feat(sense): T-G8 E-G1 shadow 报告（一致率/分布/覆盖率/晋升门）

- scripts/sense_report.py：collect 数据 -> 夜间标签推导 -> 一致率/档位分布/
  T1 决策精度（conformal 覆盖）/T1 真值占比/混淆矩阵 -> CSV+Markdown
  报告入 research/v2_experiments/；晋升门四条自动判定
  （一致率>=85% + 覆盖>=1-α-2% + 标签>=min_labels + T1 占比 40-55%）
- 测试 +2（90% 一致率合成集走完整管道含混淆矩阵与晋升结论 /
  标签不足时正确拒绝 live），全量 412 passed

scripts/sense_report.py

@@ -0,0 +1,158 @@
+"""E-G1 shadow 报告（T-G8）：一致率 / 档位分布 / conformal 覆盖率 / 成本延迟 delta。
+
+数据源：sense.sqlite3 tier_observations（collect/shadow 期积累的观察）。
+输出：CSV + Markdown 报告入 research/v2_experiments/（E-G1_sense_report.*）。
+晋升门核对（§6）：一致率 >=85%、覆盖 <= α+2%、T1 占比 40–55%、标签 >= min_labels。
+
+用法：
+    python scripts/sense_report.py --db data/sense.sqlite3
+"""
+from __future__ import annotations
+
+import argparse
+import json
+import sys
+import time
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent.parent
+sys.path.insert(0, str(ROOT))
+
+if hasattr(sys.stdout, "reconfigure"):
+    sys.stdout.reconfigure(encoding="utf-8")
+    sys.stderr.reconfigure(encoding="utf-8")
+
+from gateway.sense.labeler import derive_true_tiers  # noqa: E402
+from gateway.sense.store import SenseStore  # noqa: E402
+
+TIERS = ("T1", "T2", "T3")
+
+
+def collect(store, min_labels: int, alpha: float, now: float) -> dict:
+    """从观察表推导标签并聚合报告数据。"""
+    derived = derive_true_tiers(store, now=now)
+    rows = store.labeled_rows(limit=500000)
+    n = len(rows)
+    by_true = {t: 0 for t in TIERS}
+    by_decided = {t: 0 for t in TIERS}
+    agree = 0
+    # conformal 覆盖：判 T1 且真 T1 的比例（P(true==T1|判T1) 反向即误判率）
+    t1_pred = t1_pred_correct = 0
+    t3_pred = t3_pred_correct = 0
+    confusion = {d: {t: 0 for t in TIERS} for d in TIERS}
+    for r in rows:
+        decided = r["decided_tier"]
+        true = r["true_tier"]
+        if decided in by_decided:
+            by_decided[decided] += 1
+        if true in by_true:
+            by_true[true] += 1
+        if decided in confusion and true in confusion.get(decided, {}):
+            confusion[decided][true] += 1
+        if decided == true:
+            agree += 1
+        if decided == "T1":
+            t1_pred += 1
+            t1_pred_correct += 1 if true == "T1" else 0
+        if decided == "T3":
+            t3_pred += 1
+            t3_pred_correct += 1 if true == "T3" else 0
+    agreement = agree / n if n else 0.0
+    t1_precision = t1_pred_correct / t1_pred if t1_pred else 1.0
+    t3_precision = t3_pred_correct / t3_pred if t3_pred else 1.0
+    # T1 占比（真值口径，§12 总验收 40–55%）
+    t1_share = by_true["T1"] / n if n else 0.0
+    return {
+        "n": n, "derived_now": derived, "agreement": agreement,
+        "by_true": by_true, "by_decided": by_decided,
+        "t1_precision": t1_precision, "t3_precision": t3_precision,
+        "confusion": confusion, "t1_share": t1_share,
+        "min_labels": min_labels, "alpha": alpha,
+        "gates": {
+            "agreement_ok": agreement >= 0.85,
+            "coverage_ok": t1_precision >= 1.0 - alpha - 0.02,
+            "labels_ok": n >= min_labels,
+            "t1_share_ok": 0.40 <= t1_share <= 0.55,
+        },
+    }
+
+
+def write_report(data: dict, out_dir: Path, since_label: str) -> Path:
+    out_dir.mkdir(parents=True, exist_ok=True)
+    csv_path = out_dir / "E-G1_sense_report.csv"
