argo-search 1.0.1 → 2.9.7

package/scripts/engines.py

@@ -1,24 +1,23 @@
 #!/usr/bin/env python3
-"""engines.py — Unified Search v2 引擎适配层（精简版，< 400 行）
+"""engines.py — Unified Search v2 引擎适配层（门面）
 
-配置驱动 + 声明式 output_map 字段提取 + 通用 parser 兜底。
-支持 cli / http(GET/POST) 类型，所有异常吞没返回 []。
+配置驱动 + 声明式 output_map 字段提取 + 通用 parser 保底。
+实现拆分：
+  - engines_base.py      公共工具 / cli / http / html / 通用解析
+  - engines_builders.py  专用引擎构建器
+本文件仅负责 BUILDERS 注册表与 search() 入口。
 """
 
 from __future__ import annotations
 
-import functools
+import contextlib
+import inspect
 import json
-import logging
-import os
 import re
-import subprocess
 import sys
 import time
-import urllib.error
-import urllib.request
 from pathlib import Path
-from typing import Any, Callable
+from typing import Any
 
 try:
     from config import load_config, get_engines
@@ -26,344 +25,829 @@ except ImportError:
     sys.path.insert(0, str(Path(__file__).parent))
     from config import load_config, get_engines
 
-logger = logging.getLogger("unified_search.engines")
-if not logger.handlers:
-    logger.setLevel(logging.WARNING)
-    logger.addHandler(logging.StreamHandler(sys.stderr))
-
-
-def safe_search(fn: Callable) -> Callable:
-    """统一错误处理装饰器 — 所有异常返回 []。"""
-    @functools.wraps(fn)
-    def wrapper(*args, **kwargs) -> list[dict[str, Any]]:
-        name = fn.__name__.replace("_engine", "").strip("_")
-        try:
-            return fn(*args, **kwargs)
-        except subprocess.TimeoutExpired:
-            logger.warning(f"引擎 {name} 超时")
-        except FileNotFoundError as e:
-            logger.warning(f"引擎 {name} 命令不存在: {e}")
-        except Exception as e:
-            logger.error(f"引擎 {name} 异常: {type(e).__name__}: {e}", exc_info=True)
+try:
+    from single_flight import engine_coalescer
+except ImportError:  # pragma: no cover
+    def engine_coalescer():  # type: ignore
+        class _Noop:
+            def run(self, key, fn, **kw):
+                return fn()
+        return _Noop()
+
+from engines_base import (
+    safe_search,
+    _build_cli_engine,
+    _build_http_engine,
+    _build_html_engine,
+    _parse_generic,
+    _parse_text_output,
+    _parse_xml,
+    _parse_uapi,
+    _parse_semantic_scholar,
+    _ensure_engine_source,
+    _CUSTOM_JSON_PARSERS,
+)
+from recovery import strip_structured
+from cli_io import dumps
+
+# 对外/测试兼容：专用解析器与 source 纠正
+__all__ = [
+    "search",
+    "available_engines",
+    "get_registry",
+    "safe_search",
+    "_parse_uapi",
+    "_parse_semantic_scholar",
+    "_ensure_engine_source",
+]
+from engines_builders import (
+    _build_who_don_engine,
+    _build_who_gho_engine,
+    _build_pubmed_engine,
+    _build_core_engine,
+    _build_gdacs_engine,
+    _build_obis_engine,
+    _build_worms_engine,
+    _build_energy_charts_engine,
+    _build_jolpica_engine,
+    _build_openf1_engine,
+    _build_openligadb_engine,
+    _build_artic_engine,
+    _build_cleveland_engine,
+    _build_tvmaze_engine,
+    _build_jikan_engine,
+    _build_deezer_engine,
+    _build_listenbrainz_engine,
+    _build_egov_law_engine,
+    _build_k10plus_engine,
+    _build_ror_engine,
+    _build_soilgrids_engine,
+    _build_noaa_swpc_engine,
+    _build_satnogs_engine,
+    _build_tle_mirror_engine,
+    _build_nhtsa_vpic_engine,
+    _build_exa_engine,
+    _build_anysearch_engine,
+    _build_parallel_engine,
+    _build_parallel_free_engine,
+    _build_seltz_engine,
+    _build_you_engine,
+    _build_em_miaoxiang_engine,
+    _build_cninfo_engine,
+    _build_sina_quote_engine,
+    _build_tencent_quote_engine,
+    _build_em_flow_engine,
+    _build_wechat_sogou_engine,
+    _build_hackernews_engine,
+    _build_stackoverflow_engine,
+    _build_google_scholar_engine,
+    _build_v2ex_engine,
+    _build_tineye_engine,
+    _build_bing_rss_engine,
+    _build_ths_hot_engine,
+    _build_cls_telegraph_engine,
+    _build_em_global_news_engine,
+    _build_eastmoney_engine,
+    _build_itotii_engine,
+    _build_baidu_hot_engine,
+    _build_toutiao_hot_engine,
+    _build_bilibili_hot_engine,
+    _build_zhihu_global_engine,
+    _build_zhihu_user_engine,
+    _build_bocha_engine,
+    _build_bocha_ai_engine,
+    _build_std_samr_engine,
+    _build_openstd_engine,
+    _build_bangumi_engine,
+    _build_so_engine,
+    _build_shenma_engine,
+    _build_qwant_engine,
+    _build_ecosia_engine,
+    _build_douban_movie_engine,
+    _build_zdic_engine,
+    _build_iplant_engine,
+    _build_people_daily_engine,
+    _build_flk_law_engine,
+    _build_wikisource_engine,
+    _build_weibo_hot_engine,
+    _build_douyin_hot_engine,
+    _build_csdn_engine,
+    _build_wallstreetcn_engine,
+    _build_weather_cn_engine,
+    _build_google_news_engine,
+    _build_met_museum_engine,    _build_open_library_engine,
+    _build_weread_engine,
+    _build_douban_book_engine,
+    _build_fred_engine,
+    _build_fx_rate_engine,
+    _build_worldbank_engine,
+    _build_nbs_stats_engine,
+    _build_pubchem_engine,
+    _build_eurostat_engine,
+    _build_gbif_engine,
+    _build_rfc_editor_engine,
+    _build_twitter_syndication_engine,
+    _build_deps_dev_engine,
+    _build_endoflife_engine,
+    _build_biorxiv_engine,
+    _build_un_comtrade_engine,
