derived layer 引入只读 spreading activation shadow path

Blind spot (现状盲点)

[`docs/specs/continuum-memory.md`](../specs/continuum-memory.md) 已经把 memory owner 内部重分层成 `learned core + artifact store + derived index`，并且 retrieval 已升级为 owner-side learned-core-guided recall。但当前证明面更偏向"统一 recall signal + artifact anchor"，对**显式多跳联想召回**的机制描述还不够硬。CMA 的 associative routing 探针显示，spreading activation 能在 query 没有词法重叠时，通过 semantic / temporal / structural edges 找到多跳相关记忆；这正好对应我们 derived layer 未来要承担的"关系图谱 / 聚合索引 / 跨 owner readout"能力。 但这条不能直接做成主路径。CMA 的参考实现允许 retrieval 驱动 graph mutation，如果照搬到 VZ，会带来两个风险：(1) 自由 edge mutation 绕过 `MemoryModule` 唯一 owner；(2) 不受控的 activation feedback loop 导致 memory drift。正确姿势是先做**只读 shadow path**，只用于候选重排和 evidence，不写 durable memory。

Adoptable suggestions (可落地动作)

1. 1.在 [`docs/specs/continuum-memory.md`](../specs/continuum-memory.md) 增加 "Read-only Spreading Activation Shadow Path" 小节，明确该路径只属于 derived layer 的 SHADOW 评估，不参与 durable write。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

2. 2.定义最小 edge 类型：semantic、temporal（FOLLOWED_BY）、associative、structural（owner-to-owner readout），每条 edge 必须带 provenance。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

3. 3.实现 damped propagation：activation 只在 bounded hop 内传播；每次 query 输出 activation trace，作为可审计 telemetry。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

4. 4.只允许把 activation 结果用于 retrieval candidate re-ranking；任何新 edge / edge weight 更新都必须经 session-post slow loop + ModificationGate，不能在 turn-time retrieve 中原地写回。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

5. 5.与 CMA-2 的 VZ-MemProbe 绑定：只有 shadow path 在 associative recall / disambiguation probe 上有稳定收益，才考虑进入 ACTIVE。PROPOSED

   Not a runnable A/B candidate — evaluated by the path above, not ablation.

Traceability

No plugins / runs linked yet. Scaffold a suggestion to start.

Expected benefit (预期收益)

- 给 R5/R6 的"联想记忆"增加可解释、多跳、可审计的工程路径，而不是只依赖向量相似度。 - 在不破坏 R8 SSOT 的前提下，探索 CMA 的 strongest mechanism。 - 为 VZ-MemProbe 的 associative recall / contextual disambiguation 两类测试提供可对照实验组。

Cited paper (引用论文)

**D1. Logan J. *Continuum Memory Architectures for Long-Horizon LLM Agents*. arXiv:2601.09913, 2026.** - 文档位置：[`research/openai-frontier-2026/papers/D1_continuum_memory_architectures.pdf`](../../research/openai-frontier-2026/papers/D1_continuum_memory_architectures.pdf) - 深入分析：[`research/openai-frontier-2026/notes/D1_continuum_memory_deep_analysis.md`](../../research/openai-frontier-2026/notes/D1_continuum_memory_deep_analysis.md) - 摘要原文（精炼）： > We define the Continuum Memory Architecture (CMA), a class of systems that maintain and update internal state across interactions through persistent storage, selective retention, associative routing, temporal chaining, and consolidation into higher-order abstractions. ... Preliminary evaluation across four behavioral probes (knowledge updates, temporal association, associative recall, contextual disambiguation) provides initial evidence that CMA-class behaviors yield advantages on tasks that stress memory dynamics. - 关键观点：CMA 的 spreading activation / associative routing 对多跳召回有直接价值；但 VZ 不能照搬 free-form graph mutation，必须先做 read-only shadow path。 ---