diff --git a/.env.example b/.env.example
index 2ac80de..ff2224e 100644
--- a/.env.example
+++ b/.env.example
@@ -41,7 +41,8 @@ RAKUTEN_REQUEST_TIMEOUT_SECONDS=30
 RAKUTEN_MAX_SITE_CONCURRENCY=8
 # 单次抓取的最大尝试次数（含首次）：1 次直发，2 次换 cookie 重试，3 次起动用浏览器兜底
 RAKUTEN_HTTP_MAX_ATTEMPTS=3
-# Akamai cookie 最长复用时长（秒），超时后重新预热
+# Akamai cookie 最长复用时长（秒），超时后清空、由下一次请求的响应重新建立。
+# cookie 是随目标页响应下发的，正常路径不会额外访问站点首页。
 RAKUTEN_SESSION_TTL_SECONDS=1800
 
 # 浏览器兜底：纯 HTTP 被反爬拦截时用 Playwright 取回 cookie 再回灌重试。
diff --git a/README.md b/README.md
index e091b35..aa89111 100644
--- a/README.md
+++ b/README.md
@@ -93,8 +93,8 @@ PC UA 在搜索页、详情页、店铺页上都能拿到完整模板。因此 
 
 - FastAPI 提供 HTTP API
 - Bearer Token 接口鉴权（`Authorization: Bearer <token>`）
-- Akamai cookie 自动预热与复用，过期自动重新预热（仅乐天链路）
-- 失败逐级升级：重新预热 → Playwright 兜底取 cookie → 结构化报错
+- Akamai cookie 随目标页响应自动建立与复用，过期自动清空重建（仅乐天链路；正常路径不额外访问首页）
+- 失败逐级升级：首页换一套 cookie → Playwright 兜底取 cookie → 结构化报错
 - 搜索结果自动剔除混入的 CPC 广告位（乐天）
 - 识别站点深翻页「静默回绕到第 1 页」的行为，避免上游重复入库（乐天）
 - 商品详情按落地域名分派解析，覆盖楽天ブックス / Rakuten Fashion / ビックカメラ 等官方旗舰子站
@@ -660,6 +660,30 @@ trading 加购时的字段选择策略：多规格挑第一个非售罄的 varia
 （账号只读查询通道）**可以重试**——查询只读，重发没有副作用
 （见 [docs/order-gateway.md §11](docs/order-gateway.md#11-账号只读查询通道)）。
 
+## 链路追踪
+
+可选，默认关闭。`RAKUTEN_OTEL_ENABLED=true` + `RAKUTEN_OTEL_ENDPOINT` 指向 OTLP/HTTP
+端点后，三个服务分别以 `rakuten-scraping` / `rakuten-trading` / `rakuten-gateway` 上报。
+
+FastAPI 与 httpx 走自动埋点，但那只覆盖「收到 HTTP 请求」与「发出 httpx 请求」两类
+边界。交易侧的实际工作两者都不是——站点交互走 Playwright，worker 主循环是后台任务，
+所以这一侧手工埋点：
+
+- `order.task`：一笔下单任务的**根 span**（一个 `task_id` 一条 trace），带
+  `order.route`（`execute` / `recovery` / `already_finished`）与终态
+  `order.terminal_status`；被闸门或风控拦下时置 ERROR 并带 `error.code`
+- `order.step.*`：清车 → 加购 → 校验 → 确认页 → 提交 → 付款，每步一个子 span，
+  带 `order.evidence_ref`（可直接定位落盘证据）
+- `site.*`：Playwright 站点交互（`site.add_to_cart`、`site.enter_checkout`、
+  `site.submit_order`、`site.pay` 等）
+- `account_query`：一张只读查询单的根 span，带 `query.outcome`
+
+worker 空转的长轮询（每 30 秒问一次网关有没有活干）刻意不埋点——它们没有信息量，
+量却极大，会把观测后台刷满。
+
+> 解析失败时页面 HTML 会作为 span event 上报（`RAKUTEN_OTEL_SNAPSHOT_MAX_BYTES`
+> 控制上限，默认 2MB），用于事后复现「抓到的内容为什么解析不出预期字段」。
+
 ## 常用环境变量
 
 完整列表见 [.env.example](.env.example)。配置项前缀统一为 `RAKUTEN_`，三个服务共用同一份
@@ -675,6 +699,7 @@ trading 加购时的字段选择策略：多规格挑第一个非售罄的 varia
 - 抓取：`RAKUTEN_MAX_SITE_CONCURRENCY`（默认 `8`，两站各自独立计数）、`RAKUTEN_HTTP_MAX_ATTEMPTS`（默认 `3`）、`RAKUTEN_SESSION_TTL_SECONDS`（默认 `1800`，仅乐天）
 - 浏览器兜底（仅乐天）：`RAKUTEN_BROWSER_FALLBACK_ENABLED`、`RAKUTEN_BROWSER_HEADLESS`、`RAKUTEN_BROWSER_CHANNEL`
 - 代理（需日本 IP 时）：`RAKUTEN_PROXY_SERVER`、`RAKUTEN_PROXY_USERNAME`、`RAKUTEN_PROXY_PASSWORD`、`RAKUTEN_PROXY_BYPASS`
+- 链路追踪（可选，默认关闭；三个服务共用）：`RAKUTEN_OTEL_ENABLED`、`RAKUTEN_OTEL_ENDPOINT`、`RAKUTEN_OTEL_HEADERS`、`RAKUTEN_OTEL_SNAPSHOT_MAX_BYTES`
 
 > 从中国大陆直连实测可用、无需代理；`RAKUTEN_PROXY_SERVER` 留空即可。设置后，所有面向
 > 外部站点的 HTTPX 与 Playwright 流量都会经代理；本机和 Docker 服务间地址由
diff --git a/app/gateway/main.py b/app/gateway/main.py
index 72f13d9..831fd5a 100644
--- a/app/gateway/main.py
+++ b/app/gateway/main.py
@@ -27,6 +27,7 @@ from app.gateway.task_queue import TaskQueue
 from app.shared.api import register_exception_handlers
 from app.shared.config import get_settings
 from app.shared.logging_setup import configure_logging
+from app.shared.telemetry import instrument_app, setup_telemetry, shutdown_telemetry
 
 logger = logging.getLogger(__name__)
 
@@ -96,6 +97,11 @@ async def lifespan(app: FastAPI):
     app.state.container = container
 
     configure_logging(container.settings)
+    # 网关这一侧原先完全没接 telemetry，导致 worker 出站请求带过来的 traceparent
+    # 没人接上：一条下单链路在网关这里断掉，观测后台上只看得到 worker 侧那半截。
+    # CallbackNotifier 的 httpx 客户端是每次投递时才建的（不是 __init__ 里），
+    # 所以放在 build_container() 之后仍然赶在客户端创建之前。
+    setup_telemetry(container.settings, service_name="rakuten-gateway")
     logger.info(
         "网关启动：%s:%s", container.settings.gateway_host, container.settings.gateway_port
     )
@@ -141,6 +147,7 @@ async def lifespan(app: FastAPI):
             # 等在途回调发完（各次发送有超时兜底），避免关停时静默丢通知
             await container.notifier.aclose()
         await container.db.close()
+        shutdown_telemetry()
 
 
 def create_app() -> FastAPI:
@@ -151,6 +158,7 @@ def create_app() -> FastAPI:
     app.include_router(queries_router)
     app.include_router(account_router)
     register_exception_handlers(app)
+    instrument_app(app)
     return app
 
 
diff --git a/app/scraping/services/site_session.py b/app/scraping/services/site_session.py
index 21d643e..44c5994 100644
--- a/app/scraping/services/site_session.py
+++ b/app/scraping/services/site_session.py
@@ -8,7 +8,13 @@
 画像（详情页只有手机 UA 才返回带 __INITIAL_STATE__ 的统一模板）。两条通道
 各自持有独立 cookie 罐，避免把 PC 指纹拿到的 cookie 混用到手机请求上。
 