+    _build_uniprot_engine,
+    _build_rcsb_pdb_engine,
+    _build_courtlistener_engine,
+    _build_gutenberg_engine,
+    _build_wayback_cdx_engine,
+    _build_usgs_engine,
+    _build_nasa_cmr_engine,
+    _build_free_dictionary_engine,
+    _build_baidu_baike_engine,
+    _build_pypi_engine,
+    _build_clinicaltrials_engine,
+    _build_openfda_engine,
+    _build_juejin_engine,
+    _build_models_dev_engine,
+    _build_finviz_engine,
+    _build_yahoo_finance_engine,
+    _build_sec_companyfacts_engine,
+    _build_frankfurter_engine,
+    _build_seeking_alpha_engine,
+    _build_qweather_engine,
+    _build_wenshu_engine,
+    _build_jin10_engine,
+    _build_octen_engine,
+    _build_imdb_engine,
+    _build_thesportsdb_engine,
+    _build_itunes_engine,
+    _build_opencorporates_engine,
+    _build_google_patents_engine,
+    _build_marginalia_engine,
+    _build_wiby_engine,
+    _build_cnii_engine,
+    _build_ndl_engine,
+    _build_kor_law_engine,
+    _build_hatena_bookmark_engine,
+    _build_dnb_engine,
+    _build_doaj_engine,
+    _build_europeana_engine,
+    _build_hal_engine,
+    _build_eu_opendata_engine,
+    _build_open_meteo_engine,
+    _build_searchmysite_engine,
+    _build_lieu_engine,
+    _build_opensky_engine,
+    _build_electricity_maps_engine,
+    _build_usda_engine,
+    _build_tatoeba_engine,
+    _build_figshare_engine,
+    _build_tencent_kline_engine,
+    _build_qq_music_engine,
+    _build_github_engine,
+    _build_rss_feed_engine,
+)
+# 批次十（engines_builders_batch10）：engines_builders.py 聚合层暂未转出该模块，
+# 先直连其源模块——engines_builders.py 补转出后可并入上方聚合 import。
+# zhihu_global 用别名：与上方 engines_builders 转出的 cn 旧版同名，F811 门禁
+# 禁止静默重定义；别名让「batch10 取代 cn 版」在注册表处显式可见。
+from engines_builders_batch10 import (  # noqa: F401
+    _build_zhihu_global_engine as _build_zhihu_global_engine_v2,
+)
+# 批次十一（engines_builders_batch11）：图源补强。Wikimedia Commons 的
+# query.pages 是 pageid 字典，声明式 output_map 取不到字段，必须有 builder。
+from engines_builders_batch11 import (  # noqa: F401
+    _build_wikimedia_commons_engine,
+)
+
+
+
+class _LazyLogger:
+    """延迟创建 logger，避免 import logging 的重量级 import 链（traceback→dataclasses→inspect）。"""
+    _logger = None
+
+    def __getattr__(self, name: str):
+        if _LazyLogger._logger is None:
+            import logging
+            _LazyLogger._logger = logging.getLogger("unified_search.engines")
+            if not _LazyLogger._logger.handlers:
+                _LazyLogger._logger.setLevel(logging.WARNING)
+                _LazyLogger._logger.addHandler(logging.StreamHandler(sys.stderr))
+        return getattr(_LazyLogger._logger, name)
+
+
+logger = _LazyLogger()
+
+
+# 「未知引擎」警告按名字每进程只报一次。
+# 有界（2026-09-29）：键可能来自用户/路由输入，长驻 MCP server 收到大量
+# 非常规引擎名时，无界集合会随进程寿命线性增长。超限即整体清空——去重是
+# best-effort，丢一轮去重只多几行重复警告，远比无界增长便宜。
+_UNKNOWN_ENGINE_WARNED_MAX = 256
+_unknown_engine_warned: set[str] = set()
+
+
+def _warn_unknown_engine(engine: str) -> None:
+    """未注册引擎的日志警告（同名每进程一次）。
+
+    why 去重：同一次调用里，主 combo、次域补充、恢复链会各试一遍同一个名字，
+    实测同一条警告重复 3–4 行。重复并不增加信息，只会把真信号淹掉。
+    机器可读的那一份不在这里——它走 outcome 的 `status=unknown-engine` + detail，
+    最终进响应的 `errors`（见 engine_dispatch 的前置拦截）。
+    """
+    if engine in _unknown_engine_warned:
+        return
+    if len(_unknown_engine_warned) >= _UNKNOWN_ENGINE_WARNED_MAX:
+        _unknown_engine_warned.clear()
+    _unknown_engine_warned.add(engine)
+    logger.warning(
+        "未知引擎: %s（registry 查无此名：拼写错误，或该源未随本版发布）", engine)
+
+
+def is_registered(engine: str) -> bool:
+    """该名字在本版 registry 里是否存在。
+
+    唯一判据就是 `get_registry()`，与 `run_engine` 内部的 `registry.get()` 同源：
+    调用方（engine_dispatch 的前置拦截）必须与执行层看到同一个注册表，各写一份
+    `in` 判断迟早漂移。
+    """
+    return engine in get_registry()
+
+
+def unknown_requested_engines(engine_override: str) -> list[str]:
+    """显式点名的引擎里，本版 registry 查无的那些（按输入顺序，去重）。
+
+    why：`--engine nope` 此前产出
+    `status=completed / errors=[] / engines_used=['nope'] / count=0`——调用方
+    （Agent）据此判定「网上没有这个信息」并停止追问，而真相是这个名字在本版
+    registry 里根本不存在（拼写错、或该源未随本版发布）。它与「缺 API Key 的
+    静默 no-results」是同一根因，处置也应一致：变成可行动的 error。
+
+    **为什么判在请求层而不是执行层**：执行层收到的引擎名可以是测试替身
+    （test_serial_hedge 的 quick_primary、test_race_fast_budget 的 fast_garbage
+    都从 dispatch 入口走），在那里按 registry 判「未注册」会误伤替身——实测
+    11 个 dispatch 用例集体变红。而「用户点了个不存在的名字」是**请求级**的
+    事实，判在请求层才既不误伤、又落在正确的抽象层。
+
+    住在 engines.py 而不是 search.py：判据本体（is_registered）在这里，而
+    search.py 已压着 1000 行硬上限（本函数曾在那边，被体积门禁拦下）。这里只做
+    「请求串 → 查无的名字」这一层薄封装，不重复 `in` 判断。
+    """
+    if not engine_override or engine_override == "auto":
         return []
-    return wrapper
+    unknown: list[str] = []
+    for raw in str(engine_override).split(","):
+        name = raw.strip()
+        if name and name not in unknown and not is_registered(name):
+            unknown.append(name)
+    return unknown
+
+
+@contextlib.contextmanager
+def _sys_path_tmp(path: str):
+    """临时把 path 加入 sys.path，退出时移除——只移除自己添加的那一次。
+
+    append 而非 insert(0)：sub-skills 顶层有与 scripts 同名的模块
+    （health_check 等），插到前面会劫持 scripts 下同名模块的解析。
+    """
+    added = path not in sys.path
+    if added:
+        sys.path.append(path)
+    try:
+        yield
+    finally:
+        if added:
+            sys.path.remove(path)
 
