argo-search 2.9.7 → 2.9.9

package/scripts/http_client.py

@@ -36,11 +36,22 @@ from contextlib import contextmanager
 from typing import Any
 
 try:
-    from url_safety import check_url
+    from url_safety import check_url, check_redirect_target
+    from url_safety import max_redirects as _max_redirects
+    from url_safety import resolve_redirect_chain
 except ImportError:  # pragma: no cover
     def check_url(url: str) -> tuple[bool, str]:
         return True, ""
 
+    def check_redirect_target(url: str, start: float) -> dict | None:
+        return None
+
+    def _max_redirects() -> int:
+        return 5
+
+    def resolve_redirect_chain(url, headers, timeout, limit=None):
+        return None
+
 
 # ─── User-Agent 轮换池 ───────────────────────────────────────────────────────
 
@@ -525,7 +536,7 @@ class HttpClient:
         """实际执行 GET 请求（使用 http.client，不自动解压）。"""
         start = time.time()
         current_url = url
-        redirects_left = 5 if follow_redirects else 0
+        redirects_left = _max_redirects() if follow_redirects else 0
 
         while True:
             # 构建请求头
@@ -593,6 +604,11 @@ class HttpClient:
                         ch if ord(ch) < 128 else urllib.parse.quote(ch)
                         for ch in current_url
                     )
+                # 每跳重定向目标过 SSRF 校验（与 _safe_follow_redirects 对齐）：
+                # 旧实现只校验首跳，重定向到内网即可绕过（SSRF）
+                blocked = check_redirect_target(current_url, start)
+                if blocked:
+                    return blocked
                 redirects_left -= 1
                 continue
 
@@ -664,7 +680,7 @@ class HttpClient:
         start = time.time()
         headers = _random_headers(extra_headers)
 
-        final_url = self._safe_follow_redirects(url, headers)
+        final_url = resolve_redirect_chain(url, headers, self.timeout)
         if final_url is None:
             return {"status": 0, "headers": {}, "text": "", "url": url,
                     "elapsed_ms": int((time.time() - start) * 1000),
@@ -868,7 +884,7 @@ class HttpClient:
         """
         start = time.time()
         current_url = url
-        redirects_left = 5 if follow_redirects else 0
+        redirects_left = _max_redirects() if follow_redirects else 0
         while True:
             headers = _random_headers(extra_headers)
             cookie_str = self._cookies.get_cookie_header(current_url)
@@ -914,6 +930,10 @@ class HttpClient:
                 if not loc:
                     break
                 current_url = urllib.parse.urljoin(current_url, loc)
+                # 每跳重定向目标过 SSRF 校验（与 _do_get/_safe_follow_redirects 对齐）
+                blocked = check_redirect_target(current_url, start)
+                if blocked:
+                    return blocked
                 redirects_left -= 1
                 continue
             raw_body = resp.read()
@@ -978,40 +998,6 @@ class HttpClient:
             return {"status": 0, "headers": {}, "text": "", "url": url,
                     "elapsed_ms": 0, "error": str(e)[:200]}
 
-    def _safe_follow_redirects(self, url: str, headers: dict,
-                               max_redirects: int = 5) -> str | None:
-        """逐跳安全跟随重定向：每跳目标都过 SSRF 校验。
-
-        返回最终 URL；任一跳被拦截或超过跳数返回 None。
-        """
-        current = url
-        for _ in range(max_redirects):
-            ok, _reason = check_url(current)
-            if not ok:
-                return None
-            parsed = urllib.parse.urlparse(current)
-            if parsed.scheme not in ("http", "https"):
-                return None
-            try:
-                from net_proxy import open_connection, request_selector, resolve_proxy
-                proxy_url = resolve_proxy(current)
-                conn, via_proxy = open_connection(parsed, self.timeout, proxy_url)
-                path = parsed.path or "/"
-                if parsed.query:
-                    path += "?" + parsed.query
-                conn.request("HEAD", request_selector(parsed, path, via_proxy),
-                             headers=headers)
-                resp = conn.getresponse()
-                status = resp.status
-                location = resp.getheader("Location")
-                conn.close()
-            except Exception:
-                return None
-            if status not in (301, 302, 303, 307, 308) or not location:
-                return current
-            current = urllib.parse.urljoin(current, location)
-        return None
-
 
