Merge branch 'main' into codex/user-draw-ecpm-scope

## 问题

比价记录进入中途退出后未触发 LLM 成本回填，周期补偿也未扫描 cancelled，导致实际已有 LLM 调用的记录长期显示成本、LLM、TOKEN 为空。

## 修改

- finalize 落库后立即追加 LLM 成本回填
- 周期补偿范围加入 cancelled
- 保持无有效调用和全调用失败记录不伪造成本
- 增加即时回填和周期补偿回归测试

## 验证

- ruff 检查通过
- 相关测试 28 项通过
- 全仓 626 项通过；主干既有失败已在未修改的 origin/main 复现

---------

Co-authored-by: unknown <798648091@qq.com>
Reviewed-on: #191
Co-authored-by: linkeyu <linkeyu@wonderable.ai>
Co-committed-by: linkeyu <linkeyu@wonderable.ai>

app/admin/repositories/ad_revenue.py

@@ -43,7 +43,7 @@ _KNOWN_PROD_BUSINESS_CODE_IDS = frozenset({"104098712", "104099389"})
 _TEST_BUSINESS_CODE_IDS = frozenset({"104127529", "104127626", "104137445"})
 
 
-def _business_code_ids(db: Session, app_env: str | None) -> set[str]:
+def business_code_ids(db: Session, app_env: str | None) -> set[str]:
     """返回指定应用环境下可用于业务收益对账的 GroMore 聚合代码位。"""
     prod_config = app_config.get_ad_config(db)
     prod_ids = set(_KNOWN_PROD_BUSINESS_CODE_IDS) | {
@@ -320,10 +320,10 @@ def ad_revenue_report(
 
     # 业务口径仅保留正式配置/测试业务链路实际使用的代码位。穿山甲“全量”还包含广告测试
     # demo、插屏等没有客户端收益上报的曝光，两边直接比较会天然产生假差额。
-    business_code_ids: set[str] | None = None
+    business_ids: set[str] | None = None
     if revenue_scope == "business":
-        business_code_ids = _business_code_ids(db, app_env)
-        events = [e for e in events if e.get("our_code_id") in business_code_ids]
+        business_ids = business_code_ids(db, app_env)
+        events = [e for e in events if e.get("our_code_id") in business_ids]
 
     # 排序:time=按时间倒序(新→旧);ecpm=按 eCPM 数值倒序(eCPM 原值是字符串「分」,转数值排;
     # 纯发奖行用其发奖采用的 eCPM,缺失/非法计 0 排末尾)。
@@ -381,7 +381,7 @@ def ad_revenue_report(
             date_from=date_from,
             date_to=date_to,
             app_env=app_env,
-            our_code_ids=business_code_ids,
+            our_code_ids=business_ids,
         )
         if pangle_aggs:
             by_date = {a["date"]: a for a in pangle_aggs}

app/admin/repositories/queries.py

@@ -14,6 +14,7 @@ from sqlalchemy.orm import Session
 
 from app.core import rewards
 from app.core.config import settings
+from app.models.ad_ecpm import AdEcpmRecord
 from app.models.ad_feed_reward import AdFeedRewardRecord
 from app.models.ad_reward import AdRewardRecord
 from app.models.admin import AdminAuditLog
@@ -1257,11 +1258,15 @@ def user_reward_stats(
     date_from: datetime | None = None,
     date_to: datetime | None = None,
     withdraw_source: str | None = None,
+    app_env: str | None = None,
+    revenue_scope: str = "all",
+    feed_scene: str | None = None,
 ) -> dict:
     """提现详情「用户统计区」10 项。窗口作用于除「现金余额」外的所有项(余额是当前快照)。
 
