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

data/net_locomo/manifest.json

@@ -0,0 +1,25 @@
+{
+  "source": "V2_dpskw\\data\\net_locomo\\locomo10.json",
+  "dataset": "LoCoMo (snap-research/locomo) locomo10.json",
+  "provenance": "real published benchmark; dialogue, questions and answers are the dataset's own",
+  "conversations": 10,
+  "written": 195,
+  "pool_cap": 16,
+  "per_category_cap": 40,
+  "seed": 20260915,
+  "by_category": {
+    "multi_hop": 40,
+    "temporal": 39,
+    "open_domain": 36,
+    "single_hop": 40,
+    "adversarial": 40
+  },
+  "skipped": {
+    "no_evidence_in_index": 5
+  },
+  "caveats": [
+    "candidate pool is bounded per question, so this is not a full 5,882-turn haystack run",
+    "category adversarial is scored as 'must refuse': the dataset provides a plausible wrong answer",
+    "answers phrased differently from the benchmark string fail exact-anchor matching; metadata.answer_tokens supports a paraphrase-tolerant secondary metric"
+  ]
+}