 # ─── 便捷函数 ────────────────────────────────────────────────────────────────
 

package/scripts/install.ps1

@@ -77,6 +77,15 @@ if (Test-Path (Join-Path $InstallDir ".git")) {
     Write-Host "==> 克隆仓库…"
     New-Item -ItemType Directory -Force -Path (Split-Path $InstallDir) | Out-Null
     git clone --depth 1 --branch $Branch $Repo $InstallDir
+    if ($LASTEXITCODE -ne 0 -and $Repo -like "*github.com*") {
+        $MirrorRepo = "https://ghfast.top/" + $Repo
+        Write-Host "[warn] 直连 GitHub 失败，尝试加速镜像: $MirrorRepo …"
+        git clone --depth 1 --branch $Branch $MirrorRepo $InstallDir
+        if ($LASTEXITCODE -ne 0) {
+            Write-Error "克隆失败。若处于受限网络环境，可尝试手动克隆或设置 ARGO_REPO 环境变量。"
+            exit 1
+        }
+    }
 }
 
 # 供应链加固：固定 commit 校验（ARGO_PIN），与 install.sh 对齐
@@ -94,11 +103,20 @@ if ($Pin) {
 if (-not $SkipPip) {
     Write-Host "==> 安装依赖 (PyYAML)…"
     & $script:py -m pip install pyyaml
+    if ($LASTEXITCODE -ne 0) {
+        & $script:py -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple pyyaml
+    }
     if ($LASTEXITCODE -ne 0) { Write-Host "[warn] pip 安装 PyYAML 失败，可手动: pip install pyyaml" }
     Write-Host "==> 安装可选增强 (curl_cffi: TLS 指纹伪造；rapidfuzz: 去重模糊近重复复核；缺失不影响核心功能)…"
     & $script:py -m pip install curl_cffi
+    if ($LASTEXITCODE -ne 0) {
+        & $script:py -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple curl_cffi
+    }
     if ($LASTEXITCODE -ne 0) { Write-Host "[warn] pip 安装 curl_cffi 失败（可选依赖）" }
     & $script:py -m pip install rapidfuzz
+    if ($LASTEXITCODE -ne 0) {
+        & $script:py -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple rapidfuzz
+    }
     if ($LASTEXITCODE -ne 0) { Write-Host "[warn] pip 安装 rapidfuzz 失败（可选依赖，缺失时去重回退纯 minhash）" }
 }
 

package/scripts/install.sh

@@ -104,7 +104,19 @@ elif [[ -d "$INSTALL_DIR" ]] && [[ -f "$INSTALL_DIR/scripts/search.py" ]]; then
 else
   echo "==> 克隆仓库…"
   mkdir -p "$(dirname "$INSTALL_DIR")"
-  git clone --depth 1 --branch "$BRANCH" "$REPO" "$INSTALL_DIR"
+  if ! git clone --depth 1 --branch "$BRANCH" "$REPO" "$INSTALL_DIR"; then
+    if [[ "$REPO" == *"github.com"* ]]; then
+      MIRROR_REPO="https://ghfast.top/${REPO}"
+      echo "[warn] 从 GitHub 直连克隆失败，尝试加速镜像: $MIRROR_REPO …"
+      if ! git clone --depth 1 --branch "$BRANCH" "$MIRROR_REPO" "$INSTALL_DIR"; then
+        echo "!! 克隆失败。若处于受限网络环境，可尝试手动克隆或设置 ARGO_REPO 环境变量。" >&2
+        exit 1
+      fi
+    else
+      echo "!! 克隆失败，请检查网络连接。" >&2
+      exit 1
+    fi
+  fi
 fi
 
 # 供应链加固：校验 HEAD 与预期 commit 一致（ARGO_PIN 固定）
@@ -122,13 +134,16 @@ if [[ "$SKIP_PIP" != "1" ]]; then
   echo "==> 安装依赖 (PyYAML)…"
   python3 -m pip install --user -q pyyaml 2>/dev/null \
     || python3 -m pip install -q pyyaml 2>/dev/null \
+    || python3 -m pip install --user -q -i https://pypi.tuna.tsinghua.edu.cn/simple pyyaml 2>/dev/null \
     || echo "[warn] pip 安装 PyYAML 失败，可手动: pip install pyyaml"
   echo "==> 安装可选增强 (curl_cffi: TLS 指纹伪造；rapidfuzz: 去重模糊近重复复核；缺失不影响核心功能)…"
   python3 -m pip install --user -q curl_cffi 2>/dev/null \
     || python3 -m pip install -q curl_cffi 2>/dev/null \
+    || python3 -m pip install --user -q -i https://pypi.tuna.tsinghua.edu.cn/simple curl_cffi 2>/dev/null \
     || echo "[warn] pip 安装 curl_cffi 失败（可选依赖）。需要反爬 TLS 指纹伪造时手动: pip install curl_cffi"
   python3 -m pip install --user -q rapidfuzz 2>/dev/null \
     || python3 -m pip install -q rapidfuzz 2>/dev/null \
+    || python3 -m pip install --user -q -i https://pypi.tuna.tsinghua.edu.cn/simple rapidfuzz 2>/dev/null \
     || echo "[warn] pip 安装 rapidfuzz 失败（可选依赖）。缺失时去重回退纯 minhash，需要模糊近重复复核时手动: pip install rapidfuzz"
 fi
 

