
+        if repo_readme.is_file() and "回填矩阵" in repo_readme.read_text(encoding="utf-8", errors="ignore"):
+            handoff = root / "docs" / "reference" / "handoff.md"
+            if not handoff.is_file() or wid not in handoff.read_text(encoding="utf-8", errors="ignore"):
+                print("ERROR: close gate — docs/reference/handoff.md must reference wish id (会话收尾必更 handoff，禁脚本盲写)", file=sys.stderr)
+                errs += 1
+            product_text = product_md.read_text(encoding="utf-8", errors="ignore") if product_md.is_file() else ""
+            dec_lines = [l for l in section_lines(product_text, "产品决策记录") if l.strip() and "持续补录" not in l]
+            if dec_lines:  # 有实质决策 → 要求至少一处落库（gap-register 或 09-decisions）
+                gap_reg = root / "docs" / "architecture" / "11-gap-register.md"
+                dec_dir = root / "docs" / "architecture" / "09-decisions"
+                found = False
+                if gap_reg.is_file() and wid in gap_reg.read_text(encoding="utf-8", errors="ignore"):
+                    found = True
+                if not found and dec_dir.is_dir():
+                    for f in dec_dir.iterdir():
+                        if f.is_file() and wid in f.read_text(encoding="utf-8", errors="ignore"):
+                            found = True
+                            break
+                if not found:
+                    print("ERROR: close gate — product.md records real decisions but gap-register/09-decisions does not reference wish id", file=sys.stderr)
+                    errs += 1
+        else:
+            print("check_spec: close gate skipped (repo has no docs/README.md 回填矩阵)")
+
+    if errs:
+        print(f"check_spec FAIL: {wid} phase={phase} ({errs} error(s))", file=sys.stderr)
+        return 1
+    print(f"check_spec OK: {wid} phase={phase}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())