Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据

- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线
- 修复 prepare_memory_router_dataset 候选池重建缺陷（mega 家族 3568x 加速，输出逐字节相同）
- 修复 v5 被破坏的拒答与多跳标签（train 未知样本 319 -> 16319，multi_hop 平均正例 1.00 -> 2.00）
- 同存储预算下 V2-128 v6 逐轴 22/22 通过：Top-1 41.12% -> 94.62%，未知拒答 0.00% -> 100.00%
- 记录三条被实测推翻的显然优化（logits_to_keep=1 反而慢 55%、XL 容量未带来收益）
- 记忆手术跨架构可移植性 14/14，读写关闭时与原生模型逐位相同

verify_threshold_sweep.py

@@ -0,0 +1,111 @@
+"""Verify dominance on the policy axes across the *whole* threshold sweep, not just 0.50.
+
+The scorecard's headline verdict compares the refusal axes at the single default
+read threshold (0.50).  A candidate that merely happens to sit on the right side of one
+threshold could still be worse at 0.30 or 0.80.  This script checks every threshold in
+the recorded sweep, so "not worse than the baseline on 未知拒答率 / 已知问题被误拒率 /
+未知问题被误读率" is established across the curve rather than at one point.
+
+Usage::
+
+    python -m V2_dpskw.verify_threshold_sweep --scorecard router_scorecard_final.json ^
+        --baseline "V2-128 deployed(v3)" --candidate "REPLAY-128 v7 final"
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+#: (metric key, label, direction) -- "max" means higher is better.
+AXES = [
+    ("need_f1", "need F1", "max"),
+    ("need_recall", "need 召回", "max"),
+    ("need_precision", "need 精确率", "max"),
+    ("specificity_unknown_refusal", "未知拒答率", "max"),
+    ("known_question_refusal_rate", "已知问题被误拒率", "min"),
+    ("unknown_question_read_rate", "未知问题被误读率", "min"),
+    ("abstention_accuracy", "仲裁准确率", "max"),
+]
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--scorecard", default="router_scorecard_final.json")
+    parser.add_argument("--baseline", default="V2-128 deployed(v3)")
+    parser.add_argument("--candidate", action="append", required=True)
+    parser.add_argument("--output", default="threshold_sweep_check.json")
+    parser.add_argument("--markdown", default="threshold_sweep_check.md")
+    args = parser.parse_args()
+
+    data = json.loads(Path(args.scorecard).read_text(encoding="utf-8"))
+    if args.baseline not in data:
+        raise SystemExit(f"baseline {args.baseline!r} not in scorecard: {list(data)}")
+    thresholds = sorted(k for k in data[args.baseline]["metrics"] if k.startswith("thr"))
+
+    lines = ["# 拒答策略轴 · 全门槛曲线验证", "",
+             f"基线：`{args.baseline}`。门槛取值：{', '.join(t[3:] for t in thresholds)}。", "",
+             "全部为百分比。`通过` 表示该候选在每个门槛上都不劣于基线。", ""]
+    report: dict = {"baseline": args.baseline, "thresholds": thresholds, "candidates": {}}
+
+    for candidate in args.candidate:
+        if candidate not in data:
+            raise SystemExit(f"candidate {candidate!r} not in scorecard")
+        rows = []
+        all_pass = True
+        for key, label, direction in AXES:
+            cells = []
+            axis_pass = True
+            for threshold in thresholds:
+                base_value = data[args.baseline]["metrics"][threshold].get(key)
+                cand_value = data[candidate]["metrics"][threshold].get(key)
+                if base_value is None or cand_value is None:
+                    cells.append("-")
+                    continue
+                if direction == "max":
+                    ok = cand_value >= base_value - 1e-12
+                    delta = cand_value - base_value
+                else:
+                    ok = cand_value <= base_value + 1e-12
+                    delta = cand_value - base_value
+                axis_pass = axis_pass and ok
+                cells.append(f"{100 * cand_value:.2f}%" + ("" if ok else " ✗"))
+            all_pass = all_pass and axis_pass
+            rows.append({"axis": label, "direction": direction,
+                         "values_pct": cells, "axis_passed": axis_pass})
+        report["candidates"][candidate] = {"rows": rows, "sweep_dominates": all_pass}
+
+    lines.append("| 轴 | " + " | ".join(f"门槛 {t[3:]}" for t in thresholds) + " | 全门槛通过 |")
+    lines.append("|---" * (len(thresholds) + 2) + "|")
+    for candidate in args.candidate:
+        body = report["candidates"][candidate]
+        for row in body["rows"]:
+            lines.append(
+                f"| {candidate} · {row['axis']} | " + " | ".join(row["values_pct"])
+                + f" | {'是' if row['axis_passed'] else '**否**'} |"
+            )
+    lines.append("")
+    lines.append("## 基线在各门槛上的值（作对照）")
+    lines.append("")
+    lines.append("| 轴 | " + " | ".join(f"门槛 {t[3:]}" for t in thresholds) + " |")
+    lines.append("|---" * (len(thresholds) + 1) + "|")
+    for key, label, _ in AXES:
+        cells = []
+        for threshold in thresholds:
+            value = data[args.baseline]["metrics"][threshold].get(key)
+            cells.append("-" if value is None else f"{100 * value:.2f}%")
+        lines.append(f"| {args.baseline} · {label} | " + " | ".join(cells) + " |")
+
+    text = "\n".join(lines) + "\n"
+    Path(args.markdown).write_text(text, encoding="utf-8")
+    Path(args.output).write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
+    print(text)
+    for candidate in args.candidate:
+        verdict = "通过" if report["candidates"][candidate]["sweep_dominates"] else "未通过"
+        print(f"[{candidate}] 全门槛 7 轴判定: {verdict}")
+    return 0 if all(report["candidates"][c]["sweep_dominates"] for c in args.candidate) else 2
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())