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_pairing.py

@@ -0,0 +1,62 @@
+"""Prove that stored run rows pair correctly with corpus cases.
+
+Pairing is by position (``build_cases`` emits categories in sorted order, file
+order inside a category), so any drift between the corpus as it is *now* and the
+corpus as it was *when the run happened* would silently pair the wrong anchor
+with the wrong reply -- and repeated queries inside a category would hide it.
+
+The stored rows carry ``matched``, the list the legacy scorer produced at run
+time.  Recomputing it from the paired case must reproduce that list exactly;
+that is an independent witness of correct pairing.
+"""
+
+from __future__ import annotations
+
+import sys
+from collections import defaultdict
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from V2_dpskw.rescore_e2e import load_corpus, load_run
+
+
+def legacy_accept(reply: str, case: dict) -> list[str]:
+    lowered = reply.lower()
+    return [v for v in case["acceptable"] if v and v.lower() in lowered][:3]
+
+
+def check(corpus_path: Path, run_path: Path, per_category: int = 25) -> bool:
+    cases = load_corpus(corpus_path, per_category)
+    rows = load_run(run_path)
+    print(f"{run_path.name}: {len(rows)} rows / {len(cases)} cases")
+    if len(rows) != len(cases):
+        print("  !! length mismatch")
+        return False
+    bad = []
+    for i, (row, case) in enumerate(zip(rows, cases)):
+        expect = legacy_accept(row.get("reply", ""), case)
+        got = row.get("matched") or []
+        if expect != got:
+            bad.append((i, case, row, expect, got))
+    if not bad:
+        print("  pairing VERIFIED: every stored 'matched' field reproduces exactly")
+        return True
+    print(f"  !! {len(bad)} rows disagree with the corpus")
+    per_cat = defaultdict(int)
+    for _, case, _, _, _ in bad:
+        per_cat[case["category"]] += 1
+    print(f"  disagreement by category: {dict(per_cat)}")
+    for i, case, row, expect, got in bad[:6]:
+        print(f"   row#{i} cat={case['category']}")
+        print(f"     corpus query={case['query']!r} acceptable={case['acceptable']!r}")
+        print(f"     row    query={row.get('query')!r}")
+        print(f"     expect matched={expect!r} stored={got!r}")
+    return False
+
+
+if __name__ == "__main__":
+    ok = True
+    for run in sys.argv[2:]:
+        ok &= check(Path(sys.argv[1]), Path(run))
+    print("\nALL PAIRINGS OK" if ok else "\nPAIRING PROBLEM")