 
-def _run(cmd: list[str], timeout: float = 8, engine_name: str = "?") -> str:
-    """执行命令，超时/异常不抛。"""
-    try:
-        r = subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)
-        if r.returncode == 0:
-            return r.stdout
-        tail = (r.stderr or "").strip()[:200]
-        logger.warning(f"引擎 {engine_name} 失败 (rc={r.returncode}): {tail}")
-        return r.stdout if r.stdout.strip() else ""
-    except subprocess.TimeoutExpired:
-        logger.warning(f"引擎 {engine_name} 超时 (>{timeout}s)")
-    except FileNotFoundError as e:
-        logger.error(f"引擎 {engine_name} CLI 缺失: {e}")
-    except Exception as e:
-        logger.error(f"引擎 {engine_name} 异常: {type(e).__name__}: {e}")
-    return ""
-
-
-def _resolve(template: list[str] | str, query: str, n: int, **extra: Any) -> list[str] | str:
-    """替换模板占位符。"""
-    if isinstance(template, list):
-        return [_resolve(item, query, n, **extra) for item in template]
-    s = template.replace("{query}", query).replace("{n}", str(n))
-    s = s.replace("{TIMESTAMP}", str(int(time.time())))
-    for key, val in extra.items():
-        s = s.replace(f"{{{key}}}", str(val))
-    if s.startswith("~"):
-        s = str(Path.home() / s[1:])
-    return re.sub(r"\{([A-Z_][A-Z0-9_]*)\}", lambda m: os.environ.get(m.group(1), m.group(0)), s)
-
-
-def _extract_items(data: Any, path: str) -> list[dict]:
-    """从嵌套 dict 按路径提取列表。"""
-    obj = data
-    for part in path.split("."):
-        if isinstance(obj, dict):
-            obj = obj.get(part, [])
-        else:
-            return []
-    return obj if isinstance(obj, list) else []
+def _build_local_search_engine(spec: dict[str, Any]) -> Any:
+    """进程内调用 local-search 子技能，避免 subprocess 冷启动（~300-500ms/次）。
 
+    直接 import search_v3.search_engines，复用其智能路由/健康过滤/批量并行，
+    输出与 unified-search 一致的 schema。安全降级：import 失败时回退
+    原 subprocess 调用，保证功能不丢。
+    """
+    import os as _os
+    import subprocess as _subprocess
 
-def _make_field_parser(path: str, fields: dict[str, str]) -> Callable:
-    """构造声明式 parser。"""
-    def parser(data: Any) -> list[dict[str, Any]]:
-        items = _extract_items(data, path) if isinstance(data, dict) else []
-        results = []
-        for item in items:
-            if not isinstance(item, dict):
-                continue
-            r = {ok: iv[:300] if ok == "snippet" and isinstance(iv, str) else iv
-                 for ok, ik in fields.items() if (iv := item.get(ik, ""))}
-            if r.get("title") or r.get("url"):
-                results.append(r)
-        return results[:10]
-    return parser
-
-
-def _build_cli_engine(spec: dict[str, Any]) -> Any:
     cmd_template = spec.get("cmd", [])
     search_args = spec.get("search_args", [])
-    env_overrides = spec.get("env", {})
 
