feat: 预算系统6项技术改进（告警归因/实际值自动归集/真零基/派生规则/告警路径统一/现金流分类）

P1-③ 告警归因: budget_deviation_alerts+alert_type/attribution/scenario_id, 归因引擎alert_attribution.py(子KPI/科目/量价差/趋势), deviation-check统一写归因+场景, GET /deviation-alerts/{id}/attribution详情(旧告警现场组装)
P1-④ 实际值自动归集: kpi_value_sources/kpi_value_collect_logs表+CRUD+试跑+覆盖率, 采集器kpi_value_collector.py(voucher_details/进销存/cash_plans按entity+period汇总, 幂等upsert不覆盖人工), crontab每日06:30
P2-① 真零基: budget_zero_based_items逐项论证表+generate, method-comparison有论证项逐项求和is_demo=false否则fallback
P2-② 派生规则: budget_derivation_rules配置表, apply-method优先读规则rule_source=configured
P2-⑤ 告警双路径合并: deviation_engine.build_deviation_alert统一函数, 方向列表配置化kpi_alert_higher_better+alert-direction接口
P2-⑥ 现金流分类: cash_plan_classify_rules规则表+cash_plan_unclassified待分类队列, sync-cash-plans未命中进队列不静默跳过
新增: GET /kpis/{kpi_id}/values + 前端kpiApi.values(归集标签页数据源), scenario_suggestions幂等seed(init_db)
测试: test_budget_tech_improve.py 15用例, 预算相关96 passed, 全量646 passed