package/scripts/job.py

@@ -895,13 +895,13 @@ def _is_quota_exhausted(exc: Exception) -> bool:
     的模块静默生效。403/429 各有归属：engine_failure.classify 分 auth /
     blocked / rate_limited，健康问题交熔断器自愈，本函数不接手。
 
-    关键词表复用 engine_dispatch 的 `_QUOTA_ERROR_KEYWORDS`（该表自述为
-    「唯一来源」，配额分类与自适应跳过共用），不另起一份。
+    关键词表真源是 cache_guard 的 `_QUOTA_ERROR_KEYWORDS`（配额分类与自适应
+    跳过共用），不另起一份。
     真欠费的文案（`You do not have enough money or package quota`）含
     `quota`，仍然照常命中。
     """
     try:
-        from engine_dispatch import _QUOTA_ERROR_KEYWORDS
+        from cache_guard import _QUOTA_ERROR_KEYWORDS
     except ImportError:
         return False
     text = str(exc).lower()

package/scripts/local_seek.py

@@ -151,6 +151,37 @@ def _seek_disk_cache_put(key: str, results: list[dict[str, Any]]) -> None:
         pass
 
 
+# 本地代码/工程/配置意图正则模式
+_LOCAL_INTENT_PATTERN = re.compile(
+    r"(?i)(?:\b(?:def|class|function|fn|import|export|const|var|struct|enum)\b|"
+    r"\.(?:py|js|ts|json|yaml|yml|md|rs|go|c|cpp|h|sh|sql)\b|"
+    r"(?:readme|config|package\.json|cargo\.toml|dockerfile|makefile)|"
+    r"(?:本工程|本项目|代码库|本地文件|架构实现|函数|方法|源码|接口定义))"
+)
+
+
+def has_local_code_intent(query: str) -> bool:
+    """是否包含本地工程/代码/配置检索意图。"""
+    if not query:
+        return False
+    return bool(_LOCAL_INTENT_PATTERN.search(query))
+
+
+def resolve_include_local(include_local: bool | None, mode: str, query: str = "") -> bool:
+    """include-local 三态解析：True/False 显式，None=自动。
+
+    自适应感知：除 fast/budget 模式外，当查询具有明显的本地代码/文件特征
+    且处于非宽泛根项目目录时，自动开启本地命中。
+    """
+    if include_local is not None:
+        return bool(include_local)
+    if mode in ("fast", "budget"):
+        return True
+    if has_local_code_intent(query) and _local_seek_dir() is not None:
+        return True
+    return False
+
+
 def _is_broad_local_root(path: str) -> bool:
     """path 是否是「不能当作本地检索范围」的宽泛根。
 
@@ -200,7 +231,8 @@ def _local_seek_dir() -> str | None:
 
 
 def _run_local_seek(query: str, max_n: int = 5,
-                    search_dir: str | None = None) -> list[dict[str, Any]]:
+                    search_dir: str | None = None,
+                    warn: bool = True) -> list[dict[str, Any]]:
     """本机文件命中（include-local 用）：进程内调 seek.run_query，JSON 并入。
 
     fast/budget 模式下由 super_search 自动调用（_resolve_include_local），
@@ -215,10 +247,14 @@ def _run_local_seek(query: str, max_n: int = 5,
         # 静默跳过会让用户以为本地没匹配，而事实是根本没查。返回一个
         # 占位「结果」又会污染 results（它只是个说明，不是命中），所以
         # 说明走 stderr：与 CLI 既有的 include-local 异常提示同一条通道。
-        sys.stderr.write(
-            "  [include-local] 当前目录不适合本地检索（home / 根目录过宽，或位于 argo "
-            "自身安装树内——只会命中自身源码），已跳过；"
-            "要搜本机文件请 cd 到目标目录后重试，或用 local-seek 的 seek.py --path <目录>\n")
+        # 只在**显式开启**时提示（warn=False 是 fast/budget 的自动路径）：
+        # 自动路径在 home 等宽根下每次 search 都会打印，Agent 会把它当成
+        # 输出的一部分，属于噪音而非归因。
+        if warn:
+            sys.stderr.write(
+                "  [include-local] 当前目录不适合本地检索（home / 根目录过宽，或位于 argo "
+                "自身安装树内——只会命中自身源码），已跳过；"
+                "要搜本机文件请 cd 到目标目录后重试，或用 local-seek 的 seek.py --path <目录>\n")
         return []
 
