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

tools_cmp_rv2.py

@@ -0,0 +1,121 @@
+"""Per-category + per-case comparison of realistic_v2 end-to-end eval JSONs.
+
+Handles two on-disk shapes:
+  {"<router>": {"summary": {...}, "rows": [...]}}   (current)
+  {"<router>": [ ...rows... ]}                       (older)
+"""
+import json
+import sys
+from collections import defaultdict
+from pathlib import Path
+
+ROOT = Path(r"H:\Memory\V2_dpskw")
+
+
+def load(path):
+    data = json.loads(Path(path).read_text(encoding="utf-8"))
+    key = next(iter(data))
+    body = data[key]
+    if isinstance(body, list):
+        rows = body
+        summary = None
+    else:
+        rows = body.get("rows", [])
+        summary = body.get("summary")
+    return rows, summary
+
+
+def per_cat(rows):
+    agg = defaultdict(lambda: {"n": 0, "ok": 0, "abst": 0, "wrong_abst": 0})
+    for r in rows:
+        c = r["category"]
+        a = agg[c]
+        a["n"] += 1
+        if r.get("correct"):
+            a["ok"] += 1
+        if r.get("abstained"):
+            a["abst"] += 1
+        if r.get("wrongly_abstained"):
+            a["wrong_abst"] += 1
+    return agg
+
+
+def pct(x, n):
+    return 100.0 * x / n if n else 0.0
+
+
+def key_of(r, idx):
+    """Stable identity for a case row: category + query + index within category."""
+    return (r["category"], r["query"], idx)
+
+
+def main(paths):
+    loaded = []
+    for p in paths:
+        rows, summary = load(p)
+        loaded.append((p, rows, summary))
+
+    cats = sorted({r["category"] for _, rows, _ in loaded for r in rows})
+    agg = [per_cat(rows) for _, rows, _ in loaded]
+
+    head = f"{'category':<20}" + "".join(f"{Path(p).stem[:20]:>22}" for p, _, _ in loaded)
+    print(head)
+    for c in cats:
+        line = f"{c:<20}"
+        for a in agg:
+            n = a[c]["n"]
+            line += f"{a[c]['ok']:>10}/{n:<3}{pct(a[c]['ok'], n):>7.2f}%"
+        print(line)
+    line = f"{'TOTAL':<20}"
+    for a in agg:
+        n = sum(v["n"] for v in a.values())
+        ok = sum(v["ok"] for v in a.values())
+        line += f"{ok:>10}/{n:<3}{pct(ok, n):>7.2f}%"
+    print(line)
+
+    # per-case flip analysis between baseline (first) and candidate (last)
+    base_rows = loaded[0][1]
+    cand_rows = loaded[-1][1]
+    if len(base_rows) == len(cand_rows):
+        idx = defaultdict(int)
+        bk = []
+        for r in base_rows:
+            c = r["category"]
+            bk.append(key_of(r, idx[c]))
+            idx[c] += 1
+        idx2 = defaultdict(int)
+        ck = []
+        for r in cand_rows:
+            c = r["category"]
+            ck.append(key_of(r, idx2[c]))
+            idx2[c] += 1
+        bmap = {k: r for k, r in zip(bk, base_rows)}
+        cmap = {k: r for k, r in zip(ck, cand_rows)}
+        only_b = [k for k in bmap if k not in cmap]
+        only_c = [k for k in cmap if k not in bmap]
+        print(f"\nkeys only in baseline: {len(only_b)}  only in candidate: {len(only_c)}")
+        fixes, breaks = [], []
+        for k in bmap:
+            if k not in cmap:
+                continue
+            b, c = bmap[k], cmap[k]
+            if not b.get("correct") and c.get("correct"):
+                fixes.append((k, b, c))
+            elif b.get("correct") and not c.get("correct"):
+                breaks.append((k, b, c))
+        fc = defaultdict(int)
+        for k, _, _ in fixes:
+            fc[k[0]] += 1
+        bc = defaultdict(int)
+        for k, _, _ in breaks:
+            bc[k[0]] += 1
+        print(f"FIXED  (wrong->right): {len(fixes)}  {dict(fc)}")
+        print(f"BROKEN (right->wrong): {len(breaks)}  {dict(bc)}")
+        for k, b, c in fixes:
+            print(f"  [+] {k[0]:<18} {k[1][:34]:<34} {b.get('reply','')[:52]!r} -> {c.get('reply','')[:52]!r}")
+        for k, b, c in breaks:
+            print(f"  [-] {k[0]:<18} {k[1][:34]:<34} {b.get('reply','')[:52]!r} -> {c.get('reply','')[:52]!r}")
+
+
+if __name__ == "__main__":
+    main(sys.argv[1:])