     @safe_search
     def _engine(query: str, n: int = 5, timeout: float = 8, mode: str = "fast", **kwargs) -> list[dict[str, Any]]:
-        cmd = _resolve(cmd_template, query, n, mode=mode)
-        args = _resolve(search_args, query, n, mode=mode)
+        # 回退路径无条件依赖 _resolve：必须放 try 外。放 try 首行时，import 失败
+        # 会被下面的 except Exception 吞掉，回退路径读 _resolve_tpl 直接
+        # UnboundLocalError——「进程内失败回退 subprocess」的语义被架空
+        # （ef4e509 同类：分支内 import 污染同函数后续读取，静态门禁
+        # test_local_import_shadows_global 依赖此处不报）。
+        from engines_base import _resolve as _resolve_tpl
+        # 进程内优先（省 subprocess 冷启动）
+        try:
+            sub_dir = Path(__file__).resolve().parent.parent / "sub-skills" / "local-search"
+            # 临时加路径 import，完事自动清理（助手语义见定义处）
+            with _sys_path_tmp(str(sub_dir)):
+                import search_v3
+            res = search_v3.search_engines(
+                query, engines=None, n=n, timeout=float(timeout),
+                # None = 读子技能 config 的 max_parallel_engines（此前硬编码 5，
+                # config 调并发对主链路径不生效）；预算=单引擎超时，local-first
+                # 是「快」语义，墙钟不吃 30s 历史默认。
+                max_parallel=None, total_budget=float(timeout),
+                skip_cache=bool(kwargs.get("skip_cache", False)),
+                mode=mode,
+                since=kwargs.get("since"), until=kwargs.get("until"),
+                sort=kwargs.get("sort"),
+            )
+            results = res.get("results") or []
+            # 与子进程路径一致：每条带 _engine 标记，source 保持子引擎名
+            for r in results:
+                if isinstance(r, dict) and "error" not in r:
+                    r.setdefault("_engine", r.get("source") or "local_search")
+            if results:
+                return results
+        except Exception:
+            pass  # 进程内失败回退 subprocess
+        # 回退：subprocess 调用（原行为）
+        cmd = _resolve_tpl(cmd_template, query, n, mode=mode)
+        args = _resolve_tpl(search_args, query, n, mode=mode)
         if not cmd:
             return []
-        env = os.environ.copy()
-        env.update(env_overrides)
-        return _parse_text_output(_run(cmd + args, timeout=timeout, engine_name=spec.get("_name", "cli")),
-                                  spec.get("_name", "cli"))
-    return _engine
-
-
-def _build_http_engine(spec: dict[str, Any]) -> Any:
-    """统一 HTTP 引擎构造（GET/POST）。"""
-    url_template = spec.get("url", "")
-    headers = spec.get("headers", {"Content-Type": "application/json"})
-    query_param = spec.get("query_param", "q")
-    fmt = spec.get("format", "")
-    timeout = spec.get("timeout", 8)
-    extra_params = spec.get("extra_params", {})
-    output_map = spec.get("output_map", {})
-    is_get = spec.get("method", "GET") == "GET"
-    body_template = spec.get("body", {})
-
-    @safe_search
-    def _engine(query: str, n: int = 5, _timeout: float | None = None, depth: str = "fast", **kwargs) -> list[dict[str, Any]]:
-        to = _timeout or timeout
-        import urllib.parse as up
-
-        if is_get:
-            resolved_url = _resolve(url_template, query, n)
-            separator = "&" if "?" in resolved_url else "?"
-            full_url = f"{resolved_url}{separator}{query_param}={up.quote(query)}"
-            if fmt:
-                full_url += f"&format={fmt}"
-            for k, v in extra_params.items():
-                full_url += f"&{k}={up.quote(_resolve(str(v), query, n))}"
-            req = urllib.request.Request(full_url, headers={k: _resolve(v, query, n) for k, v in headers.items()})
-        else:
-            body: dict[str, Any] = {}
-            for k, v in body_template.items():
-                resolved = _resolve(str(v), query, n)
-                if k == "search_depth":
-                    body[k] = depth
-                elif resolved.lower() == "true":
-                    body[k] = True
-                elif resolved.lower() == "false":
-                    body[k] = False
-                else:
-                    try:
-                        body[k] = int(resolved)
-                    except ValueError:
-                        try:
-                            body[k] = float(resolved)
-                        except ValueError:
-                            body[k] = resolved
-            req = urllib.request.Request(url_template, data=json.dumps(body).encode("utf-8"),
-                                         headers={k: _resolve(v, query, n) for k, v in headers.items()})
-
+        env = _os.environ.copy()
+        env.update(spec.get("env", {}) or {})
+        proc = _subprocess.run(cmd + args, capture_output=True, text=True,
+                               encoding="utf-8", errors="replace",
+                               timeout=timeout, env=env)
+        if proc.returncode != 0:
+            return []
         try:
-            with urllib.request.urlopen(req, timeout=to) as resp:
-                raw = resp.read().decode("utf-8")
-                if fmt == "xml":
-                    return _parse_xml(raw, spec.get("_name", ""))
-                data = json.loads(raw)
-                if output_map:
-                    return _make_field_parser(output_map.get("items", ""), {
-                        "title": output_map.get("item_title", "title"),
-                        "url": output_map.get("item_url", "url"),
-                        "snippet": output_map.get("item_summary", "snippet"),
-                        "source": output_map.get("item_source", "source"),
-                    })(data)
-                return _parse_generic(data, spec.get("_name", ""))
-        except (urllib.error.URLError, urllib.error.HTTPError, Exception) as e:
-            logger.warning(f"HTTP 引擎失败: {e}")
+            data = json.loads(proc.stdout)
+        except (json.JSONDecodeError, ValueError):
             return []
+        results = data.get("results") or []
+        for r in results:
+            if isinstance(r, dict) and "error" not in r:
+                r.setdefault("_engine", r.get("source") or "local_search")
+        return results
     return _engine
 