     # 缓存：同 query+target 在 TTL 内直接命中，避免重复冷启动 seek.py 子进程。

package/scripts/macro_countries.py

@@ -66,6 +66,12 @@ _NON_US_COUNTRIES: dict[str, str] = {
     name: code for name, code in WORLDBANK_COUNTRIES.items() if code != "USA"
 }
 
+# 美国子集：给「统计局/国内」这类中国语境别名做前置排除。
+# 同一张主表派生，避免「美国 统计局」被当成中国口径。
+_US_COUNTRIES: dict[str, str] = {
+    name: code for name, code in WORLDBANK_COUNTRIES.items() if code == "USA"
+}
+
 
 def is_foreign_macro_query(query: str) -> bool:
     """查询是否明确指向非美国国家/地区。
@@ -78,5 +84,14 @@ def is_foreign_macro_query(query: str) -> bool:
     英文词——实测 `focus on growth`、`consensus 预测`、`business survey`
     全被判成「外国宏观查询」，FRED 因而被无谓地排除。同一个概念（国家词识别）
     在同一模块里有两套行为不同的实现，正是这类漏判的来源。
+
+    补充（2026-10-23）：「国家统计局」「国内」「我国」是明确的中国语境，
+    但不在国家名表里。实测「国家统计局 一季度 GDP」因此没让位，FRED 领队
+    返回美国 GDP。这几个词只指向中国，直接判为 foreign（非美国）。
+    但「美国 统计局」「美国国内 CPI」同时含美国词，必须先排除美国口径。
     """
+    q = query or ""
+    if (not match_country(q, _US_COUNTRIES)
+            and any(term in q for term in ("国家统计局", "统计局", "国内", "我国"))):
+        return True
     return bool(match_country(query, _NON_US_COUNTRIES))