-抓取失败时的升级路径：重新预热 → 浏览器兜底取 cookie → 放弃。
+**不预热首页**：Akamai 的 cookie 是随任意一个页面响应下发的，目标页自己就会带回
+来，专门先打一次 `www.rakuten.co.jp/` 除了多一个出站请求（以及多一次被风控计数
+的机会）之外没有额外收益——首个请求无论打哪个 URL 都是冷的 ~11s，之后都复用
+cookie。首页只在**失败修复**路径上使用：目标页已经被挑战时，拿首页换一套干净
+cookie 比继续拿目标页去撞更安全（见 `_rewarm_on_home`）。
+
+抓取失败时的升级路径：换 cookie（首页重新预热）→ 浏览器兜底取 cookie → 放弃。
 """
 from __future__ import annotations
 
@@ -62,12 +68,18 @@ class _Profile:
     mobile: bool
     client: httpx.AsyncClient
     lock: asyncio.Lock = field(default_factory=asyncio.Lock)
-    warmed_at: float = 0.0
+    # Akamai cookie 罐的建立时刻（monotonic）。0 表示当前没有可复用的 cookie。
+    # 超过 session_ttl_seconds 就主动清空：拿着过期 cookie 去撞反而更容易被挑战。
+    cookies_at: float = 0.0
 
     @property
     def cookie_names(self) -> set[str]:
         return {cookie.name for cookie in self.client.cookies.jar}
 
+    @property
+    def akamai_cookies(self) -> set[str]:
+        return self.cookie_names & set(site.AKAMAI_COOKIE_NAMES)
+
 
 class SiteSession:
     """乐天站点抓取会话，管理 cookie 预热、并发限流与失败升级"""
@@ -115,13 +127,17 @@ class SiteSession:
     # ---- 状态 ----
 
     def profile_status(self) -> dict[str, dict[str, Any]]:
-        """各通道的预热状态，供健康检查展示"""
+        """各通道的 cookie 状态，供健康检查展示
+
+        `warmed` 保留原字段名（上游健康检查看板在用），语义是「当前有可复用的
+        Akamai cookie」——不再代表「已专门预热过首页」，因为正常路径不打首页了。
+        """
         now = time.monotonic()
         return {
             name: {
-                "warmed": profile.warmed_at > 0,
-                "age_seconds": round(now - profile.warmed_at, 1) if profile.warmed_at else None,
-                "cookies": sorted(profile.cookie_names & set(site.AKAMAI_COOKIE_NAMES)),
+                "warmed": profile.cookies_at > 0,
+                "age_seconds": round(now - profile.cookies_at, 1) if profile.cookies_at else None,
+                "cookies": sorted(profile.akamai_cookies),
             }
             for name, profile in self._profiles.items()
         }
@@ -142,7 +158,8 @@ class SiteSession:
     ) -> FetchedPage:
         """抓取页面，返回 HTML 与最终落地地址
 
-        失败时按 重新预热 → 浏览器兜底 的顺序逐级升级重试。
+        直接打目标页（不先访问首页），失败时按 换 cookie → 浏览器兜底 的顺序
+        逐级升级重试。
 
         Args:
             validator: 页面校验器，默认要求页面含 __INITIAL_STATE__。跨站抓取时
@@ -181,7 +198,8 @@ class SiteSession:
             try:
                 for attempt in range(1, max_attempts + 1):
                     span.set_attribute("scrape.attempts", attempt)
-                    await self._ensure_warm(profile)
+                    # 过期 cookie 主动丢掉：带着它去撞比裸请求更容易吃挑战页
+                    await self._drop_expired_cookies(profile)
                     try:
                         response = await profile.client.get(url)
                     except httpx.HTTPError as exc:
@@ -192,6 +210,10 @@ class SiteSession:
                         )
                         continue
 
+                    # 任何响应都可能带 set-cookie（Akamai 不保证每次下发），拿到就
+                    # 记下时刻，后续请求复用到 TTL 为止。
+                    self._note_cookies(profile)
+
                     if response.status_code == 404:
                         err = ItemNotFoundError(f"Page not found: {url}")
                         span.record_exception(err)
@@ -223,9 +245,11 @@ class SiteSession:
                     if attempt >= max_attempts:
                         break
 
-                    # 第一次失败先便宜地换一套 cookie；仍失败才动用浏览器
+                    # 第一次失败先便宜地换一套 cookie（丢掉旧的，用首页换新的）；
+                    # 仍失败才动用浏览器
                     if attempt == 1:
-                        await self._invalidate(profile)
+                        await self._rewarm_on_home(profile)
+                        span.set_attribute("scrape.rewarmed_on_home", True)
                     else:
                         page = await self._escalate_to_browser(profile, url, validate)
                         if page is not None:
@@ -265,44 +289,53 @@ class SiteSession:
     def _is_server_error(reason: str) -> bool:
         return reason.startswith("upstream status") or reason.startswith("httpx") or "Error:" in reason
 
-    async def _ensure_warm(self, profile: _Profile) -> None:
-        """确保通道有一次新鲜的首页预热；过期时重新访问首页"""
-        if self._is_warm(profile):
+    def _note_cookies(self, profile: _Profile) -> None:
+        """目标页响应带回 Akamai cookie 时记下时刻，作为 TTL 起点
+
+        已经在计时的不重置：TTL 要从「这套 cookie 第一次出现」算起，每次响应都
+        刷新会让一套 cookie 被无限续命，反而绕过了 session_ttl_seconds 的本意。
+        """
+        if profile.cookies_at:
             return
+        if profile.akamai_cookies:
+            profile.cookies_at = time.monotonic()
 
-        async with profile.lock:
-            if self._is_warm(profile):
-                return
-            try:
-                response = await profile.client.get(self._settings.home_url)
-                # Akamai 不保证每次都下发 cookie；首页探测成功本身就是可复用的
-                # 预热结果，cookie 只用于观测和失败升级时的回灌。
-                if response.status_code < 400:
-                    profile.warmed_at = time.monotonic()
-                logger.info(
-                    "会话预热完成：profile=%s status=%s cookies=%s",
-                    profile.name,
-                    response.status_code,
-                    sorted(profile.cookie_names & set(site.AKAMAI_COOKIE_NAMES)),
-                )
-            except httpx.HTTPError as exc:
-                # 预热失败不阻断本次抓取：直连目标页仍可能成功，只是慢
-                logger.warning("会话预热失败：profile=%s err=%s", profile.name, exc)
-                profile.warmed_at = 0.0
-
-    def _is_warm(self, profile: _Profile) -> bool:
-        if not profile.warmed_at:
-            return False
-        if time.monotonic() - profile.warmed_at > self._settings.session_ttl_seconds:
-            return False
-        return True
-
-    async def _invalidate(self, profile: _Profile) -> None:
-        """清空通道 cookie 并强制下次重新预热"""
+    async def _drop_expired_cookies(self, profile: _Profile) -> None:
+        """cookie 罐超过 TTL 时清空，让下一次请求裸奔换一套新的"""
+        if not profile.cookies_at:
+            return
+        if time.monotonic() - profile.cookies_at <= self._settings.session_ttl_seconds:
+            return
         async with profile.lock:
             profile.client.cookies.clear()
-            profile.warmed_at = 0.0
-        logger.info("已清空会话 cookie，将重新预热：profile=%s", profile.name)
+            profile.cookies_at = 0.0
+        logger.info("会话 cookie 已过期，已清空：profile=%s", profile.name)
+
+    async def _rewarm_on_home(self, profile: _Profile) -> None:
+        """首次失败后的修复：丢掉旧 cookie，用首页换一套新的
+
+        这是首页 URL 唯一的用途。目标页已经吃了挑战页，继续拿同一个 URL 去撞
+        只会把挑战坐实；首页是站点最"无害"的入口，换 cookie 的成功率更高。
+
+        失败不阻断本次抓取（下一次 attempt 会裸请求目标页，只是慢），所以这里
+        只记日志。
+        """
+        async with profile.lock:
+            profile.client.cookies.clear()
+            profile.cookies_at = 0.0
+            try:
+                response = await profile.client.get(self._settings.home_url)
+            except httpx.HTTPError as exc:
+                logger.warning("首页换 cookie 失败：profile=%s err=%s", profile.name, exc)
+                return
+            if profile.akamai_cookies:
+                profile.cookies_at = time.monotonic()
+        logger.info(
+            "已用首页换一套新 cookie：profile=%s status=%s cookies=%s",
+            profile.name,
+            response.status_code,
+            sorted(profile.akamai_cookies),
+        )
 
     async def _escalate_to_browser(
         self, profile: _Profile, url: str, validate: PageValidator
@@ -332,7 +365,7 @@ class SiteSession:
                     domain=cookie.get("domain") or "",
                     path=cookie.get("path") or "/",
                 )
-            profile.warmed_at = time.monotonic()
+            profile.cookies_at = time.monotonic()
 
