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

holdout_6.json

@@ -0,0 +1,29 @@
+{
+  "held_out_attributes": [
+    "午餐偏好",
+    "客户名称",
+    "紧急联系人姓氏",
+    "运动习惯",
+    "邮箱域名",
+    "项目代号"
+  ],
+  "trained_on_attributes": 18,
+  "eval_held_out_queries": 56,
+  "eval_kept_queries": 194,
+  "max_prob_held_out": {
+    "p10": 0.1102,
+    "p50": 0.1771,
+    "p90": 0.1854
+  },
+  "max_prob_kept": {
+    "p10": 0.2125,
+    "p50": 0.3445,
+    "p90": 0.4934
+  },
+  "openset_auc_held_out_vs_kept": 0.9629,
+  "closed_set_false_claim_pct": 100.0,
+  "reject_threshold_from_kept_p05": 0.1752,
+  "held_out_rejected_pct_at_threshold": 44.64,
+  "kept_retained_pct_at_threshold": 100.0,
+  "note": "held-out attributes never appear as training labels, so this is the open-set case a real user bank faces"
+}