package/scripts/matrix_search_eval.py

@@ -254,6 +254,65 @@ ROUTE_MATRIX: list[dict[str, Any]] = [
      "domain": "org_entity", "primary": "wikidata", "scenario": "org"},
     {"id": "R_zh_org2", "q": "清华大学 创办年份", "lang": "zh",
      "domain": "org_entity", "primary": "wikidata", "scenario": "org"},
+    # 中文企业总部不是地理问题（2026-10-22 修：geo_places 的方位词 unless 增补
+    # 「总部|母公司|子公司|注册地|办公地」后落到本域；此前落 geo_places，用 OSM+维基
+    # 只拿到广汽/新凯来/五粮液等无关结果）
+    {"id": "R_zh_org_hq", "q": "华为 总部在哪", "lang": "zh",
+     "domain": "org_entity", "primary": "wikidata", "scenario": "org"},
+
+    # ── 企业/工商（company_search）──
+    # 2026-10-22 补：此前 ROUTE_MATRIX 无本域用例，导致「中国企业股权/工商信息」
+    # 被路由到本域、而该域 combo 只有国际法人库（gleif/opencorporates）+ 百科的
+    # 问题长期无人发现——实测「长亭科技 股东 持股比例」返回 5 条同名「XX科技
+    # 有限公司」干扰。元宝进 combo 第 1 位后（a5a46ad）才具中文交付能力。
+    {"id": "R_zh_company_shareholder", "q": "长亭科技 股东 持股比例", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
+    {"id": "R_zh_company_regcap", "q": "字节跳动 注册资本 变更", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
+    # 2026-10-25 补（企业主体查询 MECE 类目）：尽调风险（注销/经营异常）、开票
+    # （开票信息/税号）、人查公司（担任法定代表人/名下公司）三类此前落
+    # org_entity 的 wikidata 组合——wikidata 对中文企业 7 天空结果率 82%，
+    # company_search 的元宝才是交付源。config.yaml 两域词表成对改动
+    # （org_entity 让位 unless + company_search 收口 match/unless），改一处
+    # 必须同步另一处。企业名/人名一律合成名（对外脱敏纪律，v2.9.8 同款）。
+    {"id": "R_zh_company_risk", "q": "晨曦合成科技有限公司 注销了吗", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
+    {"id": "R_zh_company_invoice", "q": "晨曦合成科技有限公司 开票信息 税号", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
+    {"id": "R_zh_company_person", "q": "李合成 担任法定代表人的公司", "lang": "zh",
+     "domain": "company_search", "primary": "opencorporates", "scenario": "company"},
+    # 双向锁（v2.9.8 纪律）：新词表不得误吸非主体查询——办事指南是 howto、
+    # 招聘是求职，都不该进尽调域
+    {"id": "R_zh_company_not_howto", "q": "公司注销流程 怎么办", "lang": "zh",
+     "domain_not": ["company_search", "org_entity"], "scenario": "company"},
+    {"id": "R_zh_company_not_job", "q": "晨曦合成科技有限公司 招聘", "lang": "zh",
+     "domain_not": ["company_search"], "scenario": "company"},
+    # ── 2026-10-25 六族 MECE 扫描（学术/金融财务/法律/医药）──
+    # 全部为真实场景类目，企业名/人名保持合成名（脱敏纪律）。每处 config 改动
+    # 都有成对的 unless 让位，反例见 tests/test_domain_rule_schema.py。
+    {"id": "R_zh_stock_financials", "q": "贵州茅台 2025 年报 营收", "lang": "zh",
+     "domain": "stock_query", "primary": "sina_quote", "scenario": "finance"},
+    {"id": "R_zh_stock_valuation", "q": "贵州茅台 市盈率 估值", "lang": "zh",
+     "domain": "stock_query", "primary": "sina_quote", "scenario": "finance"},
+    # zh 财务词形不得落美股域（sec_companyfacts 只认 CIK，A 股全废）
+    {"id": "R_zh_usfund_yields_cn", "q": "宁德时代 毛利率 现金流", "lang": "zh",
+     "domain_not": ["us_fundamentals"], "scenario": "finance"},
+    # 民事实务（劳动/借贷）此前落 chinese_general，flk_law 才是交付源
+    {"id": "R_zh_legal_labor", "q": "劳动合同解除 赔偿 N+1", "lang": "zh",
+     "domain": "legal", "primary": "anysearch", "scenario": "legal"},
+    {"id": "R_zh_legal_lending", "q": "民间借贷 利率 司法保护上限", "lang": "zh",
+     "domain": "legal", "primary": "anysearch", "scenario": "legal"},
+    # 药品说明书此前被 code_search 的「用法」子串抢走（GitHub/MDN 答不了药）
+    {"id": "R_zh_medical_drug_label", "q": "布洛芬 说明书 用法用量", "lang": "zh",
+     "domain": "medical", "primary": "clinicaltrials", "scenario": "medical"},
+    # 综述语境不被模型目录抢（models_dev 答不了综述）
+    {"id": "R_zh_survey_not_modelcard", "q": "大模型 长上下文 综述", "lang": "zh",
+     "domain_not": ["ai_model"], "scenario": "academic"},
+    # 论文写作 howto / 年报模板 howto 不进学术域与行情域
+    {"id": "R_zh_academic_not_howto", "q": "怎么写论文 查重 降重", "lang": "zh",
+     "domain_not": ["academic"], "scenario": "academic"},
+    {"id": "R_zh_stock_not_howto", "q": "年报 模板 怎么写", "lang": "zh",
+     "domain_not": ["stock_query", "us_fundamentals"], "scenario": "finance"},
 