         # 浏览器不回报最终 URL，这里以请求地址为准；跨站跳转场景下 HTTP 通道已先行
         # 报错，走不到这一步。
diff --git a/app/shared/config.py b/app/shared/config.py
index 268ad78..dd9f9db 100644
--- a/app/shared/config.py
+++ b/app/shared/config.py
@@ -73,7 +73,9 @@ class Settings(BaseSettings):
     request_timeout_seconds: float = 30.0
     max_site_concurrency: int = 8  # 对站点的最大并发请求数
     http_max_attempts: int = 3  # 单次抓取的最大尝试次数（含首次）
-    session_ttl_seconds: float = 1800.0  # Akamai cookie 会话最长复用时长，超时后重新预热
+    # Akamai cookie 最长复用时长；超时后清空 cookie 罐，由下一次响应重新建立。
+    # cookie 随目标页响应下发，正常路径不额外访问站点首页（见 site_session.py）。
+    session_ttl_seconds: float = 1800.0
 
     # ---- 浏览器兜底配置（仅抓取服务）----
     # 纯 HTTP 被 Akamai 拦截时，用 Playwright 打开页面取回 cookie 再回灌给
diff --git a/app/shared/telemetry.py b/app/shared/telemetry.py
index adbf54b..357a268 100644
--- a/app/shared/telemetry.py
+++ b/app/shared/telemetry.py
@@ -11,28 +11,45 @@ instrumentation（FastAPI、httpx）。失败时（如 endpoint 不可达）不
 OTel SDK 默认的 ProxyTracerProvider 在 setup 之前就能用（noop span），所以
 其它代码里直接 `trace.get_tracer(__name__)` + `start_as_current_span` 即可，
 不必关心 telemetry 是否启用——禁用时 span 不会真正产生与上报。
+
+自动 instrumentation（FastAPI + httpx）只覆盖「进程收到 HTTP 请求」与「进程发出
+httpx 请求」两类边界。交易侧的实际工作两者都不是：站点交互走 Playwright（不经
+httpx），worker 主循环是后台 asyncio 任务（没有 HTTP 入口）。所以那一侧必须手工
+埋点，否则 trace 里只剩 worker 与网关之间的往返记录，看不到任何业务链路。本模块
+为此提供三件东西：
+
+- `traced`：给 async 方法套一层 span，异常自动记录（站点交互各步骤在用）
+- `set_attributes` / `record_error`：批量写属性、统一记异常
+- `suppressed`：屏蔽空转长轮询产生的孤立 trace
 """
 from __future__ import annotations
 
+import functools
 import logging
-from typing import TYPE_CHECKING
+from collections.abc import Awaitable, Callable, Iterator, Mapping
+from contextlib import contextmanager
+from typing import TYPE_CHECKING, ParamSpec, TypeVar
 
 from opentelemetry import trace
 from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
 from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor
 from opentelemetry.instrumentation.httpx import HTTPXClientInstrumentor
+from opentelemetry.instrumentation.utils import suppress_instrumentation
 from opentelemetry.sdk.resources import SERVICE_NAME, Resource
 from opentelemetry.sdk.trace import TracerProvider
 from opentelemetry.sdk.trace.export import BatchSpanProcessor
 from opentelemetry.sdk.trace.sampling import ALWAYS_ON
-from opentelemetry.trace import Span
+from opentelemetry.trace import Span, SpanKind, Status, StatusCode
 from opentelemetry.util.types import AttributeValue
 
-from app.shared.config import Settings
+from app.shared.config import Settings, get_settings
 
 if TYPE_CHECKING:
     from fastapi import FastAPI
 
+P = ParamSpec("P")
+R = TypeVar("R")
+
 logger = logging.getLogger(__name__)
 
 # 全局 provider 引用，用于 instrument_app / shutdown 时判断当前是否已初始化。
@@ -88,8 +105,24 @@ def setup_telemetry(settings: Settings, *, service_name: str) -> None:
 
 
 def instrument_app(app: "FastAPI") -> None:
-    """FastAPI 应用打桩；未初始化时 noop，调用顺序无要求。"""
-    if _provider is None:
+    """FastAPI 应用打桩。必须在应用开始服务之前调用，与 setup_telemetry 的先后无关。
+
+    **不能用 `_provider is None` 做前置判断**：三个服务都在模块导入时执行
+    `app = create_app()`，而 `setup_telemetry` 要等 lifespan 启动才跑，那时
+    `_provider` 还是 None——照着判断就会直接 return，FastAPI 永远没被打桩，
+    观测后台里一条 server span 都不会有。
+
+    反过来「等 lifespan 里再打桩」也不行：instrument_app 是往应用上加中间件，
+    应用一旦开始服务，加进去的中间件不生效（实测 lifespan 内调用后 server span
+    为空）。所以只能在这里、在导入期就装上。
+
+    provider 尚未设置时拿到的是 ProxyTracer，它在 `set_tracer_provider` 之后会
+    自动委托到真实 provider（实测：导入期打桩 + lifespan 内设 provider，请求
+    照样产生 span），所以顺序不构成问题。
+
+    otel 关闭时跳过：省掉一层用不上的中间件。
+    """
+    if not get_settings().otel_enabled:
         return
     FastAPIInstrumentor.instrument_app(app)
 
@@ -127,6 +160,74 @@ def snapshot(span: Span, name: str, html: str | None, max_bytes: int) -> None:
     span.add_event(name, attributes=attributes)
 
 
+@contextmanager
+def suppressed() -> Iterator[None]:
+    """在这个上下文里不产生任何自动 instrumentation span。
+
+    给「空转的长轮询」用：worker 每 30 秒问一次网关有没有活干，绝大多数时候
+    返回空。这些请求各自成为一条孤立 trace，量大且没有信息量——把观测后台刷满
+    的正是它们。领到任务后的每一次网关调用都在任务根 span 底下，不受影响。
+    """
+    with suppress_instrumentation():
+        yield
+
+
+def record_error(span: Span, exc: BaseException) -> None:
+    """把异常记到 span 上并置 ERROR 状态。
+
+    单独抽出来是因为 AppError 带 `err_code`（对外错误码），排查时按码筛比按
+    异常类名筛更贴近上游看到的东西，值得单独落一个属性。
+    """
+    span.record_exception(exc)
+    span.set_attribute("error.type", type(exc).__name__)
+    err_code = getattr(exc, "err_code", None)
+    if isinstance(err_code, int):
+        span.set_attribute("error.code", err_code)
+    span.set_status(Status(StatusCode.ERROR, f"{type(exc).__name__}: {exc}"))
+
+
+def traced(
+    name: str,
+    *,
+    kind: SpanKind = SpanKind.INTERNAL,
+) -> Callable[[Callable[P, Awaitable[R]]], Callable[P, Awaitable[R]]]:
+    """给 async 方法套一层 span，异常自动记录后原样抛出。
+
+    交易侧的实际工作是 Playwright 页面操作，httpx 自动 instrumentation 完全看不到
+    （浏览器请求不走 httpx），所以这些步骤必须手工埋点，否则 trace 里只剩 worker
+    与网关之间的 HTTP 往返。用装饰器而不是在每个方法里写 with 块，是因为这些方法
+    的函数体都已经很长，再加一层缩进不利于阅读。
+
+    未启用 telemetry 时 tracer 是 noop，装饰器只多一次函数调用，可以无条件套。
+    """
+
+    def decorate(fn: Callable[P, Awaitable[R]]) -> Callable[P, Awaitable[R]]:
+        @functools.wraps(fn)
+        async def wrapper(*args: P.args, **kwargs: P.kwargs) -> R:
+            tracer = trace.get_tracer(fn.__module__)
+            with tracer.start_as_current_span(name, kind=kind) as span:
+                try:
+                    return await fn(*args, **kwargs)
+                except Exception as exc:
+                    record_error(span, exc)
+                    raise
+
+        return wrapper
+
+    return decorate
+
+
+def set_attributes(span: Span, attributes: Mapping[str, AttributeValue | None]) -> None:
+    """批量设置属性，跳过 None 值。
+
+    站点交互里大量字段是可选的（site_order_id 要到提交后才有、payable_yen 只在
+    确认页解析后才有），逐个 if 判断会把埋点代码写得比业务逻辑还长。
+    """
+    for key, value in attributes.items():
+        if value is not None:
+            span.set_attribute(key, value)
+
+
 def _parse_headers(raw: str | None) -> list[tuple[str, str]] | None:
     """解析 "k1=v1,k2=v2" 形式的 header 配置；空输入返回 None。"""
     if not raw or not raw.strip():
diff --git a/app/trading/worker/query_runner.py b/app/trading/worker/query_runner.py
index 8a989be..a3532e2 100644
--- a/app/trading/worker/query_runner.py
+++ b/app/trading/worker/query_runner.py
@@ -21,8 +21,12 @@ import logging
 from datetime import datetime, timezone
 from typing import TYPE_CHECKING, Any
 
