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

inspect_locomo_adversarial.py

@@ -0,0 +1,34 @@
+"""Check how LoCoMo's adversarial questions carry evidence, before trusting the mapping."""
+import json
+import sys
+from collections import Counter
+from pathlib import Path
+
+path = Path(sys.argv[1] if len(sys.argv) > 1 else r"H:\Memory\V2_dpskw\data\net_locomo\locomo10.json")
+data = json.loads(path.read_text(encoding="utf-8"))
+
+with_evidence = Counter()
+without_evidence = Counter()
+keys_seen = Counter()
+examples = []
+for item in data:
+    for qa in item["qa"]:
+        cat = qa.get("category")
+        keys_seen.update(qa.keys())
+        ev = qa.get("evidence") or []
+        if ev:
+            with_evidence[cat] += 1
+        else:
+            without_evidence[cat] += 1
+        if cat == 5 and len(examples) < 4:
+            examples.append(qa)
+
+print("qa field names:", dict(keys_seen))
+print()
+print(f"{'category':>9}{'with evidence':>15}{'without':>10}")
+for cat in sorted(set(with_evidence) | set(without_evidence)):
+    print(f"{cat:>9}{with_evidence[cat]:>15}{without_evidence[cat]:>10}")
+print()
+print("=== adversarial examples ===")
+for qa in examples:
+    print(json.dumps(qa, ensure_ascii=False, indent=1)[:600])