feat(loop): explicit mitigation for retrieval failure modes

[raymondginger2018-sudo]

Summary

Adds explicit handling for three failure modes an agent memory loop must face:

1. Retrieved nothing — no similar entry in store → similarity threshold + explicit no-memory fallback (never best-effort guessing)
2. Retrieved the wrong thing — pure semantic recall misses identifiers/API names → hybrid keyword + vector recall, threshold enforcement
3. Retrieved but not used — chunks in history but never injected → assert_all_injected() self-check

Design

• Pure mechanism, no LLM
• Works with any store that returns scored entries {similarity, content, ...}
• Uses the data-boundary wrapper from core.loop.injection_regression (PR feat(loop): prompt-injection regression suite (pure mechanism) #216) for the no-memory fallback, keeping untrusted memory content isolated

File

• core/loop/memory_retrieval.py (new, 135 lines)

Part of GenAI lesson 15 memory-loop quality family.

[raymondginger2018-sudo]

设计说明

问题：RAG/记忆检索存在三类典型失败：相似度阈值过严（漏召回）、过松（噪声召回）、排序不稳定（同一查询多次结果不同）。失败时 agent 通常静默降级或直接报错，没有显式处理。

解法：memory_retrieval.py——定义默认相似度阈值 0.45，提供显式的失败分类和降级策略。

关键设计决策：

• 显式失败模式：把"静默错误"变成"可观察的失败类型"，调用方可以根据类型决定重试/放宽/放弃
• 复用 PR feat(loop): prompt-injection regression suite (pure mechanism) #216 的 render_data_block() 来编码检索结果——检索内容是不可信数据，必须走同一套注入防护编码
• 默认阈值 0.45 来自经验值，可通过参数覆盖

依赖说明：本文件 import 了 core.loop.injection_regression.render_data_block（即 PR #216）。建议 #216 先合入，本 PR 随后。如果维护者希望独立合入，我可以移除该依赖并内联一个最小版本。

测试建议：对三类失败模式各构造一组检索结果，验证分类正确

[Zongwei9888]

Thank you for the submission, @raymondginger2018-sudo. Closing because the module fails at import: it depends on core.loop.injection_regression, which is not in the repository, and it assumes a similarity-scored retrieval store that DeepCode does not have — memory here is a file-based pointer index in core/harness/memory.py with on-demand topic reads (fetch_memory_topic, merged from your #233). The three failure modes you describe are real, but they need to be addressed in that reader; a change there with tests would be welcome.