+from opentelemetry import trace
+from opentelemetry.trace import SpanKind
+
 from app.shared.errors import AppError, InvalidRequestError
 from app.shared.task_state import AccountQueryKind
+from app.shared.telemetry import record_error, set_attributes, suppressed
 from app.trading.worker.client import GatewayClient
 from app.trading.worker.models import QueryTask
 from app.trading.worker.site_interact import SiteInteractor
@@ -32,6 +36,8 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+tracer = trace.get_tracer(__name__)
+
 # 没给 since 时的窗口下界。用 epoch 而不是 None，是为了让翻页终止条件
 # （`order_date < since`）与「不设下界」共用同一条代码路径，不多一个分支。
 _EPOCH = datetime(1970, 1, 1, tzinfo=timezone.utc)
@@ -100,7 +106,9 @@ class QueryRunner:
         logger.info("账号查询 worker 启动：worker_id=%s", self.worker_id)
         while self._running:
             try:
-                query = await self._gateway.lease_query(self.worker_id, wait=30)
+                # 与下单主循环同样：空转的长轮询不埋点，见 runner.py::run
+                with suppressed():
+                    query = await self._gateway.lease_query(self.worker_id, wait=30)
             except AppError as exc:
                 logger.warning("查询 lease 失败：%s (err=%s)", exc.message, exc.err_code)
                 await asyncio.sleep(5)
@@ -122,7 +130,27 @@ class QueryRunner:
     # ---- 单张查询单 ----
 
     async def handle(self, query: QueryTask) -> None:
-        """执行一张查询单并回报结果。任何失败都转成一次「失败回报」，不抛出去"""
+        """执行一张查询单并回报结果。任何失败都转成一次「失败回报」，不抛出去
+
+        这里开的 span 是一张查询单的根：底下挂着站点读取（`site.*`）与回报网关的
+        HTTP 调用，一个 query_id 对应一条 trace。查询失败不抛出去（都转成失败
+        回报），所以失败信息由各分支显式记到 span 上——否则 trace 里会显示成功。
+        """
+        with tracer.start_as_current_span("account_query", kind=SpanKind.CONSUMER) as span:
+            set_attributes(
+                span,
+                {
+                    "query.query_id": query.query_id,
+                    "query.kind": query.kind,
+                    "query.site": query.site,
+                    "query.attempt": query.attempt,
+                    "query.worker_id": self.worker_id,
+                },
+            )
+            await self._handle_traced(query, span)
+
+    async def _handle_traced(self, query: QueryTask, span: "trace.Span") -> None:
+        """handle() 的实际执行体，拆出来只为让根 span 的 with 块保持一层缩进"""
         logger.info(
             "领到查询单：query_id=%s kind=%s attempt=%s",
             query.query_id, query.kind, query.attempt,
@@ -132,11 +160,13 @@ class QueryRunner:
                 self.execute(query),
                 timeout=self._settings.account_query_timeout_seconds,
             )
-        except asyncio.TimeoutError:
+        except asyncio.TimeoutError as exc:
             logger.warning(
                 "查询超时（%s 秒，多半是下单任务正占着账号锁）：query_id=%s",
                 self._settings.account_query_timeout_seconds, query.query_id,
             )
+            span.set_attribute("query.outcome", "timeout")
+            record_error(span, exc)
             await self._report_safe(
                 query,
                 success=False,
@@ -152,12 +182,16 @@ class QueryRunner:
                 "查询失败：query_id=%s code=%s msg=%s",
                 query.query_id, exc.err_code, exc.message,
             )
+            span.set_attribute("query.outcome", "failed")
+            record_error(span, exc)
             await self._report_safe(
                 query, success=False, error_code=exc.err_code, error_message=exc.message
             )
             return
         except Exception as exc:  # noqa: BLE001
             logger.exception("查询未预期异常：query_id=%s", query.query_id)
+            span.set_attribute("query.outcome", "unexpected_error")
+            record_error(span, exc)
             await self._report_safe(
                 query,
                 success=False,
@@ -168,10 +202,12 @@ class QueryRunner:
 
         payload, oversized = self._enforce_result_size(result)
         if oversized is not None:
+            span.set_attribute("query.outcome", "oversized")
             await self._report_safe(
                 query, success=False, error_code=1003, error_message=oversized
             )
             return
+        span.set_attribute("query.outcome", "succeeded")
         await self._report_safe(query, success=True, result=payload)
 
     async def execute(self, query: QueryTask) -> dict[str, Any]:
diff --git a/app/trading/worker/runner.py b/app/trading/worker/runner.py
index 9843e4b..f34b2c8 100644
--- a/app/trading/worker/runner.py
+++ b/app/trading/worker/runner.py
@@ -28,6 +28,9 @@ import contextlib
 import logging
 from typing import TYPE_CHECKING
 
+from opentelemetry import trace
+from opentelemetry.trace import SpanKind, Status, StatusCode
+
 from app.shared.errors import (
     AppError,
     BrowserDeadError,
@@ -35,6 +38,7 @@ from app.shared.errors import (
     OrderGuardError,
 )
 from app.shared.task_state import OrderState, TaskStatus
+from app.shared.telemetry import record_error, set_attributes, suppressed
 from app.trading.worker import verify
 from app.trading.worker.client import GatewayClient
 from app.trading.worker.evidence import EvidenceStore
@@ -47,6 +51,8 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+tracer = trace.get_tracer(__name__)
+
 
 def _coerce_state(value: str | None) -> OrderState:
     """把网关返回的 state 字符串安全地包成 OrderState；None 或非法值回落到 CREATED"""
@@ -115,7 +121,11 @@ class WorkerRunner:
         logger.info("worker 启动：worker_id=%s", self.worker_id)
         while self._running:
             try:
-                task = await self._gateway.lease(self.worker_id, wait=30)
+                # 空转的长轮询不埋点：30 秒一次、绝大多数返回空，每次都会变成一条
+                # 孤立 trace 把观测后台刷满。领到任务后的调用都在 handle() 的根
+                # span 底下，不受这里影响。
+                with suppressed():
+                    task = await self._gateway.lease(self.worker_id, wait=30)
             except AppError as exc:
                 logger.warning("lease 失败：%s (err=%s)", exc.message, exc.err_code)
                 await asyncio.sleep(5)
@@ -137,13 +147,42 @@ class WorkerRunner:
     # ---- 单任务调度 ----
 