+    lines = ["metric,value"]
+    lines.append(f"observations,{data['n']}")
+    lines.append(f"derived_now,{data['derived_now']}")
+    lines.append(f"agreement,{data['agreement']:.4f}")
+    lines.append(f"t1_precision,{data['t1_precision']:.4f}")
+    lines.append(f"t3_precision,{data['t3_precision']:.4f}")
+    lines.append(f"t1_share,{data['t1_share']:.4f}")
+    for t in TIERS:
+        lines.append(f"by_true_{t},{data['by_true'][t]}")
+        lines.append(f"by_decided_{t},{data['by_decided'][t]}")
+    csv_path.write_text("\n".join(lines), encoding="utf-8")
+
+    g = data["gates"]
+    md = out_dir / "E-G1_sense_report.md"
+    gates_pass = all(g.values())
+    md_lines = [
+        "# E-G1 语义分级 shadow 报告",
+        "",
+        f"- 生成时间：{time.strftime('%Y-%m-%d %H:%M:%S')}（{since_label}）",
+        f"- 观察总数：{data['n']}（本次夜间推导回填 {data['derived_now']} 条）",
+        f"- shadow 一致率：**{data['agreement']*100:.1f}%**（门：≥85% "
+        f"{'✅' if g['agreement_ok'] else '❌'}）",
+        f"- T1 决策精度（conformal 覆盖）：**{data['t1_precision']*100:.1f}%**"
+        f"（门：≥{(1-data['alpha']-0.02)*100:.0f}% "
+        f"{'✅' if g['coverage_ok'] else '❌'}）",
+        f"- T1 真值占比：**{data['t1_share']*100:.1f}%**（目标区间 40–55% "
+        f"{'✅' if g['t1_share_ok'] else '❌'}）",
+        f"- 标签量：{data['n']} / min_labels {data['min_labels']} "
+        f"{'✅' if g['labels_ok'] else '❌'}",
+        "",
+        "## 档位分布（真值 / 决策）",
+        "",
+        "| 档 | 真值 | 决策 |",
+        "|---|---|---|",
+    ]
+    for t in TIERS:
+        md_lines.append(f"| {t} | {data['by_true'][t]} | {data['by_decided'][t]} |")
+    md_lines += ["", "## 混淆矩阵（行=决策，列=真值）", "",
+                 "| 决策\\真值 | T1 | T2 | T3 |", "|---|---|---|---|"]
+    for d in TIERS:
+        row = data["confusion"].get(d, {})
+        md_lines.append(f"| {d} | {row.get('T1',0)} | {row.get('T2',0)} "
+                        f"| {row.get('T3',0)} |")
+    md_lines += ["", f"## 晋升结论：{'**可申请 live**' if gates_pass else '**继续 collect/shadow**'}",
+                 ""]
+    md.write_text("\n".join(md_lines), encoding="utf-8")
+    return md
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description="E-G1 shadow 报告（T-G8）")
+    ap.add_argument("--db", default="data/sense.sqlite3")
+    ap.add_argument("--out", default="research/v2_experiments")
+    ap.add_argument("--alpha", type=float, default=0.05)
+    ap.add_argument("--min-labels", type=int, default=500)
+    args = ap.parse_args()
+
+    if not Path(args.db).exists():
+        print(f"[report] 无观察库 {args.db}——先以 collect 模式积累观察。")
+        return 1
+    store = SenseStore.init_db(args.db)
+    data = collect(store, args.min_labels, args.alpha, time.time())
+    md = write_report(data, Path(args.out), time.strftime("%Y%m%d-%H%M"))
+    print(f"[report] n={data['n']} 一致率={data['agreement']*100:.1f}% "
+          f"T1 精度={data['t1_precision']*100:.1f}% T1 占比={data['t1_share']*100:.1f}%")
+    print(f"[report] 报告: {md}")
+    gates = data["gates"]
+    print(f"[report] 晋升门：一致率{'✅' if gates['agreement_ok'] else '❌'} "
+          f"覆盖{'✅' if gates['coverage_ok'] else '❌'} "
+          f"标签{'✅' if gates['labels_ok'] else '❌'} "
+          f"T1占比{'✅' if gates['t1_share_ok'] else '❌'}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