 
-_BUILDERS = {"cli": _build_cli_engine, "http": _build_http_engine}
-
-
-# ── 通用解析器 ─────────────────────────────────────────────────────────────────
-
-def _parse_text_output(text: str, engine_name: str) -> list[dict[str, Any]]:
-    """通用 CLI 文本解析：优先 JSON，其次结构化文本。"""
-    try:
-        data = json.loads(text.strip())
-        if isinstance(data, list):
-            return [{"title": i.get("title", ""), "url": i.get("url", ""),
-                     "snippet": i.get("snippet", i.get("content", ""))[:300],
-                     "source": engine_name} for i in data if isinstance(i, dict)]
-        if isinstance(data, dict):
-            items = data.get("results", data.get("items", data.get("data", [])))
-            if isinstance(items, list):
-                return [{"title": i.get("title", ""), "url": i.get("url", ""),
-                         "snippet": i.get("snippet", i.get("content", ""))[:300],
-                         "source": engine_name} for i in items if isinstance(i, dict)]
-    except (json.JSONDecodeError, ValueError):
-        pass
-
-    results, cur = [], {}
-    seen_url = False
-    for line in text.split("\n"):
-        s = line.strip()
-        if s.startswith("### "):
-            if cur:
-                results.append(cur)
-            cur = {"title": re.sub(r'^\d+\.\s*', '', s[4:].strip()), "source": engine_name,
-                   "score": max(1.0 - len(results) * 0.1, 0.1)}
-            seen_url = False
-        elif s.startswith("- **URL**: ") and cur:
-            cur["url"] = s[11:].strip()
-            seen_url = True
-        elif s.startswith("- ") and not s.startswith("- **") and seen_url and cur:
-            cur["snippet"] = " ".join(s[2:].strip().split())[:300]
-            seen_url = False
-    if cur:
-        results.append(cur)
-    return results[:10]
-
-
-def _parse_xml(text: str, engine_name: str) -> list[dict[str, Any]]:
-    """解析 Atom XML（arXiv 等）。"""
-    import xml.etree.ElementTree as ET
-    results = []
-    try:
-        root = ET.fromstring(text)
-        ns = {"atom": "http://www.w3.org/2005/Atom"}
-        entries = root.findall(".//atom:entry", ns) or root.findall(".//{http://www.w3.org/2005/Atom}entry")
-        for entry in entries:
-            title = entry.findtext("atom:title", "", ns).strip().replace("\n", " ")[:200]
-            summary = entry.findtext("atom:summary", "", ns).strip().replace("\n", " ")[:300]
-            entry_id = entry.findtext("atom:id", "", ns)
-            url = entry_id
-            for link in entry.findall("atom:link", ns):
-                if link.get("title") == "pdf":
-                    url = link.get("href", url)
-                    break
-            if title:
-                results.append({"title": title, "url": url, "snippet": summary, "source": engine_name})
-    except ET.ParseError:
-        pass
-    return results
-
-
-def _parse_generic(data: dict[str, Any], engine_name: str = "?") -> list[dict[str, Any]]:
-    """通用 JSON 解析：自动探测常见字段。"""
-    items = None
-    for key in ["results", "items", "data", "works", "message.items"]:
-        if "." in key:
-            parts = key.split(".")
-            obj = data
-            for p in parts:
-                obj = obj.get(p, {}) if isinstance(obj, dict) else {}
-            if isinstance(obj, list):
-                items = obj
-                break
-        elif isinstance(data, dict) and key in data and isinstance(data[key], list):
-            items = data[key]
-            break
-
-    if items is None and isinstance(data, dict):
-        for v in data.values():
-            if isinstance(v, dict):
-                for key in ["results", "items", "value"]:
-                    if key in v and isinstance(v[key], list):
-                        items = v[key]
-                        break
-            if items:
+_BUILDERS = {
+    "cli": _build_cli_engine,
+    "http": _build_http_engine,
+    "html": _build_html_engine,
+    "local_search": _build_local_search_engine,
+    "exa": _build_exa_engine,
+    "anysearch": _build_anysearch_engine,
+    "em_miaoxiang": _build_em_miaoxiang_engine,
+    "cninfo": _build_cninfo_engine,
+    "sina_quote": _build_sina_quote_engine,
+    "tencent_quote": _build_tencent_quote_engine,
+    "em_flow": _build_em_flow_engine,
+    "wechat_sogou": _build_wechat_sogou_engine,
+    "hackernews": _build_hackernews_engine,
+    "stackoverflow": _build_stackoverflow_engine,
+    "std_samr": _build_std_samr_engine,
+    "openstd": _build_openstd_engine,
+    "bangumi": _build_bangumi_engine,
+    "douban_movie": _build_douban_movie_engine,
+    "zdic": _build_zdic_engine,
+    "iplant": _build_iplant_engine,
+    "people_daily": _build_people_daily_engine,
+    "flk_law": _build_flk_law_engine,
+    "wikisource": _build_wikisource_engine,
+    "weibo_hot": _build_weibo_hot_engine,
+    "douyin_hot": _build_douyin_hot_engine,
+    "csdn": _build_csdn_engine,
+    "wallstreetcn": _build_wallstreetcn_engine,
+    "weather_cn": _build_weather_cn_engine,
+    "google_news": _build_google_news_engine,
+    "met_museum": _build_met_museum_engine,
+    "github": _build_github_engine,
+    "google_scholar": _build_google_scholar_engine,
+    "v2ex": _build_v2ex_engine,
+    "tineye": _build_tineye_engine,
+    "bing_rss": _build_bing_rss_engine,
+    "ths_hot": _build_ths_hot_engine,
+    "cls_telegraph": _build_cls_telegraph_engine,
+    "twitter_syndication": _build_twitter_syndication_engine,
+    "deps_dev": _build_deps_dev_engine,
+    "endoflife": _build_endoflife_engine,
+    "biorxiv": _build_biorxiv_engine,
+    "un_comtrade": _build_un_comtrade_engine,
+    "em_global_news": _build_em_global_news_engine,
+    "eastmoney": _build_eastmoney_engine,
+    "itotii": _build_itotii_engine,
+    "baidu_hot": _build_baidu_hot_engine,
+    "toutiao_hot": _build_toutiao_hot_engine,
+    "bilibili_hot": _build_bilibili_hot_engine,
+    "open_library": _build_open_library_engine,
+    "weread": _build_weread_engine,
+    "douban_book": _build_douban_book_engine,
+    "zhihu_global": _build_zhihu_global_engine_v2,  # 补强版在 batch10（候选池取满 + 错误行动提示；cn 版被取代，cn 冻结在行数门禁上限）
+    "zhihu_user": _build_zhihu_user_engine,
+    "fred": _build_fred_engine,
+    "so": _build_so_engine,
+    "shenma": _build_shenma_engine,
+    "qwant": _build_qwant_engine,
+    "ecosia": _build_ecosia_engine,
+    "fx_rate": _build_fx_rate_engine,
+    "worldbank": _build_worldbank_engine,
+    "nbs_stats": _build_nbs_stats_engine,
+    "pubchem": _build_pubchem_engine,
+    "eurostat": _build_eurostat_engine,
+    "gbif": _build_gbif_engine,
+    "rfc_editor": _build_rfc_editor_engine,
+    "uniprot": _build_uniprot_engine,