     async def handle(self, task: LeaseTask) -> None:
-        """单任务调度入口：本地幂等闸门 → 恢复核对 → 执行"""
+        """单任务调度入口：本地幂等闸门 → 恢复核对 → 执行
+
+        这里开的 span 是**整条下单链路的根**：往下的每一步站点交互、每一次回报
+        网关都挂在它底下，一个 task_id 对应观测后台里的一条 trace。worker 是后台
+        asyncio 任务，没有 HTTP 入口，不开这个根 span 的话下游 httpx 调用会各自
+        散成孤立 trace（这正是「只有请求记录、没有链路」的原因）。
+        """
+        with tracer.start_as_current_span("order.task", kind=SpanKind.CONSUMER) as span:
+            intent = task.intent or {}
+            set_attributes(
+                span,
+                {
+                    "order.task_id": task.task_id,
+                    "order.site": task.site,
+                    "order.worker_id": self.worker_id,
+                    "order.lease_count": task.lease_count,
+                    "order.known_state": task.known_state,
+                    "order.item_url": intent.get("item_url"),
+                    "order.quantity": intent.get("quantity"),
+                },
+            )
+            try:
+                await self._dispatch(task, span)
+            except Exception as exc:
+                record_error(span, exc)
+                raise
+
+    async def _dispatch(self, task: LeaseTask, span: "trace.Span") -> None:
+        """handle() 的实际分支逻辑，拆出来只为让根 span 的 with 块保持一层缩进"""
         # 本地幂等闸门：之前已完成过的任务不再执行
         if await self._db.has_finished(task.task_id):
             final_state = await self._db.final_state(task.task_id)
             logger.info(
                 "本地已完成，补报终态：task_id=%s state=%s", task.task_id, final_state
             )
+            span.set_attribute("order.route", "already_finished")
             await self._report_safe(
                 task,
                 state=_coerce_state(final_state),
@@ -154,10 +193,12 @@ class WorkerRunner:
 
         # 恢复领取：lease_count > 1 表示 stale → reclaim，必须先核对站点订单
         if task.lease_count > 1:
+            span.set_attribute("order.route", "recovery")
             await self._handle_recovery(task)
             return
 
         # 常规执行
+        span.set_attribute("order.route", "execute")
         await self._execute_with_renewal(task)
 
     async def _handle_recovery(self, task: LeaseTask) -> None:
@@ -197,11 +238,24 @@ class WorkerRunner:
 
     # ---- 常规执行 ----
 
+    async def _execute_recording_errors(self, task: LeaseTask) -> None:
+        """execute() 外面加一层：异常先记到任务根 span 上，再原样抛出
+
+        下面那一串 except 分支会把异常**吞掉**转成 needs_human / failed 回报，
+        异常冒不到 handle() 的根 span。在这里统一记一次，比每个分支各写一遍省事，
+        也不会漏掉新增的分支。
+        """
+        try:
+            await self.execute(task)
+        except Exception as exc:
+            record_error(trace.get_current_span(), exc)
+            raise
+
     async def _execute_with_renewal(self, task: LeaseTask) -> None:
         """在租约自动续期的上下文里执行任务"""
         async with self._renew_lease_every(task, interval=60):
             try:
-                await self.execute(task)
+                await self._execute_recording_errors(task)
             except NotImplementedError as exc:
                 # 站点交互未实现（规格 §10）：上报 needs_human，不视为 worker 失败
                 logger.warning(
@@ -521,38 +575,65 @@ class WorkerRunner:
         证据载体）时，其 html / screenshot 即本步骤要落盘的页面与整页截图；
         显式传入的 `html` / `png` 优先级更高，供 step3/step4 等已经单独拿到
         页面的调用点使用。
+
+        每步一个 span，挂在 handle() 的任务根 span 底下：一条 trace 就是一单的
+        完整流水（清车 → 加购 → 校验 → 确认 → 提交 → 付款），卡在哪一步、每步
+        耗时多少、证据落在哪个 evidence_ref 都能直接读出来。
         """
-        result = await action()
-        if isinstance(result, PageSnapshot):
-            if html is None:
-                html = result.html or None
-            if png is None:
-                png = result.screenshot or None
-        elif html is None and isinstance(result, str):
-            # 旧契约兼容：站点方法返回字符串时视为页面 HTML
-            html = result
-        meta = {
-            "step": step_name,
-            "state": state.value,
-            **(evidence_meta or {}),
-        }
-        evidence_ref = self._evidence.write_step(
-            task.task_id, step_no, step_name, html=html, png=png, meta=meta
-        )
-        await self._db.index_evidence(task.task_id, step_no, step_name, evidence_ref)
-        await self._db.record_event(
-            task.task_id, state.value, detail=detail, evidence_ref=evidence_ref
-        )
-        await self._gateway.report(
-            task.task_id,
-            self.worker_id,
-            state=state,
-            payable_yen=payable_yen,
-            pay_deadline=pay_deadline,
-            site_order_id=site_order_id,
-            evidence_ref=evidence_ref,
-            detail=detail,
-        )
+        with tracer.start_as_current_span(f"order.step.{step_name}") as span:
+            set_attributes(
+                span,
+                {
+                    "order.task_id": task.task_id,
+                    "order.step_no": step_no,
+                    "order.step_name": step_name,
+                    "order.state": state.value,
+                },
+            )
+            try:
+                result = await action()
+            except Exception as exc:
+                record_error(span, exc)
+                raise
+
+            if isinstance(result, PageSnapshot):
+                if html is None:
+                    html = result.html or None
+                if png is None:
+                    png = result.screenshot or None
+            elif html is None and isinstance(result, str):
+                # 旧契约兼容：站点方法返回字符串时视为页面 HTML
+                html = result
+            meta = {
+                "step": step_name,
+                "state": state.value,
+                **(evidence_meta or {}),
+            }
+            evidence_ref = self._evidence.write_step(
+                task.task_id, step_no, step_name, html=html, png=png, meta=meta
+            )
+            set_attributes(
+                span,
+                {
+                    "order.evidence_ref": evidence_ref,
+                    "order.site_order_id": site_order_id,
+                    "order.payable_yen": payable_yen,
+                },
+            )
+            await self._db.index_evidence(task.task_id, step_no, step_name, evidence_ref)
+            await self._db.record_event(
+                task.task_id, state.value, detail=detail, evidence_ref=evidence_ref
+            )
+            await self._gateway.report(
+                task.task_id,
+                self.worker_id,
+                state=state,
+                payable_yen=payable_yen,
+                pay_deadline=pay_deadline,
+                site_order_id=site_order_id,
+                evidence_ref=evidence_ref,
+                detail=detail,
+            )
 
     async def _report_safe(
         self,
@@ -566,7 +647,25 @@ class WorkerRunner:
         payable_yen: int | None = None,
         pay_deadline: str | None = None,
     ) -> None:
-        """回报 gateway，失败只记日志不抛——主循环不能因为回报失败退出"""
+        """回报 gateway，失败只记日志不抛——主循环不能因为回报失败退出
+
+        顺带把终态标到当前 span 上。这一步是必要的：`_execute_with_renewal` 会把
+        闸门拦截、风控拦截、浏览器掉线等异常**吞掉**转成 needs_human 回报，异常
+        不会冒到 handle() 的根 span，于是一笔被拦下的单在 trace 里看起来跟成功
+        下单一模一样。所有这些分支都汇到这个方法，标在这里最省事也最不容易漏。
+
+        `is_recording()` 那道判断不是多余的：付款后监控（_monitor_order）是
+        `create_task` 起的后台任务，而 asyncio 在创建时就把当时的 context 复制了
+        进去——等它真正跑起来，任务根 span 早已结束，但 `get_current_span()` 拿到
+        的仍是那个**已结束**的 span（不是 INVALID_SPAN）。往上写属性会打
+        "Setting attribute on ended span" 警告。当前监控路径不传 terminal_status
+        走不到这里，但这道判断保证以后传了也不会污染已完成的任务 span。
+        """
+        span = trace.get_current_span()
+        if terminal_status is not None and span.is_recording():
+            span.set_attribute("order.terminal_status", terminal_status.value)
+            if terminal_status in (TaskStatus.NEEDS_HUMAN, TaskStatus.FAILED):
+                span.set_status(Status(StatusCode.ERROR, detail))
         try:
             await self._gateway.report(
                 task.task_id,
diff --git a/app/trading/worker/site_interact.py b/app/trading/worker/site_interact.py
index 187d27e..cbd11ea 100644
--- a/app/trading/worker/site_interact.py
+++ b/app/trading/worker/site_interact.py
@@ -96,6 +96,8 @@ from dataclasses import dataclass, field
 from datetime import datetime
 from typing import TYPE_CHECKING, TypeVar
 