     # ── 媒体/音乐 ──
     {"id": "R_en_media", "q": "Taylor Swift album", "lang": "en",
@@ -384,6 +443,81 @@ ROUTE_MATRIX: list[dict[str, Any]] = [
                 "worldbank", "nbs_stats", "un_comtrade", "endoflife",
                 "deps_dev", "openalex", "europepmc"],
      "forbid_cn": True, "scenario": "negative"},
+
+    # ── 域覆盖（2026-10-22）：补 53 个此前零用例的域 ────────────────────────
+    # 零用例 = 路由回归无人挡。company_search 零用例，「中国企业股权/工商信息」
+    # 被路由到本域却无中文能力的问题因此潜伏 8 天；geo_places 的「在哪」劫持
+    # 企业总部查询同理（两者都是靠人工撞出来的）。lang 取路由实测 primary_lang，
+    # 与 _route_ok 的 lang 断言对齐；primary 取域声明的主引擎。
+    {"id": "R_ai_model", "q": "models.dev claude 价格 上下文", "lang": "en", "domain": "ai_model", "primary": "models_dev", "scenario": "tech"},
+    {"id": "R_aviation_weather", "q": "北京首都机场 航空气象 METAR", "lang": "zh", "domain": "aviation_weather", "primary": "aviation_weather", "scenario": "weather"},
+    {"id": "R_chinese_tech_deep", "q": "中文 编程 算法 架构 深度对比", "lang": "zh", "domain": "chinese_tech_deep", "primary": "byted", "scenario": "tech"},
+    {"id": "R_claim_check", "q": "辟谣 传言 事实核查", "lang": "zh", "domain": "claim_check", "primary": "factcheck_org", "scenario": "factcheck"},
+    {"id": "R_cn_encyclopedia", "q": "长征 百度百科 词条解释", "lang": "zh", "domain": "cn_encyclopedia", "primary": "baidu_baike", "scenario": "media"},
+    {"id": "R_cn_tech_community", "q": "掘金 前端文章 技术博客", "lang": "zh", "domain": "cn_tech_community", "primary": "juejin", "scenario": "tech"},
+    {"id": "R_code_search", "q": "github 代码 实现 用法", "lang": "zh", "domain": "code_search", "primary": "github", "scenario": "code"},
+    {"id": "R_crypto_search", "q": "比特币 以太坊 加密货币 价格", "lang": "zh", "domain": "crypto_search", "primary": "coingecko", "scenario": "finance"},
+    {"id": "R_dataset_search", "q": "科研数据集 开放数据 下载", "lang": "zh", "domain": "dataset_search", "primary": "datacite", "scenario": "academic"},
+    {"id": "R_dictionary_search", "q": "英文单词 词典 意思", "lang": "zh", "domain": "dictionary_search", "primary": "free_dictionary", "scenario": "media"},
+    {"id": "R_earth_science", "q": "地震 震级 余震 震中", "lang": "zh", "domain": "earth_science", "primary": "usgs", "scenario": "science"},
+    {"id": "R_em_news_search", "q": "东财 全球资讯 市场", "lang": "zh", "domain": "em_news_search", "primary": "em_global_news", "scenario": "news"},
+    {"id": "R_energy_grid", "q": "电网 发电结构 发电量", "lang": "zh", "domain": "energy_grid", "primary": "energy_charts", "scenario": "science"},
+    {"id": "R_entity_search", "q": "维基数据 知识图谱 实体查询", "lang": "zh", "domain": "entity_search", "primary": "wikidata", "scenario": "media"},
+    {"id": "R_fact_check", "q": "长江 长度 是多少", "lang": "zh", "domain": "fact_check", "primary": "wikipedia", "scenario": "factcheck"},
+    {"id": "R_futures_market", "q": "沪铜 期货 行情", "lang": "zh", "domain": "futures_market", "primary": "shfe", "scenario": "finance"},
+    {"id": "R_fx_query", "q": "人民币 汇率 中间价", "lang": "zh", "domain": "fx_query", "primary": "fx_rate", "scenario": "finance"},
+    {"id": "R_game_search", "q": "steam 游戏推荐 好玩的游戏", "lang": "zh", "domain": "game_search", "primary": "steam", "scenario": "media"},
+    {"id": "R_global_event", "q": "国际事件 地缘政治 局势", "lang": "zh", "domain": "global_event", "primary": "google_news", "scenario": "news"},
+    {"id": "R_hk_filings", "q": "港股 港交所 披露易 公告", "lang": "zh", "domain": "hk_filings", "primary": "hkex", "scenario": "finance"},
+    {"id": "R_hot_trending", "q": "微博热搜 今日热搜", "lang": "zh", "domain": "hot_trending", "primary": "baidu_hot", "scenario": "community"},
+    {"id": "R_image_search", "q": "雪景 壁纸 图片", "lang": "zh", "domain": "image_search", "primary": "openverse", "scenario": "media"},
+    {"id": "R_intl_news_flash", "q": "国际新闻 外媒 国际要闻", "lang": "zh", "domain": "intl_news_flash", "primary": "guardian_rss", "scenario": "news"},
+    {"id": "R_japan_law", "q": "日本 法令 民法 条文", "lang": "zh", "domain": "japan_law", "primary": "egov_law", "scenario": "legal"},
+    {"id": "R_kor_law", "q": "韩国 法令 民法 条文", "lang": "zh", "domain": "kor_law", "primary": "kor_law", "scenario": "legal"},
+    {"id": "R_law_text", "q": "法条 法律条文 全文", "lang": "zh", "domain": "law_text", "primary": "flk_law", "scenario": "legal"},
+    {"id": "R_legal", "q": "刑法 司法解释 判例", "lang": "zh", "domain": "legal", "primary": "anysearch", "scenario": "legal"},
+    {"id": "R_meme_slang", "q": "网络梗 流行语 黑话", "lang": "zh", "domain": "meme_slang", "primary": "itotii", "scenario": "media"},
+    {"id": "R_ml_models", "q": "huggingface 预训练模型 开源模型", "lang": "zh", "domain": "ml_models", "primary": "huggingface", "scenario": "tech"},