+    "rcsb_pdb": _build_rcsb_pdb_engine,
+    "courtlistener": _build_courtlistener_engine,
+    "gutenberg": _build_gutenberg_engine,
+    "wayback_cdx": _build_wayback_cdx_engine,
+    "usgs": _build_usgs_engine,
+    "nasa_cmr": _build_nasa_cmr_engine,
+    "free_dictionary": _build_free_dictionary_engine,
+    "baidu_baike": _build_baidu_baike_engine,
+    "pypi": _build_pypi_engine,
+    "clinicaltrials": _build_clinicaltrials_engine,
+    "openfda": _build_openfda_engine,
+    "juejin": _build_juejin_engine,
+    "models_dev": _build_models_dev_engine,
+    "finviz": _build_finviz_engine,
+    "yahoo_finance": _build_yahoo_finance_engine,
+    "sec_companyfacts": _build_sec_companyfacts_engine,
+    "frankfurter": _build_frankfurter_engine,
+    "seeking_alpha": _build_seeking_alpha_engine,
+    "qweather": _build_qweather_engine,
+    "wenshu": _build_wenshu_engine,
+    "jin10": _build_jin10_engine,
+    "octen": _build_octen_engine,
+    "bocha": _build_bocha_engine,
+    "parallel": _build_parallel_engine,
+    "parallel_free": _build_parallel_free_engine,
+    "seltz": _build_seltz_engine,
+    "you": _build_you_engine,
+    "bocha_ai": _build_bocha_ai_engine,
+    "imdb": _build_imdb_engine,
+    "thesportsdb": _build_thesportsdb_engine,
+    "itunes": _build_itunes_engine,
+    "opencorporates": _build_opencorporates_engine,
+    "google_patents": _build_google_patents_engine,
+    "marginalia": _build_marginalia_engine,
+    "wiby": _build_wiby_engine,
+    "rss_feed": _build_rss_feed_engine,
+    "cnii": _build_cnii_engine,
+    "ndl": _build_ndl_engine,
+    "kor_law": _build_kor_law_engine,
+    "hatena_bookmark": _build_hatena_bookmark_engine,
+    "dnb": _build_dnb_engine,
+    "doaj": _build_doaj_engine,
+    "europeana": _build_europeana_engine,
+    "hal": _build_hal_engine,
+    "eu_opendata": _build_eu_opendata_engine,
+    "open_meteo": _build_open_meteo_engine,
+    "searchmysite": _build_searchmysite_engine,
+    "lieu": _build_lieu_engine,
+    "opensky": _build_opensky_engine,
+    "electricity_maps": _build_electricity_maps_engine,
+    "usda": _build_usda_engine,
+    "tatoeba": _build_tatoeba_engine,
+    "figshare": _build_figshare_engine,
+    "tencent_kline": _build_tencent_kline_engine,
+    "qq_music": _build_qq_music_engine,
+    "who_don": _build_who_don_engine,
+    "who_gho": _build_who_gho_engine,
+    "pubmed": _build_pubmed_engine,
+    "core": _build_core_engine,
+    "gdacs": _build_gdacs_engine,
+    "obis": _build_obis_engine,
+    "worms": _build_worms_engine,
+    "energy_charts": _build_energy_charts_engine,
+    "jolpica": _build_jolpica_engine,
+    "openf1": _build_openf1_engine,
+    "openligadb": _build_openligadb_engine,
+    "artic": _build_artic_engine,
+    "cleveland": _build_cleveland_engine,
+    "wikimedia_commons": _build_wikimedia_commons_engine,
+    "tvmaze": _build_tvmaze_engine,
+    "jikan": _build_jikan_engine,
+    "deezer": _build_deezer_engine,
+    "listenbrainz": _build_listenbrainz_engine,
+    "egov_law": _build_egov_law_engine,
+    "k10plus": _build_k10plus_engine,
+    "ror": _build_ror_engine,
+    "soilgrids": _build_soilgrids_engine,
+    "noaa_swpc": _build_noaa_swpc_engine,
+    "satnogs": _build_satnogs_engine,
+    "tle_mirror": _build_tle_mirror_engine,
+    "nhtsa_vpic": _build_nhtsa_vpic_engine,
+}
+
+# 语义型引擎：把 query 当自然语言语义检索，不识别平台原生结构化语法
+# （from:/repo:/site:/until: 等）。对其剥掉字段只留核心词，避免污染相关性。
+# 透传型/垂直源（local_*、social/academic/code、github 等）保持原 query，让平台语法生效。
+# local_search 是聚合器（子引擎 local_bing 等透传平台语法），不在语义集内。
+_SEMANTIC_ENGINES = frozenset({
+    "byted", "bocha", "anysearch", "tavily", "exa", "octen",
+    "uapi", "searxng", "parallel", "you", "bocha_ai",
+})
+
+# ── 实体型引擎的查询规范化（2026-09-06 live 金标教训）─────────────────────
+# 自然语言句直达实体搜索接口会全军覆没：「NASA founding year」wikidata 0 条
+# （裸「NASA」7 条）、「where is Eiffel Tower」同型且拖满 11s 超时、
+# 「Cristiano Ronaldo club」thesportsdb 抖动放大。接口只吃实体名——分发层
+# 统一剥疑问前缀与属性词，保留核心实体。只作用于实体型引擎；通用 web 引擎
+# 不动（属性词对全文检索是有用信号）。
+_ENTITY_QUERY_ENGINES = frozenset({"wikidata", "thesportsdb", "local_openstreetmap"})
+_ENTITY_Q_PREFIX_RE = re.compile(
+    r"(?i)^\s*(where\s+(is|are|was|were)|where'?s|who\s+(is|are|was)|who'?s"
+    r"|what\s+(is|are)|when\s+(was|is|did)|which\s+\w+\s+(is|was)"
+    r"|哪里|在哪[里儿]?|是谁?|是什么|什么时候|哪一[个国年])\s*",
+)
+# 英文属性词带 \b（空格分词下词边界天然成立）；中文属性词不进此表——
+# CJK 字符全是 \w，「周杰伦专辑」「周杰伦的专辑」这类无空格连写永远撞不上
+# 词边界（2026-09-06 审查实锤漏剥），改由下方尾部循环剥离。
+_ENTITY_ATTR_RE = re.compile(
+    r"(?i)\b(founding year|founded|established|headquarters?|located( in)?"
+    r"|population|capital|address|club|team|stadium|league|album|discography"
+    r"|song|movie|film|director|cast)\b",
+)
+# 中文属性词（含繁体）：只在查询尾部剥离——中文自然语序是实体名在前、
+# 属性词在后；不碰串中，「电影频道」「歌曲排行榜」这类实体名内含属性词，
+# 串中剥离会误伤。长词优先，避免「成立年份」被「成立」截断残留「年份」。
+_ENTITY_ATTR_TAILS = (
+    "成立年份", "成立", "创立", "創立", "总部", "總部", "位置", "球队", "球隊",
+    "俱乐部", "俱樂部", "职能", "職能", "职责", "職責", "专辑", "專輯",
+    "歌曲", "电影", "電影", "导演", "導演", "主演",
+)
+
+
+def _entity_core_query(query: str) -> str:
+    """实体搜索接口的查询规范化：剥疑问前缀+属性词，保留核心实体名。"""
+    raw = (query or "").strip()
+    q = _ENTITY_Q_PREFIX_RE.sub(" ", raw)
+    # 中文疑问词多为尾部后置（「清华大学 在哪里」），与英文前缀形态分开剥
+    q = re.sub(r"(?i)(在哪里|在哪|哪儿|哪里|是什么|是谁|什么时候|有哪些)\s*[?？]?\s*$",
+               " ", q)
+    q = _ENTITY_ATTR_RE.sub(" ", q)
+    # 空白归一化必须在尾部循环前：前缀/属性词剥除用空格替换，尾空格会让
+    # endswith 判空（「清华大学总部 」漏剥总部）
+    q = re.sub(r"\s+", " ", q).strip()
+    # 中文属性词尾部循环剥 + 「的」后缀收尾：「周杰伦的专辑」→「周杰伦的」
+    # →「周杰伦」。剩单字不剥（「美的」是实体，剥成「美」即误伤）
+    changed = True
+    while changed:
+        changed = False
+        for tail in _ENTITY_ATTR_TAILS:
+            if q.endswith(tail) and len(q) > len(tail) + 1:
+                q = q[:-len(tail)].strip()
+                changed = True
                 break
-
-    if not items or not isinstance(items, list):
-        return []
-
-    results = []
-    for i in items:
-        if not isinstance(i, dict):
-            continue
-        title = i.get("title", "")
-        if isinstance(title, list):
-            title = title[0] if title else ""
-        url = i.get("url", i.get("URL", i.get("html_url", "")))
-        snippet = (i.get("snippet", i.get("content", i.get("summary", i.get("description", "")))))[:300]
-        score = i.get("score", i.get("relevance_score", 0.5))
-        results.append({"title": str(title)[:200], "url": str(url),
-                        "snippet": str(snippet), "score": score, "source": engine_name})
-    return results[:10]
+        if not changed and q.endswith("的") and len(q) > 3:
+            q = q[:-1].strip()
+            changed = True
+    q = re.sub(r"[?？:：，,]+", " ", q)
+    q = re.sub(r"\s+", " ", q).strip()
+    return q or raw
 
