feat: 四层架构全面增强

安全与稳定性:
- 移除硬编码 API Key，改用 .env + 环境变量
- LLM 调用统一重试机制（指数退避，3 次重试，处理 429/5xx/超时）
- 中文字体检测增强（CJK 关键词兜底 + 无字体时英文 fallback）
- 缺失 API Key 给出友好提示而非崩溃

分析能力提升:
- 异常检测新增 z-score 检测（标准差>2 标记异常）
- 新增变异系数 CV 检测（数据波动性）
- 新增零值/缺失检测
- 上下文管理器升级为关键词语义匹配（替代简单取最近 2 条）

用户体验:
- 报告自动保存为 Markdown（reports/ 目录）
- 新增 export 命令导出查询结果为 CSV
- 新增 reports 命令查看已保存报告
- CLI 支持 readline 命令历史（方向键翻阅）
- CSV 导入工具重写：自动列名映射、容错处理、dry-run 模式
- 新增 .env.example 配置模板

layers/insights.py

@@ -5,7 +5,7 @@ import json
 from typing import Any
 
 from core.config import LLM_CONFIG
-from core.utils import get_llm_client, extract_json_array
+from core.utils import get_llm_client, llm_chat, extract_json_array
 from layers.explorer import ExplorationStep
 
 
@@ -68,15 +68,14 @@ class InsightEngine:
             return []
 
         history = self._build_history(steps)
-        response = self.client.chat.completions.create(
-            model=self.model,
+        content = llm_chat(
+            self.client, self.model,
             messages=[
                 {"role": "system", "content": INSIGHT_SYSTEM},
                 {"role": "user", "content": f"## 用户问题\n{question}\n\n## 探索历史\n{history}\n\n请分析以上数据，输出异常和洞察。"},
             ],
             temperature=0.3, max_tokens=2048,
         )
-        content = response.choices[0].message.content.strip()
         return [Insight(d) for d in extract_json_array(content)]
 
     def format_insights(self, insights: list[Insight]) -> str:
@@ -109,9 +108,9 @@ class InsightEngine:
 
 
 def quick_detect(steps: list[ExplorationStep]) -> list[str]:
-    """基于规则的快速异常检测，不调 LLM"""
+    """基于规则的快速异常检测（零 LLM 成本）"""
     alerts = []
-    seen = set()  # 去重
+    seen = set()
 
     for step in steps:
         if not step.success or not step.rows:
@@ -119,22 +118,21 @@ def quick_detect(steps: list[ExplorationStep]) -> list[str]:
 
         for col in step.columns:
             vals = [r.get(col) for r in step.rows if isinstance(r.get(col), (int, float))]
-            if not vals:
+            if len(vals) < 2:
                 continue
 
             col_lower = col.lower()
 
-            # 占比列：某个分组占比过高
+            # ── 占比列：集中度过高 ──
             if col_lower in ("pct", "percent", "percentage", "占比"):
-                for v in vals:
-                    if v > 50:
-                        key = f"pct_{step.purpose}"
-                        if key not in seen:
-                            seen.add(key)
-                            alerts.append(f"⚠️ {step.purpose} 中某个分组占比 {v}%，集中度过高")
-                        break
+                max_pct = max(vals)
+                if max_pct > 50:
+                    key = f"pct_{step.purpose}"
+                    if key not in seen:
+                        seen.add(key)
+                        alerts.append(f"⚠️ {step.purpose}: 最高占比 {max_pct}%，集中度过高")
 
-            # 计数列：极值差异
+            # ── 计数列：极值差异 ──
             if col_lower in ("count", "cnt", "n", "total", "order_count") and len(vals) >= 3:
                 avg = sum(vals) / len(vals)
                 if avg > 0:
@@ -143,6 +141,47 @@ def quick_detect(steps: list[ExplorationStep]) -> list[str]:
                         key = f"count_{step.purpose}"
                         if key not in seen:
                             seen.add(key)
-                            alerts.append(f"⚠️ {step.purpose} 中最大值是均值的 {ratio:.1f} 倍")
+                            alerts.append(f"⚠️ {step.purpose}: 最大值是均值的 {ratio:.1f} 倍")
+
+            # ── Z-Score 异常检测 ──
+            if len(vals) >= 5 and col_lower not in ("id", "year", "month"):
+                mean = sum(vals) / len(vals)
+                variance = sum((v - mean) ** 2 for v in vals) / len(vals)
+                std = variance ** 0.5
+                if std > 0:
+                    outliers = [(i, v) for i, v in enumerate(vals) if abs(v - mean) / std > 2]
+                    if outliers:
+                        key = f"zscore_{step.purpose}_{col}"
+                        if key not in seen:
+                            seen.add(key)
+                            outlier_desc = ", ".join(f"{v:.1f}" for _, v in outliers[:3])
+                            alerts.append(
+                                f"⚠️ {step.purpose}「{col}」发现 {len(outliers)} 个异常值 "
+                                f"(均值={mean:.1f}, σ={std:.1f}, 异常值={outlier_desc})"
+                            )
+
+            # ── 离散度检测（变异系数 CV）──
+            if len(vals) >= 3 and col_lower not in ("id", "year", "month"):
+                mean = sum(vals) / len(vals)
+                if mean != 0:
+                    variance = sum((v - mean) ** 2 for v in vals) / len(vals)
+                    std = variance ** 0.5
+                    cv = std / abs(mean)
+                    if cv > 1.0:
+                        key = f"cv_{step.purpose}_{col}"
+                        if key not in seen:
+                            seen.add(key)
+                            alerts.append(f"⚠️ {step.purpose}「{col}」离散度高 (CV={cv:.2f})，数据波动大")
+
+            # ── 零值/缺失检测 ──
+            if col_lower in ("count", "cnt", "total", "amount", "sum", "关闭时长"):
+                zero_count = sum(1 for v in vals if v == 0)
+                if zero_count > 0 and zero_count < len(vals):
+                    pct = zero_count / len(vals) * 100
+                    if pct > 10:
+                        key = f"zero_{step.purpose}_{col}"
+                        if key not in seen:
+                            seen.add(key)
+                            alerts.append(f"⚠️ {step.purpose}「{col}」有 {zero_count} 个零值 ({pct:.0f}%)")
 
     return alerts