-    口径:激励视频/信息流只统计 granted;数量——视频按条数、信息流按份数(unit_count 累加);
-    平均 eCPM 用原始分值(分/千次)按记录取算术平均;各「提现」= 该来源累计金币折现。
+    口径:激励视频/信息流奖励数量只统计 granted;数量——视频按条数、信息流按份数(unit_count 累加)。
+    平均 Draw eCPM 与广告收益报表一致:基于 ad_ecpm_record 的全部 draw/feed 展示记录计算，
+    不以是否发奖为筛选条件。各「提现」= 该来源累计金币折现。
     传统任务 = 窗口内正向金币中,排除广告(reward_video/feed_ad_reward)与人工调整后的折现。
     """
     withdraw_source_conds = (
@@ -1296,30 +1301,68 @@ def user_reward_stats(
 
     # 只投影本统计实际使用的列。避免滚动发布或旧本地库尚未补齐无关新列时，
     # SQLAlchemy 因 select(ORM) 自动展开整表字段而让提现详情整体 500。
+    business_ids: set[str] | None = None
+    if revenue_scope == "business":
+        # 与广告收益报表共用正式/测试业务代码位集合，避免两个页面随配置切换后再次漂移。
+        from app.admin.repositories.ad_revenue import business_code_ids
+
+        business_ids = business_code_ids(db, app_env)
+
+    rv_conds = [
+        AdRewardRecord.user_id == user_id,
+        AdRewardRecord.reward_scene == "reward_video",
+        AdRewardRecord.status == "granted",
+        *_window_conds(AdRewardRecord.created_at, date_from, date_to),
+    ]
+    if app_env is not None:
+        rv_conds.append(AdRewardRecord.app_env == app_env)
+    if business_ids is not None:
+        rv_conds.append(AdRewardRecord.our_code_id.in_(business_ids))
     rv = db.execute(
         select(AdRewardRecord.ecpm_raw, AdRewardRecord.coin).where(
-            AdRewardRecord.user_id == user_id,
-            AdRewardRecord.reward_scene == "reward_video",
-            AdRewardRecord.status == "granted",
-            *_window_conds(AdRewardRecord.created_at, date_from, date_to),
+            *rv_conds,
         )
     ).all()
     rv_ecpms = [rewards.parse_ecpm_fen(r.ecpm_raw) for r in rv if r.ecpm_raw]
     rv_coins = sum(r.coin for r in rv)
 
-    feed = db.execute(
+    feed_reward_conds = [
+        AdFeedRewardRecord.user_id == user_id,
+        AdFeedRewardRecord.status == "granted",
+        *_window_conds(AdFeedRewardRecord.created_at, date_from, date_to),
+    ]
+    if app_env is not None:
+        feed_reward_conds.append(AdFeedRewardRecord.app_env == app_env)
+    if feed_scene is not None:
+        feed_reward_conds.append(AdFeedRewardRecord.feed_scene == feed_scene)
+    if business_ids is not None:
+        feed_reward_conds.append(AdFeedRewardRecord.our_code_id.in_(business_ids))
+    feed_rewards = db.execute(
         select(
             AdFeedRewardRecord.unit_count,
-            AdFeedRewardRecord.ecpm_raw,
             AdFeedRewardRecord.coin,
         ).where(
-            AdFeedRewardRecord.user_id == user_id,
-            AdFeedRewardRecord.status == "granted",
-            *_window_conds(AdFeedRewardRecord.created_at, date_from, date_to),
+            *feed_reward_conds,
         )
     ).all()
-    feed_ecpms = [rewards.parse_ecpm_fen(f.ecpm_raw) for f in feed if f.ecpm_raw]
-    feed_coins = sum(f.coin for f in feed)
+    feed_coins = sum(f.coin for f in feed_rewards)
+
+    feed_impression_conds = [
+        AdEcpmRecord.user_id == user_id,
+        AdEcpmRecord.ad_type.in_(("draw", "feed")),
+        *_window_conds(AdEcpmRecord.created_at, date_from, date_to),
+    ]
+    if app_env is not None:
+        feed_impression_conds.append(AdEcpmRecord.app_env == app_env)
+    if feed_scene is not None:
+        feed_impression_conds.append(AdEcpmRecord.feed_scene == feed_scene)
+    if business_ids is not None:
+        feed_impression_conds.append(AdEcpmRecord.our_code_id.in_(business_ids))
+    feed_impressions = db.execute(
+        select(AdEcpmRecord.ecpm_raw).where(*feed_impression_conds)
+    ).all()
+    # 与 ad_revenue.category_stats 相同:每次展示权重相同，非法原值按 parse_ecpm_fen 记 0。
+    feed_ecpms = [rewards.parse_ecpm_fen(row.ecpm_raw) for row in feed_impressions]
 