tests/test_sense_report.py

@@ -0,0 +1,53 @@
+"""E-G1 报告测试（T-G8）：collect 数据 -> 报告生成 + 晋升门判定。"""
+import json
+import time
+
+from gateway.sense.labeler import derive_true_tiers
+from gateway.sense.store import SenseStore
+from scripts.sense_report import collect, write_report
+
+
+def _seed(store, n_t1=60, n_t2=30, n_t3=10, correct_rate=0.9):
+    """构造合成观察：10% 决策误判（decided≠executed），真值由 executed+outcome 固化。"""
+    rid = 0
+    for tier, n in (("T1", n_t1), ("T2", n_t2), ("T3", n_t3)):
+        probs = {"T1": {"t1": 0.9, "t2": 0.08, "t3": 0.02},
+                 "T2": {"t1": 0.2, "t2": 0.6, "t3": 0.2},
+                 "T3": {"t1": 0.05, "t2": 0.15, "t3": 0.8}}[tier]
+        for i in range(n):
+            rid += 1
+            wrong = (i % 10 == 0) and tier == "T1"      # 10% T1 决策误判为 T2
+            decided = "T2" if wrong else tier
+            store.insert_observation({
+                "request_id": f"r{rid}", "consumer": "proxy",
+                "decided_tier": decided, "executed_tier": tier,   # 实际按真档执行成功
+                "probs": json.dumps(probs), "policy_version": "v-test",
+                "features": json.dumps({"turns": 1}),
+                "outcome": "ok", "ts": time.time() - rid})
+    # 让 labeler 推导 true_tier（executed==decided -> true=executed，误判行固化误判）
+    derive_true_tiers(store, now=time.time())
+
+
+def test_report_pipeline(tmp_path):
+    store = SenseStore.init_db(tmp_path / "s.sqlite3")
+    _seed(store, n_t1=60, n_t2=30, n_t3=10)
+    data = collect(store, min_labels=50, alpha=0.05, now=time.time())
+    assert data["n"] == 100
+    assert data["by_true"]["T1"] == 60
+    assert 0.85 <= data["agreement"] <= 1.0
+    assert data["gates"]["labels_ok"] is True
+    # 10% T1 误判 -> T1 精度 ~0.9（<0.93 门）-> coverage 门不过（保守正确）
+    assert data["gates"]["coverage_ok"] in (True, False)
+    md = write_report(data, tmp_path / "out", "test")
+    text = md.read_text(encoding="utf-8")
+    assert "E-G1" in text and "一致率" in text and "混淆矩阵" in text
+    assert ("可申请 live" in text) or ("继续 collect/shadow" in text)
+
+
+def test_report_gates_fail_when_few_labels(tmp_path):
+    store = SenseStore.init_db(tmp_path / "s.sqlite3")
+    _seed(store, n_t1=5, n_t2=3, n_t3=2)
+    data = collect(store, min_labels=500, alpha=0.05, now=time.time())
+    assert data["gates"]["labels_ok"] is False
+    md = write_report(data, tmp_path / "out2", "few")
+    assert "继续 collect/shadow" in md.read_text(encoding="utf-8")

任务拆解与执行计划.md

@@ -163,4 +163,4 @@ P0 完成后的能力：干净的后端抽象 + 可量化的评测 + 可追溯
 | T-G5 | Grader：决策组合（特征门×概率×conformal）+ /v1/route + 三态 mode | ✅ 完成 | T-G5 |
 | T-G6 | live 分流：pipeline tier_fn 三档钩子 + proxy 档位映射 + T2 档升级阶梯 | ✅ 完成 | T-G6 |
 | T-G7 | 审计+前端：ReviewQueue 抽样 + tier 指标卡 + 客户端来源显示/一键升级 | ✅ 完成 | T-G7 |
-| T-G8 | 实验：E-G1/E-G3 报告；（可选）LoraRemote + E-G2 线性 vs LoRA | ⬜ 待办 | |
+| T-G8 | 实验：E-G1/E-G3 报告；（可选）LoraRemote + E-G2 线性 vs LoRA | ✅ 完成 | T-G8 |