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

none_class_result.json

@@ -0,0 +1,40 @@
+{
+  "vocabulary_attributes": [
+    "主管姓名",
+    "入职年份",
+    "出生城市",
+    "办公城市",
+    "午餐偏好",
+    "咖啡口味",
+    "团队名称",
+    "客户名称",
+    "宿舍楼号",
+    "工位楼层",
+    "常住城市",
+    "常用编辑器"
+  ],
+  "none_train_attributes": [
+    "手机尾号",
+    "档案标识",
+    "紧急联系人姓氏",
+    "设备型号",
+    "课程名称",
+    "起床时间"
+  ],
+  "none_eval_attributes": [
+    "运动习惯",
+    "通勤方式",
+    "邮箱域名",
+    "阅读工具",
+    "项目代号",
+    "默认语言"
+  ],
+  "train_rows": 890,
+  "eval_vocabulary_queries": 133,
+  "eval_never_seen_queries": 52,
+  "chance_pct": 7.69,
+  "vocabulary_accuracy_pct": 100.0,
+  "never_seen_rejected_pct": 100.0,
+  "known_false_rejection_pct": 0.0,
+  "note": "an explicit NONE class trained on held-out attributes; the never-seen attributes used for testing are disjoint from both the vocabulary and the NONE training set"
+}