     trad_coins = db.execute(
         select(func.coalesce(func.sum(CoinTransaction.amount), 0)).where(
@@ -1338,7 +1381,7 @@ def user_reward_stats(
         "reward_video_count": len(rv),
         "reward_video_avg_ecpm": round(sum(rv_ecpms) / len(rv_ecpms), 2) if rv_ecpms else 0.0,
         "reward_video_cash_cents": _coins_to_cents(rv_coins),
-        "feed_count": int(sum(f.unit_count for f in feed)),
+        "feed_count": int(sum(f.unit_count for f in feed_rewards)),
         "feed_avg_ecpm": round(sum(feed_ecpms) / len(feed_ecpms), 2) if feed_ecpms else 0.0,
         "feed_cash_cents": _coins_to_cents(feed_coins),
     }

app/admin/routers/users.py

@@ -88,6 +88,11 @@ def get_user_reward_stats(
     withdraw_source: Annotated[
         str | None, Query(pattern="^(coin_cash|invite_cash)$")
     ] = None,
+    app_env: Annotated[str | None, Query(pattern="^(prod|test)$")] = None,
+    revenue_scope: Annotated[str, Query(pattern="^(business|all)$")] = "all",
+    feed_scene: Annotated[
+        str | None, Query(pattern="^(comparison|coupon|welfare)$")
+    ] = None,
 ) -> UserRewardStats:
     """提现详情抽屉「用户统计区」。date_from/date_to 都不传 = 注册至今(全量)。"""
     if not user_repo.user_exists(db, user_id):
@@ -99,6 +104,9 @@ def get_user_reward_stats(
             date_from=date_from,
             date_to=date_to,
             withdraw_source=withdraw_source,
+            app_env=app_env,
+            revenue_scope=revenue_scope,
+            feed_scene=feed_scene,
         )
     )
 

app/admin/schemas/user.py

@@ -60,7 +60,7 @@ class UserRewardStats(BaseModel):
     reward_video_avg_ecpm: float  # 平均激励视频 eCPM(分/千次)
     reward_video_cash_cents: int  # 激励视频提现(金币折现)
     feed_count: int  # 累计信息流广告数(granted 份数,unit_count 累加)
-    feed_avg_ecpm: float  # 平均信息流广告 eCPM(分/千次)
+    feed_avg_ecpm: float  # 全部 Draw/feed 实际展示的平均 eCPM(分/千次,含未发奖展示)
     feed_cash_cents: int  # 信息流广告提现(金币折现)
 
 

app/api/v1/compare.py

@@ -108,7 +108,7 @@ def _harvest_done_blocking(
 
 def _harvest_abort_blocking(
     trace_id: str, status_hint: str, reason: str | None, trace_url: str | None,
-) -> None:
+) -> int | None:
     with SessionLocal() as db:
         rec = crud_compare.harvest_abort(
             db, trace_id=trace_id, status=status_hint, reason=reason, trace_url=trace_url,
@@ -118,6 +118,7 @@ def _harvest_abort_blocking(
         extra={"phase": "harvest_abort",
                "status": (rec.status if rec else None), "reason": reason},
     )
+    return rec.id if rec is not None else None
 