+from opentelemetry.trace import SpanKind
+
 from app.shared.errors import (
     BrowserDeadError,
     CartOperationError,
@@ -113,6 +115,8 @@ from app.shared.purchase_contract import (
 )
 from app.shared.proxy import playwright_launch_proxy
 from app.shared.task_state import OrderState
+from app.shared.telemetry import traced
+
 from app.trading.core import auth_site
 from app.trading.worker.models import LeaseTask
 
@@ -973,6 +977,7 @@ class SiteInteractor:
 
     # ---- 已实现：add_to_cart / verify_cart / cart_status ----
 
+    @traced("site.add_to_cart", kind=SpanKind.CLIENT)
     async def add_to_cart(self, task: LeaseTask) -> PageSnapshot:
         """加购（worker 入口）：从 task.intent 取字段，调 _add_to_cart_with_fields
 
@@ -1015,6 +1020,7 @@ class SiteInteractor:
             screenshot=result.get("screenshot") or b"",
         )
 
+    @traced("site.add_to_cart_payload", kind=SpanKind.CLIENT)
     async def add_to_cart_payload(
         self,
         *,
@@ -1163,6 +1169,7 @@ class SiteInteractor:
             "screenshot": screenshot,
         }
 
+    @traced("site.verify_cart", kind=SpanKind.CLIENT)
     async def verify_cart(self, task: LeaseTask) -> PageSnapshot:
         """校验购物车里有没有刚加的商品
 
@@ -1196,6 +1203,7 @@ class SiteInteractor:
             # 2. 渲染 cart 页确认 item_id 在里面
             return await self._verify_item_in_cart_html(item_id, label=f"task_id={task.task_id}")
 
+    @traced("site.cart_status", kind=SpanKind.CLIENT)
     async def cart_status(self) -> dict:
         """轻量查询购物车状态：调 cart count JSONP API，不渲染整页
 
@@ -1216,6 +1224,7 @@ class SiteInteractor:
 
     # ---- 已实现：clear_cart / remove_item（Playwright UI 点击）----
 
+    @traced("site.clear_cart", kind=SpanKind.CLIENT)
     async def clear_cart(self) -> dict:
         """清空购物车：渲染 cart SPA → 反复点第一个「削除」按钮 → count API 校验
 
@@ -1304,6 +1313,7 @@ class SiteInteractor:
                 "screenshot": screenshot,
             }
 
+    @traced("site.remove_item", kind=SpanKind.CLIENT)
     async def remove_item(self, item_id: str) -> dict:
         """删除购物车里指定 item_id 的商品
 
@@ -1506,6 +1516,7 @@ class SiteInteractor:
 
     # ---- 已实现：enter_checkout（到下单确认页，中间步骤未经真实 HTML 验证）----
 
+    @traced("site.enter_checkout", kind=SpanKind.CLIENT)
     async def enter_checkout(self, task: LeaseTask) -> PageSnapshot:
         """进入下单确认页：购物车 → 点「購入手続き」→ 依次处理中间步骤 → 落地确认页
 
@@ -2017,6 +2028,7 @@ class SiteInteractor:
         """
         return _parse_checkout_summary(html)
 
+    @traced("site.submit_order", kind=SpanKind.CLIENT)
     async def submit_order(self, task: LeaseTask) -> SubmitOutcome:
         """点击下单确认页的最终确认按钮，提交订单
 
@@ -2118,6 +2130,7 @@ class SiteInteractor:
                 evidence=PageSnapshot(html=html, screenshot=screenshot),
             )
 
+    @traced("site.pay", kind=SpanKind.CLIENT)
     async def pay(self, task: LeaseTask, site_order_id: str) -> PageSnapshot:
         """检查提交下单后是否已完成付款 / 是否触发了需要人工介入的验证环节
 
@@ -2181,6 +2194,7 @@ class SiteInteractor:
                 except Exception:
                     logger.debug("pay 关闭确认页失败", exc_info=True)
 
+    @traced("site.check_order_status", kind=SpanKind.CLIENT)
     async def check_order_status(self, site_order_id: str) -> OrderStatusSnapshot:
         """付款后监控的单次探测：查一次订单详情页的配送阶段，不循环
 
@@ -2198,6 +2212,7 @@ class SiteInteractor:
         """
         return (await self.fetch_order_detail(site_order_id)).status
 
+    @traced("site.fetch_order_detail", kind=SpanKind.CLIENT)
     async def fetch_order_detail(self, site_order_id: str) -> OrderDetailSnapshot:
         """读一次订单详情页：配送阶段 + 页面原始 __INITIAL_STATE__
 
@@ -2269,6 +2284,7 @@ class SiteInteractor:
             f"check_order_status site_order_id={site_order_id}", read
         )
 
+    @traced("site.list_recent_orders", kind=SpanKind.CLIENT)
     async def list_recent_orders(
         self, *, since: datetime, max_pages: int | None = None
     ) -> OrderListWindow:
diff --git a/tests/test_site_session.py b/tests/test_site_session.py
index d650fdc..b6feb2f 100644
--- a/tests/test_site_session.py
+++ b/tests/test_site_session.py
@@ -67,15 +67,17 @@ async def build_session(handler, *, browser=None, settings=None) -> SiteSession:
     return session
 
 
-async def test_warmup_happens_before_first_fetch_and_is_reused():
+async def test_happy_path_never_touches_the_home_page():
+    """目标页自己会带回 Akamai cookie，正常路径不该多打一次首页。"""
     seen: list[str] = []
 
     def handler(request: httpx.Request) -> httpx.Response:
         seen.append(str(request.url))
-        headers = {}
-        if request.url.host == "www.rakuten.co.jp":
-            headers["set-cookie"] = "ak_bmsc=abc; Domain=.rakuten.co.jp; Path=/"
-        return httpx.Response(200, text=GOOD_PAGE, headers=headers)
+        return httpx.Response(
+            200,
+            text=GOOD_PAGE,
+            headers={"set-cookie": "ak_bmsc=abc; Domain=.rakuten.co.jp; Path=/"},
+        )
 
     session = await build_session(handler)
     try:
@@ -84,18 +86,20 @@ async def test_warmup_happens_before_first_fetch_and_is_reused():
     finally:
         await session.close()
 
-    # 首页预热只做一次，第二次抓取直接复用 cookie
-    assert seen.count("https://www.rakuten.co.jp/") == 1
-    assert seen.count(TARGET) == 2
+    assert seen == [TARGET, TARGET]
 
 
-async def test_successful_warmup_is_reused_when_upstream_sets_no_cookie():
-    """首页可能返回 200 但不下发 Akamai cookie，仍不得重复预热。"""
-    seen: list[str] = []
+async def test_cookies_from_target_page_are_reused_across_fetches():
+    """第一次响应下发的 cookie 要带到后续请求上，不必再走首页。"""
+    cookie_headers: list[str | None] = []
 
     def handler(request: httpx.Request) -> httpx.Response:
-        seen.append(str(request.url))
-        return httpx.Response(200, text=GOOD_PAGE)
+        cookie_headers.append(request.headers.get("cookie"))
+        return httpx.Response(
+            200,
+            text=GOOD_PAGE,
+            headers={"set-cookie": "ak_bmsc=abc; Domain=.rakuten.co.jp; Path=/"},
+        )
 
     session = await build_session(handler)
     try:
@@ -104,8 +108,30 @@ async def test_successful_warmup_is_reused_when_upstream_sets_no_cookie():
     finally:
         await session.close()
 