+    {"id": "R_news_realtime", "q": "最新新闻 今天发生 突发", "lang": "zh", "domain": "news_realtime", "primary": "byted", "scenario": "news"},
+    {"id": "R_outbreak_health", "q": "疫情 暴发 流行病 通报", "lang": "zh", "domain": "outbreak_health", "primary": "who_don", "scenario": "science"},
+    {"id": "R_patent_search", "q": "专利 权利要求 技术方案", "lang": "zh", "domain": "patent_search", "primary": "google_patents", "scenario": "academic"},
+    {"id": "R_prediction_market", "q": "预测市场 事件合约", "lang": "zh", "domain": "prediction_market", "primary": "polymarket", "scenario": "finance"},
+    {"id": "R_protein_search", "q": "蛋白质 基因 氨基酸序列", "lang": "zh", "domain": "protein_search", "primary": "uniprot", "scenario": "academic"},
+    {"id": "R_redskill_search", "q": "小红书 技能 排行榜", "lang": "zh", "domain": "redskill_search", "primary": "redskill", "scenario": "community"},
+    {"id": "R_scholar_search", "q": "google scholar 学术论文 期刊", "lang": "zh", "domain": "scholar_search", "primary": "google_scholar", "scenario": "academic"},
+    {"id": "R_sec_filings", "q": "SEC EDGAR 招股书 财报申报", "lang": "zh", "domain": "sec_filings", "primary": "sec_edgar", "scenario": "finance"},
+    {"id": "R_semantic_discovery", "q": "discover landscape ecosystem emerging", "lang": "en", "domain": "semantic_discovery", "primary": "exa", "scenario": "tech"},
+    {"id": "R_shopping", "q": "购物 评测 推荐 踩坑", "lang": "zh", "domain": "shopping", "primary": "zhihu", "scenario": "media"},
+    {"id": "R_skill_search", "q": "agent skills 市场 目录", "lang": "en", "domain": "skill_search", "primary": "skillsmp", "scenario": "community"},
+    {"id": "R_species_search", "q": "物种 学名 生物分类 拉丁名", "lang": "zh", "domain": "species_search", "primary": "gbif", "scenario": "academic"},
+    {"id": "R_stackoverflow_search", "q": "stackoverflow 编程问题 代码问题", "lang": "zh", "domain": "stackoverflow_search", "primary": "stackoverflow", "scenario": "code"},
+    {"id": "R_standards", "q": "国家标准 国标 标准号", "lang": "zh", "domain": "standards", "primary": "std_samr", "scenario": "academic"},
+    {"id": "R_tech_deep", "q": "deep dive 架构 实践", "lang": "zh", "domain": "tech_deep", "primary": "arxiv", "scenario": "tech"},
+    # 2026-10-25 约定更新：zh 词形财务查询统一走 stock_query（sina/eastmoney
+    # 双地覆盖，与「XX 股价」分流同构）；本域收英文词形与美股语境。代价是
+    # zh 美股财务查询损失 sec 结构化精度，换来 A 股企业不再落只认 CIK 的
+    # sec_companyfacts 全空——两害取其轻，成对改动注释见 config us_fundamentals。
+    # 英文方向的域锁定另见 R_us_fundamentals_en。
+    {"id": "R_us_fundamentals", "q": "营收 净利润 财务数据 苹果", "lang": "zh",
+     "domain": "stock_query", "primary": "sina_quote", "scenario": "finance"},
+    {"id": "R_us_fundamentals_en", "q": "apple annual revenue net income", "lang": "en",
+     "domain": "us_fundamentals", "primary": "sec_companyfacts", "scenario": "finance"},
+    {"id": "R_us_legal", "q": "supreme court 判例 precedent", "lang": "en", "domain": "us_legal", "primary": "federal_register", "scenario": "legal"},
+    {"id": "R_web_archive", "q": "网页快照 历史网页 wayback", "lang": "zh", "domain": "web_archive", "primary": "wikipedia", "scenario": "general"},
+    {"id": "R_web_docs", "q": "mozilla developer fetch 属性", "lang": "en", "domain": "web_docs", "primary": "mdn", "scenario": "code"},
+    {"id": "R_wechat_search", "q": "微信公众号 文章 搜狗微信", "lang": "zh", "domain": "wechat_search", "primary": "wechat_sogou", "scenario": "community"},
+    {"id": "R_world_news", "q": "noticias 国际 新闻", "lang": "zh", "domain": "world_news", "primary": "anysearch", "scenario": "news"},
+    {"id": "R_zhihu_content", "q": "知乎 怎么看待 如何评价", "lang": "zh", "domain": "zhihu_content", "primary": "zhihu", "scenario": "media"},
+    {"id": "R_zhihu_hot_list", "q": "知乎热榜 知乎热搜", "lang": "zh", "domain": "zhihu_hot_list", "primary": "zhihu_hot", "scenario": "community"},
+    {"id": "R_zhihu_user_data", "q": "我的知乎 我的创作 收藏", "lang": "zh", "domain": "zhihu_user_data", "primary": "zhihu_user", "scenario": "community"},
+
+    # ── 域冲突让位修复（2026-10-22）────────────────────────────────────────
+    # 这三个域的触发词此前被更宽泛的域覆盖，无法通过自动路由命中（因此曾登记在
+    # test_route_matrix_coverage 的 _NO_CASE_DEBT 里）。在宽泛侧加 unless 让位后
+    # 各补一条钉住；宽泛侧的真查询锁在 test_domain_pattern_precision 的既有用例。
+    {"id": "R_jin10_flash", "q": "金十 7x24 快讯", "lang": "zh", "domain": "jin10_flash", "primary": "jin10", "scenario": "news"},
+    {"id": "R_v2ex_search", "q": "v2ex 技术社区", "lang": "zh", "domain": "v2ex_search", "primary": "v2ex", "scenario": "code"},
+    {"id": "R_wenshu_query", "q": "裁判文书 判决书", "lang": "zh", "domain": "wenshu_query", "primary": "wenshu", "scenario": "legal"},
 ]
 