 
 async def _forward(
@@ -291,7 +292,10 @@ async def trace_epilogue(
 
 @router.post("/trace/finalize", summary="比价 trace 收尾上云 (透传 + 夭折落库)")
 async def trace_finalize(
-    request: Request, user: OptionalUser, db: DbSession
+    request: Request,
+    background_tasks: BackgroundTasks,
+    user: OptionalUser,
+    db: DbSession,
 ) -> dict[str, Any]:
     _ensure_compare_allowed(user, db)
     # 用户终止 / Phase1 未识别没到 done 帧: pricebot 打包半截上云返回 {trace_url};
@@ -302,12 +306,16 @@ async def trace_finalize(
         request, "/api/trace/finalize", user, harvest_first_frame=False,
     )
     try:
-        await run_in_threadpool(
+        record_id = await run_in_threadpool(
             _harvest_abort_blocking, trace_id,
             (meta.get("status") or "cancelled"),
             (meta.get("reason") or meta.get("information")),
             (resp.get("trace_url") if isinstance(resp, dict) else None),
         )
+        if record_id is not None:
+            background_tasks.add_task(
+                backfill_comparison_llm_cost, record_id, trace_id
+            )
     except Exception as e:  # noqa: BLE001
         logger.warning("harvest_abort failed trace=%s: %s", trace_id, e)
     return resp

app/services/comparison_llm_backfill.py

@@ -135,7 +135,7 @@ def repair_missing_comparison_llm_costs(
                 select(ComparisonRecord.id, ComparisonRecord.trace_id)
                 .where(
                     *date_conditions,
-                    ComparisonRecord.status.in_(("success", "failed")),
+                    ComparisonRecord.status.in_(("success", "failed", "cancelled")),
                     ComparisonRecord.llm_cost_yuan.is_(None),
                 )
                 .order_by(ComparisonRecord.created_at.desc(), ComparisonRecord.id.desc())

tests/test_admin_read.py

@@ -219,6 +219,115 @@ def test_user_reward_stats_can_scope_withdrawals_by_account(
     assert invite.json()["cash_balance_cents"] == 456
 
 
+def test_user_reward_stats_draw_ecpm_uses_all_filtered_impressions(
+    admin_client: TestClient, admin_token: str
+) -> None:
+    """Draw 平均 eCPM 应与收益报表一致，不能只平均成功发奖记录。"""
+    from app.models.ad_ecpm import AdEcpmRecord
+    from app.models.ad_feed_reward import AdFeedRewardRecord
+
+    uid = _seed_user_with_data("13800000024")
+    created_at = datetime(2038, 1, 15, 4, tzinfo=UTC)
+    db = SessionLocal()
+    try:
+        db.add_all(
+            [
+                AdFeedRewardRecord(
+                    client_event_id="reward-stats-granted-high",
+                    ad_session_id="reward-stats-granted-high",
+                    user_id=uid,
+                    reward_date="2038-01-15",
+                    duration_seconds=10,
+                    unit_count=1,
+                    ecpm_raw="9000",
+                    ad_type="draw",
+                    feed_scene="coupon",
+                    app_env="prod",
+                    our_code_id="104098712",
+                    coin=9,
+                    status="granted",
+                    created_at=created_at,
+                ),
+                AdEcpmRecord(
+                    user_id=uid,
+                    ad_type="draw",
+                    feed_scene="coupon",
+                    ad_session_id="reward-stats-impression-low",
+                    app_env="prod",
+                    our_code_id="104098712",
+                    ecpm_raw="1000",
+                    report_date="2038-01-15",
+                    created_at=created_at,
+                ),
+                AdEcpmRecord(
+                    user_id=uid,
+                    ad_type="feed",
+                    feed_scene="coupon",
+                    ad_session_id="reward-stats-impression-mid",
+                    app_env="prod",
+                    our_code_id="104098712",
+                    ecpm_raw="3000",
+                    report_date="2038-01-15",
+                    created_at=created_at,
+                ),
+                # 同用户但不同场景/环境/非业务代码位，均不应进入本次详情筛选。
+                AdEcpmRecord(
+                    user_id=uid,
+                    ad_type="draw",
+                    feed_scene="comparison",
+                    ad_session_id="reward-stats-other-scene",
+                    app_env="prod",
+                    our_code_id="104098712",
+                    ecpm_raw="7000",
+                    report_date="2038-01-15",
+                    created_at=created_at,
+                ),
+                AdEcpmRecord(
+                    user_id=uid,
+                    ad_type="draw",
+                    feed_scene="coupon",
+                    ad_session_id="reward-stats-test-env",
+                    app_env="test",
+                    our_code_id="104127529",
+                    ecpm_raw="8000",
+                    report_date="2038-01-15",
+                    created_at=created_at,
+                ),
+                AdEcpmRecord(
+                    user_id=uid,
+                    ad_type="draw",
+                    feed_scene="coupon",
+                    ad_session_id="reward-stats-non-business",
+                    app_env="prod",
+                    our_code_id="demo-slot",
+                    ecpm_raw="9000",
+                    report_date="2038-01-15",
+                    created_at=created_at,
+                ),
+            ]
+        )
+        db.commit()
+    finally:
+        db.close()
+
+    response = admin_client.get(
+        f"/admin/api/users/{uid}/reward-stats",
+        params={
+            "date_from": "2038-01-15T00:00:00Z",
+            "date_to": "2038-01-15T23:59:59Z",
+            "app_env": "prod",
+            "revenue_scope": "business",
+            "feed_scene": "coupon",
+        },
+        headers=_auth(admin_token),
+    )
+    assert response.status_code == 200, response.text
+    data = response.json()
+    assert data["feed_count"] == 1
+    # 全部真实展示 (1000 + 3000) / 2；不能返回成功发奖记录的 9000。
+    assert data["feed_avg_ecpm"] == 2000.0
+
+
 def test_user_coin_record_sort_accepts_mixed_timezone_datetimes() -> None:
     """线上 PostgreSQL 返回 aware，SQLite/历史转换可能返回 naive，二者必须可混排。"""
     naive = datetime(2038, 1, 1, 8, 0)