 
-# ── 引擎注册表 ─────────────────────────────────────────────────────────────────
-
 _engine_registry: dict[str, Any] = {}
+_engine_specs: dict[str, dict[str, Any]] = {}
 _engine_registry_loaded = False
+_registry_stamp: float | None = None
 
 
 def _load_registry():
-    global _engine_registry, _engine_registry_loaded
+    global _engine_registry, _engine_specs, _engine_registry_loaded, _registry_stamp
     if _engine_registry_loaded:
         return
     cfg = load_config()
     engines = get_engines(cfg)
+    # 本地个性化引擎插件（不随公开仓库提交）
+    try:
+        from local_engines import register as _register_local
+        _register_local(_BUILDERS, engines)
+    except ImportError:
+        pass
     registry = {}
     for name, spec in engines.items():
         spec = dict(spec)
         spec["_name"] = name
+        # local_search 走进程内 builder（config 里 type=cli，这里显式路由到专用实现）
+        if name == "local_search":
+            spec["type"] = "local_search"
+        # anysearch：type 已在 config.yaml 显式声明（引擎声明来源），
+        # 不再运行时硬覆盖；若旧配置缺 type 字段，保底路由到进程内 builder。
+        if name == "anysearch" and spec.get("type", "cli") not in _BUILDERS:
+            spec["type"] = "anysearch"
         builder = _BUILDERS.get(spec.get("type", "cli"))
         if builder:
             registry[name] = builder(spec)
         else:
             logger.warning(f"未知引擎类型: {spec.get('type')} (引擎 {name})")
     _engine_registry = registry
+    # spec 侧表：engine_env 缺 env 检测等需要原始声明（registry 值是闭包）
+    global _engine_specs
+    _engine_specs = engines
     _engine_registry_loaded = True
+    try:
+        from config import config_stamp
+        _registry_stamp = config_stamp()
+    except ImportError:
+        _registry_stamp = None
 
 
 def get_registry() -> dict[str, Any]:
+    """引擎注册表。config 变更（综合 mtime 指纹）时自动重建——
+    增删引擎/改 qps/换声明等配置无需重启进程。"""
+    global _engine_registry, _engine_specs, _engine_registry_loaded, _registry_stamp
+    try:
+        from config import config_stamp
+        stamp = config_stamp()
+    except ImportError:
+        stamp = _registry_stamp
+    if _engine_registry_loaded and _registry_stamp is not None and stamp != _registry_stamp:
+        logger.info("config 变更 → 重建引擎注册表")
+        _engine_registry = {}
+        _engine_specs = {}
+        _engine_registry_loaded = False
     _load_registry()
     return _engine_registry
 
 
-def available_engines() -> list[str]:
-    return sorted(get_registry().keys())
+def get_engine_spec(name: str) -> dict[str, Any] | None:
+    """返回引擎原始声明（spec），无此引擎返回 None。
 
+    registry 值是构建后的闭包，原始 spec（url/headers/required_env 等）
+    存侧表供 engine_env 缺 env 检测等调用方使用。
+    """
+    get_registry()  # 触发热重建检查，保持侧表与注册表同步
+    return _engine_specs.get(name)
 