-    assert seen.count("https://www.rakuten.co.jp/") == 1
-    assert seen.count(TARGET) == 2
+    assert cookie_headers[0] is None  # 首个请求是冷的
+    assert "ak_bmsc=abc" in (cookie_headers[1] or "")
+
+
+async def test_expired_cookies_are_dropped_before_next_fetch():
+    """cookie 罐超过 session_ttl_seconds 后要清空，不能带着过期 cookie 去撞。"""
+    cookie_headers: list[str | None] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        cookie_headers.append(request.headers.get("cookie"))
+        return httpx.Response(
+            200,
+            text=GOOD_PAGE,
+            headers={"set-cookie": "ak_bmsc=abc; Domain=.rakuten.co.jp; Path=/"},
+        )
+
+    session = await build_session(handler, settings=make_settings(session_ttl_seconds=0.0))
+    try:
+        await session.fetch_html(TARGET, mobile=False)
+        await session.fetch_html(TARGET, mobile=False)
+    finally:
+        await session.close()
+
+    assert cookie_headers == [None, None]  # 每次都从干净状态起步
 
 
 async def test_missing_state_marker_is_treated_as_blocked():
@@ -139,6 +165,32 @@ async def test_retry_recovers_when_a_later_attempt_succeeds():
     assert attempts["n"] == 2
 
 
+async def test_first_failure_swaps_cookies_via_home_page():
+    """首次失败时用首页换一套 cookie——这是首页 URL 唯一的用途。"""
+    seen: list[str] = []
+
+    def handler(request: httpx.Request) -> httpx.Response:
+        seen.append(str(request.url))
+        if request.url.host == "www.rakuten.co.jp":
+            return httpx.Response(
+                200,
+                text="home",
+                headers={"set-cookie": "ak_bmsc=fresh; Domain=.rakuten.co.jp; Path=/"},
+            )
+        if seen.count(TARGET) == 1:
+            return httpx.Response(200, text=BLOCK_PAGE)
+        return httpx.Response(200, text=GOOD_PAGE)
+
+    session = await build_session(handler)
+    try:
+        assert await session.fetch_html(TARGET, mobile=False) == GOOD_PAGE
+    finally:
+        await session.close()
+
+    # 首页只在失败之后出现一次，且排在两次目标页请求中间
+    assert seen == [TARGET, "https://www.rakuten.co.jp/", TARGET]
+
+
 async def test_browser_fallback_supplies_cookies_and_page_on_persistent_block():
     """浏览器已经取到页面时应直接采用，不再多打一次 HTTP"""
     browser = FakeBrowser(
@@ -264,9 +316,9 @@ async def test_search_and_detail_use_separate_cookie_jars():
 
     def handler(request: httpx.Request) -> httpx.Response:
         mobile = request.headers.get("sec-ch-ua-mobile") == "?1"
-        headers = {}
-        if request.url.host == "www.rakuten.co.jp":
-            headers["set-cookie"] = f"ak_bmsc={'sp' if mobile else 'pc'}; Domain=.rakuten.co.jp; Path=/"
+        headers = {
+            "set-cookie": f"ak_bmsc={'sp' if mobile else 'pc'}; Domain=.rakuten.co.jp; Path=/"
+        }
         return httpx.Response(200, text=GOOD_PAGE, headers=headers)
 
     session = await build_session(handler)
@@ -282,7 +334,9 @@ async def test_search_and_detail_use_separate_cookie_jars():
     assert sp_cookie == "sp"
 
 
-async def test_profile_status_reports_warmup_state():
+async def test_profile_status_reports_cookie_state():
+    """`warmed` 现在的语义是「当前有可复用的 Akamai cookie」，字段名为兼容保留。"""
+
     def handler(request: httpx.Request) -> httpx.Response:
         headers = {"set-cookie": "ak_bmsc=abc; Domain=.rakuten.co.jp; Path=/"}
         return httpx.Response(200, text=GOOD_PAGE, headers=headers)
diff --git a/tests/test_telemetry.py b/tests/test_telemetry.py
index 3cf449d..9574fb0 100644
--- a/tests/test_telemetry.py
+++ b/tests/test_telemetry.py
@@ -1,22 +1,47 @@
 """telemetry 模块测试
 
-验证两件事：
+验证：
 1. 默认配置（otel_enabled=False）下 setup 是 noop，不初始化任何 provider。
 2. enabled=true + endpoint 时 setup 注册真实 TracerProvider；shutdown 复位。
+3. instrument_app 在 setup 之前调用也要真的装上中间件（三个服务都是导入期打桩）。
+4. traced / set_attributes / record_error 的行为。
 
 不打真实网络：OTLPSpanExporter 创建时不发请求，BatchSpanProcessor 异步批量
-上报在没有 span 产生时也不会触发。
+上报在没有 span 产生时也不会触发。span 断言用独立的 InMemory provider，不碰
+全局 provider——OTel 的全局 provider 只允许设置一次，测试间共享会互相污染。
 """
 from __future__ import annotations
 
+import pytest
+from fastapi import FastAPI
 from opentelemetry import trace
+from opentelemetry.instrumentation.fastapi import FastAPIInstrumentor
 from opentelemetry.instrumentation.httpx import HTTPXClientInstrumentor
 from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.sdk.trace.export import SimpleSpanProcessor
+from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
+from opentelemetry.trace import StatusCode
 
+from app.shared import telemetry
 from app.shared.config import Settings
+from app.shared.errors import OrderGuardError
 from app.shared.telemetry import is_initialized, setup_telemetry, shutdown_telemetry
 
 
+@pytest.fixture
+def spans(monkeypatch) -> InMemorySpanExporter:
+    """把 telemetry 内部取到的 tracer 换成写内存的，用于断言 span
+
+    不用全局 provider：`trace.set_tracer_provider` 只生效一次，一旦别的用例先
+    设过，这里再设就被忽略（OTel 只打一条 warning），断言会莫名其妙拿不到 span。
+    """
+    exporter = InMemorySpanExporter()
+    provider = TracerProvider()
+    provider.add_span_processor(SimpleSpanProcessor(exporter))
+    monkeypatch.setattr(telemetry.trace, "get_tracer", provider.get_tracer)
+    return exporter
+
+
 def test_disabled_is_noop():
     """otel_enabled=False 时 setup/instrument/shutdown 都不初始化 provider"""
     settings = Settings(_env_file=None)  # 默认 otel_enabled=False
@@ -24,6 +49,90 @@ def test_disabled_is_noop():
     assert not is_initialized()
 
 
+def test_instrument_app_works_before_setup_telemetry(monkeypatch):
+    """回归：instrument_app 必须在 setup_telemetry 之前也能真的装上中间件
+
+    三个服务都在模块导入时执行 `app = create_app()`（内部调 instrument_app），
+    而 setup_telemetry 要等 lifespan 才跑。曾经这里用 `_provider is None` 做前置
+    判断，于是导入期一律 return，FastAPI 从来没被打桩过——一条 server span 都
+    没有。这条用例把那个顺序钉住。
+    """
+    monkeypatch.setattr(
+        telemetry, "get_settings", lambda: Settings(_env_file=None, otel_enabled=True)
+    )
+    assert not is_initialized()  # 尚未 setup，正是导入期的状态
+
+    app = FastAPI()
+    try:
+        telemetry.instrument_app(app)
+        assert app._is_instrumented_by_opentelemetry
+    finally:
+        FastAPIInstrumentor.uninstrument_app(app)
+
+
+def test_instrument_app_skipped_when_otel_disabled(monkeypatch):
+    """otel 关闭时不装中间件，省掉一层用不上的开销"""
+    monkeypatch.setattr(
+        telemetry, "get_settings", lambda: Settings(_env_file=None, otel_enabled=False)
+    )
+    app = FastAPI()
+    telemetry.instrument_app(app)
+    assert not getattr(app, "_is_instrumented_by_opentelemetry", False)
+
+
+async def test_traced_records_span_and_reraises(spans):
+    """traced 成功时留一个 span；异常时记 ERROR 状态并原样抛出"""
+
+    @telemetry.traced("unit.ok")
+    async def ok() -> str:
+        return "done"
+
+    @telemetry.traced("unit.boom")
+    async def boom() -> None:
+        raise OrderGuardError("金额超限")
+
+    assert await ok() == "done"
+    with pytest.raises(OrderGuardError):
+        await boom()
+
+    finished = {s.name: s for s in spans.get_finished_spans()}
+    assert finished["unit.ok"].status.status_code is not StatusCode.ERROR
+    failed = finished["unit.boom"]
+    assert failed.status.status_code is StatusCode.ERROR
+    # AppError 的对外错误码要落在 span 上：排查时按码筛比按异常类名筛更贴近上游
+    assert failed.attributes["error.type"] == "OrderGuardError"
+    assert failed.attributes["error.code"] == OrderGuardError("x").err_code
+
+
+async def test_traced_nests_under_caller_span(spans):
+    """traced 出来的 span 要挂在调用方的 span 底下，而不是各自成为孤立 trace"""
+
+    @telemetry.traced("unit.child")
+    async def child() -> None:
+        return None
+
+    tracer = telemetry.trace.get_tracer(__name__)
+    with tracer.start_as_current_span("unit.root"):
+        await child()
+
+    by_name = {s.name: s for s in spans.get_finished_spans()}
+    assert by_name["unit.child"].parent.span_id == by_name["unit.root"].context.span_id
+    # 同一条 trace 才能在观测后台里连成一条链路
+    assert by_name["unit.child"].context.trace_id == by_name["unit.root"].context.trace_id
+
+
+def test_set_attributes_skips_none(spans):
+    """可选字段为 None 时不落属性，避免一堆 None 噪声"""
+    tracer = telemetry.trace.get_tracer(__name__)
+    with tracer.start_as_current_span("unit.attrs") as span:
+        telemetry.set_attributes(span, {"a": 1, "b": None, "c": "x"})
+
+    attributes = spans.get_finished_spans()[0].attributes
+    assert attributes["a"] == 1
+    assert attributes["c"] == "x"
+    assert "b" not in attributes
+
+
 def test_enabled_initializes_and_shutdown_releases():
     """enabled=true 时 setup 注册 TracerProvider，shutdown 后 _provider 复位"""
     settings = Settings(
diff --git a/tests/test_worker_runner.py b/tests/test_worker_runner.py
index 04e2a0a..191b503 100644
--- a/tests/test_worker_runner.py
+++ b/tests/test_worker_runner.py
@@ -20,8 +20,14 @@ from typing import Any
 
