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_record_blend.py

@@ -0,0 +1,101 @@
+"""Dose-response for the record-ordering blend: router score vs packaged retriever.
+
+``_score_semantic_memory_records`` normally orders every record that carries a
+``semantic_key`` with the packaged ``text_retriever`` alone.  Measured on identical frozen
+keys over the zero-overlap eval (250 answerable episodes, 24 same-shape candidates,
+chance 4.17%) the two scorers differ sharply at Top-1:
+
+    packaged text_retriever   23.20%
+    frozen-key cosine         30.40%
+    trained router (REPLAY)   59.60%
+
+This probe runs the 16 router-critical end-to-end cases at several blend weights
+(``memory_config.memory_record_router_blend``: 0.0 = historical behaviour, 1.0 = router
+only) and reports answer accuracy and wrong-attribute rate for each, so the weight is
+chosen from a measured curve instead of a guess.  The default stays 0.0.
+
+Usage::
+
+    python -m V2_dpskw.probe_record_blend --package qwen3_5_4b_natural_memory_v2
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+from pathlib import Path
+
+import torch
+
+from .eval_router_critical_e2e import build_cases, run_cases
+from .qwen_integration import load_memory_config, load_qwen_dynamic, load_tokenizer
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--package", default="qwen3_5_4b_natural_memory_v2")
+    parser.add_argument("--cases", type=int, default=16)
+    parser.add_argument("--blends", default="0.0,0.25,0.5,0.75,1.0")
+    parser.add_argument("--router", default="checkpoints/router_replay_v7_v2_128/memory_router_v2.pt")
+    parser.add_argument("--output", default="record_blend_dose_response.json")
+    parser.add_argument("--markdown", default="record_blend_dose_response.md")
+    args = parser.parse_args()
+
+    blends = [float(value) for value in args.blends.split(",") if value.strip()]
+    cases = build_cases(args.cases, 20260911)
+
+    model_path = Path(args.package)
+    memory_config = load_memory_config(model_path)
+    model = load_qwen_dynamic(model_path, memory_config=memory_config, load_in_4bit=True,
+                              max_memory={0: "10.5GiB", "cpu": "48GiB"})
+    model.eval()
+    tokenizer = load_tokenizer(model_path)
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    router_path = Path(args.router)
+    if router_path.exists():
+        state = torch.load(router_path, map_location="cpu", weights_only=True)
+        model.memory_router_v2.load_state_dict(state.get("router_state_dict", state), strict=True)
+        model.memory_router_v2.to(device).eval()
+        print(json.dumps({"router": str(router_path)}), flush=True)
+
+    results = {}
+    for blend in blends:
+        model.memory_config.memory_record_router_blend = blend
+        block = run_cases(model, tokenizer, cases, device, max_new_tokens=48,
+                          label="blend=%.2f" % blend)
+        summary = block["summary"]
+        results["%.2f" % blend] = summary
+        print(json.dumps({"blend": blend, **summary}, ensure_ascii=False), flush=True)
+
+    report = {
+        "package": str(model_path),
+        "router": str(router_path),
+        "cases": len(cases),
+        "note": ("blend 0.00 is the historical behaviour (packaged retriever only); "
+                 "1.00 is the trained router only"),
+        "results": results,
+    }
+    Path(args.output).write_text(json.dumps(report, ensure_ascii=False, indent=2), encoding="utf-8")
+
+    lines = ["# 记录排序混合权重 · 端到端剂量曲线", "",
+             f"用例 {len(cases)} 条（零重叠改写，24 条同形候选）。"
+             "`blend=0.00` 为历史行为（仅打包 retriever），`1.00` 为仅训练路由器。", "",
+             "| blend | 回答正确率 | 答成别的属性 | 触发读取 | 平均选中记录 | 决策来源 |",
+             "|---|---:|---:|---:|---:|---|"]
+    for label, summary in results.items():
+        lines.append(
+            "| {0} | {1:.2f}% | {2:.2f}% | {3:.2f}% | {4:.4f} | {5} |".format(
+                label, summary["accuracy_pct"], summary["wrong_attribute_pct"],
+                summary["read_pct"], summary["mean_records_selected"],
+                ", ".join("%s:%d" % (k, v) for k, v in sorted(summary["stop_reasons"].items())),
+            )
+        )
+    text = "\n".join(lines) + "\n"
+    Path(args.markdown).write_text(text, encoding="utf-8")
+    print(text, flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())