-def search(query: str, engine: str, n: int = 5, timeout: float = 8, depth: str = "fast", mode: str = "fast") -> list[dict[str, Any]]:
-    """统一引擎调用入口；失败返回空 list，不抛异常。"""
+
+def available_engines(routable_only: bool = False) -> list[str]:
+    """可实例化的引擎名（有 builder）。
+
+    routable_only=True 时再按「真能用」过滤（密钥齐全、后端依赖就位、
+    未被熔断/禁用）——这是与 --list-engines --detail 的 routable 同一判定。
+
+    routable_only 此前没被实现（本函数不收参数），调用侧传进来会抛
+    TypeError 并被 except TypeError 吞掉回退到全量——`--list-engines
+    --routable-only` 因此静默返回全部引擎，用户以为筛过了。
+    """
+    names = sorted(get_registry().keys())
+    if not routable_only:
+        return names
+    try:
+        from engine_status import list_routable_engine_ids
+        routable = set(list_routable_engine_ids())
+        return [n for n in names if n in routable]
+    except Exception:
+        return names
+
+
+# ── 单飞合并 + 免费引擎结果数桶化 ──
+# 场景：MCP 并发请求 / 多轨道同查询 → 同一引擎调用重复打上游。
+# 桶化只对免费引擎（cost_factor >= 0.85，与 route 免费判定一致）：n 向上
+# snap 到 10/20/50/100，让同查询不同请求数共享一次执行与缓存；付费引擎
+# 保持精确 n（按结果计费不得放大）。
+_NUM_BUCKETS = (10, 20, 50, 100)
+
+
+def _free_engine(engine: str) -> bool:
+    """是否「免费档」——只有免费档才做 n 桶化。
+
+    此前判据是 cost_factor >= 0.85，而成本表没有 api 档的分支（fallthrough
+    到 1.0），于是 api 档（exa/octen/you/parallel/zhihu_global/tavily 这类
+    按量计费的源）被当成免费：要 5 条被放大到 10 条，按结果计费的接口直接
+    双倍计费。本函数注释一直写着「付费引擎保持精确 n（按结果计费不得放大）」，
+    实现却漏了这一档——现在按声明的档位判，不再依赖魔法阈值。
+    """
+    try:
+        from config import cost_tier_of
+        return cost_tier_of(engine) == "free"
+    except Exception:
+        return False
+
+
+def bucket_n(engine: str, n: int) -> int:
+    """免费引擎：n 向上 snap 到缓存友好桶（≤100）；付费引擎原样。"""
+    n = max(1, min(int(n), 100))
+    if not _free_engine(engine):
+        return n
+    for b in _NUM_BUCKETS:
+        if n <= b:
+            return b
+    return 100
+
+
+def _call_key(query: str, engine: str, n: int, kwargs: dict[str, Any],
+              *, timeout: float | None = None, depth: str = "fast",
+              mode: str = "fast") -> str:
+    """单飞 key：同 query+engine+参数指纹才合并（时间窗/域等影响结果参数
+    必须进 key，否则不同检索被错误合并）。
+
+    2026-10-15 补齐：depth / mode / timeout 同样决定结果，此前只经闭包进
+    `_execute`、没进 key —— 跨档位并发被错误合并：deep 请求可能拿到 fast
+    的结果集（反向亦然，且 follower 还会继承 leader 的 timeout）。三者入 key
+    是治本，不是给合并器打补丁。
+    """
+    import json as _json
+    try:
+        params = _json.dumps(kwargs, sort_keys=True, ensure_ascii=False,
+                             default=str)
+    except Exception:
+        params = str(sorted(kwargs.items()))
+    # timeout 归一化：8 与 8.0 是同一次执行，不该分裂成两个 key
+    if timeout is None or isinstance(timeout, bool):
+        to_seg = str(timeout)
+    elif isinstance(timeout, (int, float)):
+        to_seg = f"{float(timeout):g}"
+    else:
+        to_seg = str(timeout)
+    return f"{engine}|{n}|{query}|{params}|{depth}|{mode}|{to_seg}"
+
+
+def search(query: str, engine: str, n: int = 5, timeout: float = 8, depth: str = "fast", mode: str = "fast", **kwargs) -> list[dict[str, Any]]:
+    """统一引擎调用入口；失败返回空 list，不抛异常。kwargs 透传到引擎 builder（如 since/until 时间窗）。
+
+    并发同调用（同 query+engine+参数）经进程内单飞合并为一次上游执行；
+    免费引擎 n 桶化（见 bucket_n），leader 按桶内最大 n 执行，调用方输出
+    层截断到请求数。
+    """
     registry = get_registry()
     fn = registry.get(engine)
     if not fn:
-        logger.warning(f"未知引擎: {engine}")
+        _warn_unknown_engine(engine)
         return []
-    t0 = time.time()
-    try:
-        results = fn(query, n, timeout, depth=depth, mode=mode)
-    except TypeError:
+    # 引擎可用性由熔断器统一裁决（dispatch 的 breaker.allow / get_negative）。
+    # 这里曾有一份 `neg:{engine}` 负缓存，但两处调用（读/写）与 SearchCache
+    # 的签名都不匹配，异常被 except 吞掉，从未生效过（2026-10-18 复核）。
+    # 不修它：单次失败就全局压制该引擎 45s，比熔断器「连续两次才打开」更激进，
+    # 会在一次网络抖动后误杀健康源。保持单一真源，避免第二套状态。
+    # 语义型引擎不识别平台结构化语法，剥掉字段只留核心词；透传型保持原 query。
+    if engine in _SEMANTIC_ENGINES:
+        query = strip_structured(query)
+    elif engine in _ENTITY_QUERY_ENGINES:
+        # 实体型引擎：自然语言句（疑问前缀/属性词）会让实体搜索接口空结果
+        query = _entity_core_query(query)
+    eff_n = bucket_n(engine, n)
+
+    # 引擎自声明 timeout 是该源的硬上限（此前被调用方 timeout 位置传参整体
+    # 覆盖成死配置——OSM timeout:6 实测跑出 10s+5s curl 守卫 11.3s）。
+    # deep 模式保持原语义：研究场景宁可等，不受 spec 短超时约束。
+    spec_to = (_engine_specs or {}).get(engine) or {}
+    st = spec_to.get("timeout") if isinstance(spec_to, dict) else None
+    call_to = timeout
+    if mode != "deep" and isinstance(st, (int, float)) and 0 < st < timeout:
+        call_to = float(st)
+
+    # key 用 call_to（实际执行时的超时）而非调用方声明的 timeout：被 spec
+    # 夹到同一上限的两次调用是同一执行，应共享单飞而不是白拆两份。
+    key = _call_key(query, engine, eff_n, kwargs,
+                    timeout=call_to, depth=depth, mode=mode)
+
+    def _execute() -> list[dict[str, Any]]:
+        t0 = time.time()
         try:
-            results = fn(query, n, timeout)
+            results = fn(query, eff_n, call_to, depth=depth, mode=mode, **kwargs)
         except Exception as e:
-            logger.error(f"引擎 {engine} 异常: {type(e).__name__}: {e}")
-            results = []
-    except Exception as e:
-        logger.error(f"引擎 {engine} 异常: {type(e).__name__}: {e}")
-        results = []
-    elapsed = time.time() - t0
-    if results and isinstance(results, list):
-        for r in results:
-            if isinstance(r, dict) and "error" not in r:
-                r["_engine"] = engine
-                r["_elapsed"] = round(elapsed, 3)
-    return results if isinstance(results, list) else []
+            # TypeError 可能来自引擎内部逻辑错误而非签名不匹配。
+            # 用 inspect.signature 确认引擎是否接受 depth/mode 参数，
+            # 减少误判：只有引擎函数签名明确不接受这些参数时才回退。
+            _retry_simple = False
+            if isinstance(e, TypeError):
+                try:
+                    sig = inspect.signature(fn)
+                    params = list(sig.parameters.keys())
+                    _retry_simple = not (
+                        "depth" in params or "mode" in params or "kwargs" in params
+                        or any(
+                            p.kind == inspect.Parameter.VAR_KEYWORD
+                            for _, p in sig.parameters.items()
+                        )
+                    )
+                except (ValueError, TypeError):
+                    _retry_simple = False
+            if _retry_simple:
+                try:
+                    results = fn(query, eff_n, timeout)
+                except Exception as e2:
+                    logger.error(f"引擎 {engine} 回退失败: {type(e2).__name__}: {e2}")
+                    results = []
+            else:
+                logger.error(f"引擎 {engine} 异常: {type(e).__name__}: {e}")
+                results = []
+        elapsed = time.time() - t0
+        if results and isinstance(results, list):
+            for r in results:
+                if isinstance(r, dict) and "error" not in r:
+                    r["_engine"] = engine
+                    r["_elapsed"] = round(elapsed, 3)
+        return results if isinstance(results, list) else []
 
+    try:
+        results = engine_coalescer().run(key, _execute)
+    except Exception:
+        results = []
+    results = results or []
+    # 桶化后可能多于请求数：调用方按请求 n 截断语义由 execute_search 融合层负责；
+    # 此处保留全量（缓存键按 eff_n 存，命中时同样可截）。
+    return results
 
 def _cli():
     import argparse
@@ -375,11 +859,11 @@ def _cli():
     parser.add_argument("--list", action="store_true")
     args = parser.parse_args()
     if args.list:
-        print(json.dumps(available_engines(), ensure_ascii=False, indent=2))
+        print(dumps(available_engines()))
         return
     if not args.query:
         parser.error("必须提供 query")
-    print(json.dumps(search(args.query, args.engine, args.n, args.timeout), ensure_ascii=False, indent=2))
+    print(dumps(search(args.query, args.engine, args.n, args.timeout)))
 
 
 if __name__ == "__main__":