 
@@ -1049,7 +1183,65 @@ def _summarize(c: Checker) -> dict[str, Any]:
     }
 
 
+def _reexec_with_yaml() -> bool:
+    """缺 PyYAML 时改用带依赖的解释器重入；exec 成功则不返回。
+
+    why：config 在无 yaml 时静默回落内置默认引擎集，route 段整片假红
+    （2026-10-25 实测：同机无 yaml 的解释器跑出 94 FAIL、带 yaml 的全绿）。
+    bin/argo 启动器的 _pick_python 会自动挑解释器，但分享场景里用户会直接
+    `python3 scripts/matrix_search_eval.py`——解释器版本不该卡住人。候选序：
+    ARGO_PYTHON（与启动器同一权威覆盖）→ PATH 上的 python3.14/3.13/3.15/
+    3.12 → python3；每个候选真跑一次 `import yaml` 验证（探测约 100ms，
+    评测不是热路径）。已重入过仍缺（_ARGO_MATRIX_REEXEC 标记）则放弃，
+    避免套娃。
+    """
+    import os
+    import shutil
+    import subprocess
+
+    if os.environ.get("_ARGO_MATRIX_REEXEC") == "1":
+        return False
+    cands = []
+    env_py = os.environ.get("ARGO_PYTHON", "").strip()
+    if env_py:
+        cands.append(env_py)
+    cands += [shutil.which("python3.%d" % m) or "" for m in (14, 13, 15, 12)]
+    cands.append(shutil.which("python3") or "")
+    for cand in cands:
+        if not cand:
+            continue
+        if os.path.realpath(cand) == os.path.realpath(sys.executable):
+            continue  # 当前解释器已证明缺 yaml，跳过
+        try:
+            probe = subprocess.run([cand, "-c", "import yaml"],
+                                   capture_output=True, timeout=30)
+        except Exception:
+            continue
+        if probe.returncode == 0:
+            os.environ["_ARGO_MATRIX_REEXEC"] = "1"
+            print("[matrix] 当前解释器缺 PyYAML，改用 %s 重入" % cand,
+                  file=sys.stderr)
+            os.execv(cand, [cand, sys.argv[0], *sys.argv[1:]])
+    return False
+
+
 def main() -> int:
+    # PyYAML 缺失时 config 会**静默**回落到内置默认引擎集（垂直域引擎全不在），
+    # route 段整片假红（2026-10-25 实测：无 yaml 的解释器跑出 94 FAIL，带 yaml
+    # 重跑 197 全绿——错误根因是解释器不是路由）。先自愈换解释器重入
+    # （3.15/3.14 等版本差异不再卡用户），换无可换才硬失败退出。
+    try:
+        import yaml  # noqa: F401
+    except ImportError as exc:
+        _reexec_with_yaml()
+        print(
+            f"[matrix] PyYAML 不可用（{exc}）且候选解释器（ARGO_PYTHON 与 PATH "
+            "上的 python3.x）均缺：config 将静默降级到内置默认引擎集，route 段"
+            "会整片假红。请 pip install pyyaml，或设 ARGO_PYTHON 指向带依赖的"
+            "解释器。",
+            file=sys.stderr,
+        )
+        return 2
     ap = argparse.ArgumentParser(description="多语言×引擎×场景 矩阵测试")
     ap.add_argument("--offline", action="store_true")
     ap.add_argument("--live", action="store_true")