WikiForge 来自 Wiki-LLM 类实践，但它不是把某个私有知识库开源出来。它只保留可复制的小机制：来源隔离、置信度、reference recall、lint、exam、agent rules。
WikiForge is inspired by Wiki-LLM style systems, but it only publishes reusable mechanisms: source separation, confidence labels, reference recall, lint, exam, and agent rules.
核心思路：先把原始资料编译成稳定 wiki，再让 AI Agent 基于 wiki 工作，而不是每次临时捞几段材料拼答案。
The core idea is to compile raw material into stable wiki pages before asking an agent to reason over it.
WikiForge 保留这些结构：
WikiForge keeps:
sources/for raw inputsdomains/for distilled pageswiki/index.mdfor navigationAGENTS.mdfor agent behavior
WikiForge 吸收的是轻量卫生机制，不是重型运行时：
WikiForge absorbs lightweight hygiene patterns, not a heavy runtime:
- durable pages need metadata
wiki/facts.mdneeds source labels- decisions should have durable records
- release requires lint checks
- retrieval changes should have regression questions
默认不引入数据库、向量库、后台 worker 或常驻 agent。
By default, it does not introduce a database, vector store, background worker, or always-on agent.
tools/wiki_radar.py 是一个很小的 reference recall layer。它扫描 references/，为问题返回可能相关的外部参考。
tools/wiki_radar.py is a small reference recall layer over references/.
重要边界：召回结果只是线索，不是事实裁判。
Important boundary: a recall result is a hint, not a fact.
WikiForge 把 capture 和 distillation 分开：捕获可以自动化，蒸馏必须有边界、可审计、可回滚。
WikiForge separates capture from distillation: capture can be automated later; distillation should stay bounded and auditable.
公开模板只提供本地 ingest stub，不包含真实 message-platform bot。
The public template only includes a local ingest stub, not a production message-platform bot.
Agent 行为应该写进显式规则文件和脚本，而不是藏在不可见 hooks 或私有自动化里。
Agent behavior should live in explicit rule files and scripts, not hidden hooks or private automation.