tests/test_compare_harvest.py

@@ -263,7 +263,9 @@ def test_trace_finalize_harvests_abort(client) -> None:
     with SessionLocal() as db:                        # 先有 running 行(帧0建的)
         crud.harvest_running(db, trace_id=tid, user_id=None)
     p, _cap = _mock_pricebot({"trace_url": "https://price.shaguabijia.com/traces/fin/"})
-    with p:
+    with p, patch(
+        "app.api.v1.compare.backfill_comparison_llm_cost"
+    ) as backfill:
         r = client.post("/api/v1/trace/finalize",
                         json={"trace_id": tid, "status": "cancelled", "reason": "用户终止"})
     assert r.status_code == 200
@@ -271,6 +273,7 @@ def test_trace_finalize_harvests_abort(client) -> None:
         rec = _get(db, tid)
         assert rec is not None and rec.status == "cancelled"
         assert rec.trace_url.endswith("/fin/")
+        backfill.assert_called_once_with(rec.id, tid)
 
 
 def test_price_step_binds_user_when_authed(client) -> None:

tests/test_comparison_llm_backfill.py

@@ -72,6 +72,7 @@ def test_backfill_retries_then_persists_cost(monkeypatch):
 
 def test_repair_batch_only_targets_terminal_missing_rows(monkeypatch):
     missing_id = _record("llm-repair-missing")
+    cancelled_id = _record("llm-repair-cancelled", status="cancelled")
     running_id = _record("llm-repair-running", status="running")
     calls = [
         {
@@ -96,12 +97,15 @@ def test_repair_batch_only_targets_terminal_missing_rows(monkeypatch):
         )
         assert result["repaired"] >= 1
         assert "llm-repair-missing" in seen
+        assert "llm-repair-cancelled" in seen
         assert "llm-repair-running" not in seen
         with SessionLocal() as db:
             assert db.get(ComparisonRecord, missing_id).llm_cost_yuan is not None
+            assert db.get(ComparisonRecord, cancelled_id).llm_cost_yuan is not None
             assert db.get(ComparisonRecord, running_id).llm_cost_yuan is None
     finally:
         _delete(missing_id)
+        _delete(cancelled_id)
         _delete(running_id)