docs: 调研 70 个 Pi 社区候选并记录采用取舍

[tt-a1i]

Problem

Pi 社区有大量包和扩展，但目录收录、README 功能、固定源码、已发布 npm 产物及 OpenPI 运行兼容性容易混在一起；已有研究也需要与当前实现和公开跟踪项对照。

Value

保留 70 个包或扩展候选的第一手资料、版本身份和采用取舍，让后续实施能从具体源码与实验入口继续。Worktree 数据保全与 Web 选区反馈已分别提交到 #472、#473；不确定价值的方向进入 Discussions #474–#478。

Approach

• 保存 74 条方向研究记录，合并同名包/集合后为 70 个候选；49 个候选检查了具体实现源码，其余核对 README/manifest。
• 区分来源核实、静态机制、隔离探针、运行兼容与采用决策；补充 2026-08-09 旧研究，保留历史结论。
• 保存脱除临时机器路径的 Worktree 生产函数探针回执；研究记录与 Issue/Discussion 双向关联。
• 现有 LSP、DAP、MCP、历史搜索、Cache、Memory、后台投递与 ACP 方向补充证据，不另建重复需求。

Validation

• bun run check 通过。
• bun run test 通过：Node tests 1461 passed / 1 skipped / 0 failed；Web Vitest 113 passed。
• 新 JSON 可解析、文档相对链接存在、git diff --check 通过；公开内容检查未包含私人 Session、凭据或本机个人路径。
• 独立复核检查关键公开结论与去重；合成 Git fixture 直接调用 OpenPI reclaimWorktree 复现两种 index flags 下的数据丢失。
• 未安装或执行第三方插件，未调用模型，未做插件兼容认证或性能 Benchmark。仓库测试通过不代表社区方案已被验证采用。

Impact

本 PR 归档研究与公开跟踪回执。实现由对应 Issue 跟进；运行时、模型工具、用户配置、依赖和安装行为没有变更。

[tt-a1i]

Reviewed exact head a5fcbef.

This is valuable governed research: the 74 observations / 70 distinct candidates / 49 source-inspected candidates and 17 publication receipts reconcile, limitations remain explicit, and I found no P0/P1 or sensitive local evidence. It is not merge-ready while draft and conflicting, and the following P2 evidence-quality gaps should be resolved:

• Normalize the evidence schema. sourceURLs, sourceUrls, and sources are mixed, along with sha/commit and parallel recommendation/unknown fields; a straightforward validator can treat 18 observations as source-less.
• Preserve exact published-version identity. Twelve npm-backed candidates have only mutable registry-root links without structured version/tarball integrity; @ollama/pi-web-search similarly buries 0.0.5 and its shasum in prose.
• Make tracker backlinks durable. Current Issue/Discussion links target blob/codex/pi-ecosystem-research-20260908 and can break if the branch moves or is deleted.
• Refresh the worktree conclusion: #472 was fixed by merged #487, but the record still presents it as current action and does not link the implementation PR.
• Preserve a retrievable executable harness for the production-fixture claim; a hash of unpublished bytes identifies content but does not make it reproducible.

After refreshing provenance, resolving the README conflict, and making these identities/backlinks reproducible, this docs-only contribution is worth merging.