 import pytest
 
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.sdk.trace.export import SimpleSpanProcessor
+from opentelemetry.sdk.trace.export.in_memory_span_exporter import InMemorySpanExporter
+from opentelemetry.trace import StatusCode
+
 from app.shared.errors import BrowserDeadError, CheckoutBlockedError, OrderGuardError
 from app.shared.task_state import OrderState, TaskStatus
+from app.trading.worker import runner as runner_module
 from app.trading.worker import verify
 from app.trading.worker.evidence import EvidenceStore
 from app.trading.worker.local_db import LocalDB
@@ -900,6 +906,130 @@ async def test_monitor_order_stops_after_max_checks_without_finding_order(
     assert gateway.reports == []
 
 
+# ---- 链路追踪：一个任务一条 trace ----
+
+
+@pytest.fixture
+def spans(monkeypatch) -> InMemorySpanExporter:
+    """把 runner 模块级 tracer 换成写内存的
+
+    不动全局 provider：OTel 的全局 provider 只允许设置一次，测试间共享会互相污染。
+    """
+    exporter = InMemorySpanExporter()
+    provider = TracerProvider()
+    provider.add_span_processor(SimpleSpanProcessor(exporter))
+    monkeypatch.setattr(runner_module, "tracer", provider.get_tracer("test"))
+    return exporter
+
+
+def _stub_happy_path(runner: WorkerRunner) -> None:
+    """把整条下单流程替换成桩，让 execute() 一路走到付款完成"""
+
+    async def _clear():
+        return {"removed_count": 0, "cart_count": 0, "html": "<html/>", "screenshot": b"p"}
+
+    async def _snapshot(*args, **kwargs):
+        return PageSnapshot(html="<html/>", screenshot=b"p")
+
+    async def _parse(html: str):
+        return CheckoutSummary(payable_yen=297)
+
+    async def _submit(task):
+        return SubmitOutcome(
+            site_order_id="ord-1", evidence=PageSnapshot(html="<html/>", screenshot=b"p")
+        )
+
+    async def _unchanged(order_id: str):
+        return OrderStatusSnapshot(found=True, order_state=None)
+
+    runner._site.clear_cart = _clear  # type: ignore[assignment]
+    runner._site.add_to_cart = _snapshot  # type: ignore[assignment]
+    runner._site.verify_cart = _snapshot  # type: ignore[assignment]
+    runner._site.enter_checkout = _snapshot  # type: ignore[assignment]
+    runner._site.parse_checkout = _parse  # type: ignore[assignment]
+    runner._site.submit_order = _submit  # type: ignore[assignment]
+    runner._site.pay = _snapshot  # type: ignore[assignment]
+    runner._site.check_order_status = _unchanged  # type: ignore[assignment]
+
+
+async def test_order_task_emits_one_trace_with_a_span_per_step(
+    runner: WorkerRunner, spans: InMemorySpanExporter
+):
+    """一笔下单 = 一条 trace，每个步骤一个子 span，全部挂在任务根 span 底下
+
+    worker 是后台 asyncio 任务，没有 HTTP 入口。不开这个根 span 的话，下游每次
+    出站调用都各自成为孤立 trace——观测后台上就只剩「一堆请求记录」，看不出
+    这是同一笔单、也看不出卡在哪一步。这条用例把链路结构钉住。
+    """
+    _stub_happy_path(runner)
+
+    await runner.handle(_make_task(task_id="t1", intent={"item_url": "https://x/y/", "quantity": 2}))
+
+    finished = spans.get_finished_spans()
+    root = next(s for s in finished if s.name == "order.task")
+    steps = [s for s in finished if s.name.startswith("order.step.")]
+
+    # 步骤 span 覆盖整条流水，且顺序与执行顺序一致
+    assert [s.name for s in steps] == [
+        "order.step.cart-add",
+        "order.step.cart-check",
+        "order.step.order-confirm",
+        "order.step.order-submit",
+        "order.step.payment",
+    ]
+    # 每个步骤都挂在任务根底下，同一条 trace
+    for step in steps:
+        assert step.parent.span_id == root.context.span_id
+        assert step.context.trace_id == root.context.trace_id
+
+    assert root.attributes["order.task_id"] == "t1"
+    assert root.attributes["order.route"] == "execute"
+    assert root.attributes["order.item_url"] == "https://x/y/"
+    # 证据路径落在 span 上：看到失败步骤能直接找到现场
+    assert steps[0].attributes["order.evidence_ref"]
+    assert steps[3].attributes["order.site_order_id"] == "ord-1"
+
+
+async def test_guard_failure_marks_task_span_as_error(
+    runner: WorkerRunner, spans: InMemorySpanExporter
+):
+    """金额守卫拦下时任务根 span 要是 ERROR，并带上对外错误码
+
+    闸门拦截会被 _execute_with_renewal 吞掉转 needs_human（不抛出 handle），
+    所以根 span 很容易显示成功——这里确认它没有。
+    """
+    _stub_happy_path(runner)
+
+    async def _pricey(html: str):
+        return CheckoutSummary(payable_yen=999_999)
+
+    runner._site.parse_checkout = _pricey  # type: ignore[assignment]
+
+    await runner.handle(_make_task(task_id="t1"))
+
+    root = next(s for s in spans.get_finished_spans() if s.name == "order.task")
+    assert root.status.status_code is StatusCode.ERROR
+    assert root.attributes["error.type"] == "OrderGuardError"
+    # 提交步骤没有发生：拦在确认页之后、提交之前
+    names = [s.name for s in spans.get_finished_spans()]
+    assert "order.step.order-submit" not in names
+
+
+async def test_idle_lease_does_not_emit_spans(runner: WorkerRunner, spans: InMemorySpanExporter):
+    """空转的长轮询不产生 span——否则每 30 秒一条孤立 trace 会把观测后台刷满"""
+    runner._running = True
+
+    async def _lease_then_stop(worker_id: str, *, wait: int = 30, site: str | None = None):
+        runner.stop()  # 领到空就停，避免死循环
+        return None
+
+    runner._gateway.lease = _lease_then_stop  # type: ignore[assignment]
+
+    await runner.run()
+
+    assert spans.get_finished_spans() == ()
+
+
 async def test_monitor_order_swallows_check_errors_and_keeps_polling(
     runner: WorkerRunner,
 ):