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

probe_margin_vs_fabrications.py

@@ -0,0 +1,102 @@
+"""Would a margin gate fix the fabrications the model commits, or only echo it?
+
+The model already refuses 13 of the 25 unknown-attribute questions in its own
+words.  A gate is only worth wiring into the runtime if it catches the *other* 12
+-- the ones where the model fabricates a value -- without paying many extra false
+refusals on answerable questions.
+
+So this reports, for the stored run's unknown-attribute cases:
+
+* whether the reply was a refusal (model handled it) or a fabrication (gate needed)
+* the offline margin for that episode
+* the same for answerable episodes, as the false-refusal denominator
+
+and then sweeps the threshold to show refusals_added vs false_refusals_added.
+"""
+
+from __future__ import annotations
+
+import argparse
+import sys
+from collections import defaultdict
+from pathlib import Path
+
+import torch
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+
+from V2_dpskw.eval_scoring import is_refusal
+from V2_dpskw.fit_answerability_threshold import load_episodes, margin_of
+from V2_dpskw.make_answerability_data import FeatureBank, write_set
+from V2_dpskw.rescore_e2e import load_run
+
+
+def load_episodes_sorted(path: Path, per_category: int = 25):
+    """Corpus rows in the order the harness ran them: sorted category, then file order.
+
+    Pairing a stored run against a raw-file-order corpus silently misaligns every
+    row -- repeated queries inside a category hide it -- so the ordering used by
+    ``eval_end_to_end_memory.build_cases`` is reproduced here explicitly.
+    """
+    grouped: dict[str, list[dict]] = defaultdict(list)
+    for episode in load_episodes(path):
+        grouped[str((episode.get("metadata") or {}).get("category", ""))].append(episode)
+    out = []
+    for category in sorted(grouped):
+        out.extend(grouped[category][:per_category])
+    return out
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--bank", type=Path, default=Path(r"H:\Memory\nm_cache\nm_realistic_v2\feature_cache"))
+    parser.add_argument("--corpus", type=Path, default=Path("data/realistic_v2/eval.jsonl"))
+    parser.add_argument("--run", type=Path, required=True)
+    args = parser.parse_args()
+
+    bank = FeatureBank(args.bank)
+    rows = load_run(args.run)
+
+    unknown = []   # (margin, model_refused, query, reply)
+    answerable = []  # (margin, model_answered_correctly)
+    for episode, row in zip(load_episodes_sorted(args.corpus), rows):
+        if str(episode.get("query", "")) != str(row.get("query", "")):
+            raise SystemExit(f"row/corpus mismatch: {episode.get('query')!r} vs {row.get('query')!r}")
+        metadata = episode.get("metadata") or {}
+        candidates = [str(c.get("text", "")) for c in (episode.get("candidates") or []) if isinstance(c, dict)]
+        positives = [candidates[i] for i in (episode.get("positive_indices") or []) if 0 <= int(i) < len(candidates)]
+        records = write_set(episode, include_answer=True)
+        m = margin_of(bank, str(episode.get("query", "")), records)
+        if m is None:
+            continue
+        reply = row.get("reply", "")
+        if positives:
+            answerable.append((m, bool(row.get("correct"))))
+        else:
+            unknown.append((m, is_refusal(reply), str(episode.get("query", "")), reply))
+
+    refused_by_model = [u for u in unknown if u[1]]
+    fabricated = [u for u in unknown if not u[1]]
+    print(f"unknown-attribute cases with a usable margin: {len(unknown)}")
+    print(f"  model already refused : {len(refused_by_model)}   margins "
+          f"{sorted(round(u[0], 3) for u in refused_by_model)}")
+    print(f"  model FABRICATED      : {len(fabricated)}   margins "
+          f"{sorted(round(u[0], 3) for u in fabricated)}")
+    print("\nfabricated cases:")
+    for m, _r, q, reply in fabricated:
+        print(f"  margin={m:+.3f}  Q={q}  R={reply[:60]}")
+
+    print(f"\n{'thr':>7}{'caught_fabrications':>21}{'reflagged_refusals':>21}{'new_false_refusals':>21}")
+    ans_margins = [m for m, _ok in answerable]
+    for step in range(0, 80, 4):
+        t = step / 1000.0
+        caught = sum(1 for m, _r, _q, _x in fabricated if m < t)
+        reflagged = sum(1 for m, _r, _q, _x in refused_by_model if m < t)
+        new_false = sum(1 for m in ans_margins if m < t)
+        print(f"{t:>7.3f}{caught:>16}/{len(fabricated):<4}{reflagged:>16}/{len(refused_by_model):<4}"
+              f"{new_false:>16}/{len(ans_margins):<4}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())