backend/app/utils/alert_attribution.py

@@ -0,0 +1,203 @@
+"""告警归因引擎 — 管理会计OS (P1-③ 2026-08-28)
+
+告警从"差多少"到"差在哪+怎么办"：
+  - 子KPI维度拆解: 查 kpi_hierarchy 下级KPI各自差异（量差方向）
+  - 科目明细拆解: 查 kpi_subject_map → voucher_details 汇总科目发生额（价差方向）
+  - 趋势归因: 复用 deviation_engine.check_trend_anomaly 连续3期检测
+  - 场景建议: 按 alert_type 联查 scenario_suggestions
+
+attribution JSON 结构:
+{
+  "dimensions": [{"kpi_id":1,"kpi_name":"销售费用","deviation_value":-3.2,"deviation_rate":-18.6,"weight":0.5}],
+  "subjects": [{"subject_code":"6601","subject_name":"销售费用","amount_diff":2.1,"share_pct":34.5}],
+  "variance_type": "quantity_diff|price_diff|mixed",
+  "trend": {"anomaly":true,"type":"continuous_decline","periods":["2026-06","2026-07","2026-08"],"message":"连续3期下滑"}
+}
+"""
+import logging
+from typing import Optional
+
+from sqlalchemy.orm import Session
+from sqlalchemy import func
+
+from app.models import KPIHierarchy, KPISubjectMap, VoucherDetail, ScenarioSuggestion, KPIDefinition
+
+logger = logging.getLogger("cma.alert_attribution")
+
+# 收入型KPI特征（量差方向: 子KPI量级偏离）
+REVENUE_TYPE_CODES = (
+    "SALES_TOTAL", "REVENUE", "F_REVENUE", "SALES_PROFIT_RATE",
+    "CUSTOMER_COUNT", "NEW_CUSTOMER", "TURNOVER_RATE",
+)
+
+
+def build_dimension_attribution(db: Session, kpi_id: int, period: str) -> list:
+    """子KPI维度拆解 — 查 kpi_hierarchy 下级KPI各自差异（实际vs预算）"""
+    from app.models import KPIValue, BudgetPlan
+
+    children = db.query(KPIHierarchy).filter(
+        KPIHierarchy.parent_kpi_id == kpi_id
+    ).all()
+    if not children:
+        return []
+
+    result = []
+    for rel in children:
+        child_id = rel.child_kpi_id
+        actual = db.query(KPIValue).filter(
+            KPIValue.kpi_id == child_id,
+            KPIValue.period == period,
+            KPIValue.actual_value.isnot(None),
+        ).order_by(KPIValue.calculated_at.desc()).first()
+        budget = db.query(BudgetPlan).filter(
+            BudgetPlan.kpi_id == child_id,
+            BudgetPlan.period == period,
+            BudgetPlan.status == "active",
+        ).order_by(BudgetPlan.updated_at.desc()).first()
+
+        av = actual.actual_value if actual else None
+        bv = budget.budget_value if budget else None
+        dev_value = None
+        dev_rate = None
+        if av is not None and bv is not None and bv != 0:
+            dev_value = round(av - bv, 2)
+            dev_rate = round(dev_value / bv * 100, 2)
+
+        child_kpi = db.query(KPIDefinition).filter(KPIDefinition.id == child_id).first()
+        result.append({
+            "kpi_id": child_id,
+            "kpi_name": child_kpi.kpi_name if child_kpi else f"KPI-{child_id}",
+            "actual_value": av,
+            "budget_value": bv,
+            "deviation_value": dev_value,
+            "deviation_rate": dev_rate,
+            "weight": float(rel.weight or 0),
+        })
+    # 按偏差绝对值降序，最异常的排前面
+    result.sort(key=lambda x: -(abs(x["deviation_value"]) if x["deviation_value"] is not None else 0))
+    return result
+
+
+def build_subject_attribution(db: Session, kpi_id: int, period: str) -> list:
+    """科目明细拆解 — 查 kpi_subject_map → voucher_details 汇总科目发生额"""
+    mappings = db.query(KPISubjectMap).filter(KPISubjectMap.kpi_id == kpi_id).all()
+    if not mappings:
+        return []
+
+    result = []
+    for m in mappings:
+        q = db.query(
+            func.coalesce(func.sum(VoucherDetail.debit_amount), 0),
+            func.coalesce(func.sum(VoucherDetail.credit_amount), 0),
+        ).filter(
+            VoucherDetail.subject_code == m.subject_code,
+            VoucherDetail.period == period,
+        )
+        row = q.first()
+        debit_sum = float(row[0] or 0)
+        credit_sum = float(row[1] or 0)
+        # 方向: credit贷方(收入/流入) / debit借方(费用/流出)
+        if m.calc_type == "ratio":
+            amount = credit_sum - debit_sum
+        elif m.calc_type in ("avg", "other"):
+            amount = (credit_sum - debit_sum) / 2
+        else:  # sum
+            amount = credit_sum - debit_sum
+        amount = round(amount * float(m.weight or 1.0), 2)
+
+        result.append({
+            "subject_code": m.subject_code,
+            "subject_name": m.remark or m.subject_code,
+            "amount_diff": amount,
+            "calc_type": m.calc_type,
+            "weight": float(m.weight or 1.0),
+        })
+
+    total = sum(abs(r["amount_diff"]) for r in result) or 0
+    for r in result:
+        r["share_pct"] = round(abs(r["amount_diff"]) / total * 100, 1) if total else 0
+    result.sort(key=lambda x: -abs(x["amount_diff"]))
+    return result
+
+
+def detect_variance_type(kpi_code: str, dimensions: list, subjects: list) -> str:
+    """量价差判定简化版：
+    成本型KPI科目发生额偏离 → price_diff（价差）
+    收入型KPI子KPI量级偏离 → quantity_diff（量差）
+    两者都有 → mixed
+    """
+    has_dimension_dev = any(d.get("deviation_value") is not None and abs(d["deviation_value"]) > 0.01 for d in dimensions)
+    has_subject_dev = any(abs(s.get("amount_diff", 0)) > 0.01 for s in subjects)
+
+    is_revenue = any(code in (kpi_code or "").upper() for code in REVENUE_TYPE_CODES)
+
+    if is_revenue:
+        # 收入型: 子KPI(量)偏离为主 → quantity_diff
+        if has_dimension_dev:
+            return "quantity_diff"
+        if has_subject_dev:
+            return "price_diff"
+        return "mixed"
+    else:
+        # 成本型: 科目发生额(价)偏离为主 → price_diff
+        if has_subject_dev:
+            return "price_diff"
+        if has_dimension_dev:
+            return "quantity_diff"
+        return "mixed"
+
+
+def match_scenario(db: Session, alert_type: Optional[str]) -> Optional[dict]:
+    """按 alert_type 取 scenario_suggestions 建议（四类模板）"""
+    if not alert_type:
+        return None
+    s = db.query(ScenarioSuggestion).filter(
+        ScenarioSuggestion.alert_type == alert_type
+    ).order_by(ScenarioSuggestion.sort_order.asc(), ScenarioSuggestion.id.asc()).first()
+    if not s:
+        return None
+    return {
+        "scenario_id": s.id,
+        "alert_type": s.alert_type,
+        "title": s.title,
+        "description": s.description,
+        "action_template": s.action_template,
+        "priority": s.priority,
+    }
+
+
+def build_attribution(db: Session, kpi_id: int, period: str, alert_type: Optional[str] = None) -> dict:
+    """组装完整归因JSON（供告警生成/详情接口共用）"""
+    from app.utils.deviation_engine import check_trend_anomaly
+
+    kpi = db.query(KPIDefinition).filter(KPIDefinition.id == kpi_id).first()
+    kpi_code = kpi.kpi_code if kpi else ""
+
+    dimensions = build_dimension_attribution(db, kpi_id, period)
+    subjects = build_subject_attribution(db, kpi_id, period)
+    variance_type = detect_variance_type(kpi_code, dimensions, subjects)
+    trend = check_trend_anomaly(db, kpi_id, period, consecutive=3)
+
+    # 默认场景归类（未显式传入时按KPI名称特征推断）
+    if not alert_type:
+        alert_type = infer_alert_type(kpi_code, kpi.kpi_name if kpi else "")
+
+    attribution = {
+        "dimensions": dimensions,
+        "subjects": subjects,
+        "variance_type": variance_type,
+        "trend": trend,
+    }
+    return attribution, alert_type
+
+
+def infer_alert_type(kpi_code: str = "", kpi_name: str = "") -> str:
+    """按KPI特征推断告警场景类型（四类: cash_low/cash_critical/cost_high/revenue_drop）"""
+    text = (kpi_code or "").upper() + (kpi_name or "")
+    if any(k in text for k in ("CASH", "现金", "货币资金", "资金")):
+        return "cash_low"
+    if any(k in text for k in ("COST", "费用", "成本", "支出")):
+        return "cost_high"
+    if any(k in text for k in ("REVENUE", "收入", "销售", "营收")):
+        return "revenue_drop"
+    return "cost_high"