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

dump_e2e_category.py

@@ -0,0 +1,77 @@
+"""Dump one category of a stored run next to its corpus case.
+
+Shows the anchor, the positive (answer) fact, the distractor facts actually
+written into the bank, and the model's verbatim reply -- the minimum needed to
+tell "old value won" apart from "wrong attribute was injected".
+"""
+
+from __future__ import annotations
+
+import json
+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_run
+
+
+def load_corpus_full(path: Path):
+    """Same ordering as the harness: sorted category, file order inside a category."""
+    rows = []
+    with path.open("r", encoding="utf-8") as handle:
+        for line in handle:
+            line = line.strip()
+            if not line:
+                continue
+            row = json.loads(line)
+            meta = row.get("metadata") or {}
+            candidates = [str(c.get("text", "")) for c in (row.get("candidates") or []) if isinstance(c, dict)]
+            positives = [candidates[i] for i in (row.get("positive_indices") or []) if 0 <= int(i) < len(candidates)]
+            rows.append({
+                "category": str(meta.get("category", "")),
+                "query": str(row.get("query", "")),
+                "acceptable": [str(v) for v in (meta.get("acceptable") or []) if str(v).strip()],
+                "positives": positives,
+                "candidates": candidates,
+                "meta": meta,
+            })
+    grouped: dict[str, list[dict]] = defaultdict(list)
+    for row in rows:
+        grouped[row["category"]].append(row)
+    ordered: list[dict] = []
+    for category in sorted(grouped):
+        ordered.extend(grouped[category])
+    return ordered
+
+
+def main(corpus, run, category, limit=8, per_category=25):
+    cases = load_corpus_full(Path(corpus))
+    rows = load_run(Path(run))
+    picked = 0
+    idx_used = 0
+    seen = defaultdict(int)
+    for case, row in zip(cases, rows):
+        c = case["category"]
+        if c != category:
+            continue
+        idx_used += 1
+        picked += 1
+        if idx_used > limit:
+            break
+        print(f"--- #{idx_used} Q: {case['query']}")
+        print(f"    anchor  : {case['acceptable']}")
+        print(f"    positive: {case['positives']}")
+        print(f"    written : {row.get('written')}  correct={row.get('correct')} matched={row.get('matched')}")
+        print(f"    reply   : {row.get('reply')!r}")
+        others = [f for f in case["candidates"] if f not in case["positives"]]
+        for o in others[:4]:
+            print(f"      other : {o}")
+        extra = {k: v for k, v in case["meta"].items() if k not in {"category", "acceptable"}}
+        if extra:
+            print(f"    meta    : {json.dumps(extra, ensure_ascii=False)[:300]}")
+
+
+if __name__ == "__main__":
+    main(sys.argv[1], sys.argv[2], sys.argv[3])