docs: link design work to issue trackers

[Phoenix500526]

Summary

Several architecture documents describe a shipped foundation alongside
remaining implementation work, but that follow-up work was either not linked
to its existing GitHub issue or existed only in prose. This made progress hard
to query, assign, and connect to child PRs, and made it easy for design and
implementation status to drift through ordinary contributor turnover.

This PR:

• links existing trackers at the exact architecture boundaries they cover;
• creates and links coordination trackers tracking(runtime): harden safe-boundary continuation for production #4324, tracking(workspace): complete Gitoxide workspace continuity #4325, tracking(storage): complete managed dependencies #4326, tracking(bots): harden onboarding after V1 #4327, tracking(permissions): finish macOS Stage 2 #4328,
  tracking(mcp): complete the post-V3 roadmap #4329, tracking(runtime): finish provider-native web search #4330, and tracking(desktop): coordinate Side Conversation follow-ups #4331 for partially implemented routes that had no focused
  issue;
• keeps conditional ideas and explicit non-goals out of the tracker set.

The new issues are coordination surfaces, not requests to deliver each roadmap
as one large PR. Independent authority and failure boundaries should continue
to use focused child issues and PRs.

Verification

• git diff --check
• npm run format:check -- --files-ignore-unknown=true
• npm run check:asf-headers

Typecheck and test suites were not run because this PR changes Markdown links
and tracking notes only.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: OpenAI Codex audited tracker coverage, drafted coordination
issues, and added the documentation links.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — described under Summary above
• No

[Astro-Han]

The direction is right: docs/README.md already says issues and PRs hold plans and progress, and the eight tracking issues you opened are accurate, I read each against the document it points at. Lint, format and ASF headers pass on this head, merge-tree against main is clean. Two things before this merges, and a few smaller ones.

The line added to llm-compaction-events-log-projection-draft.md (and its zh-CN twin) describes durable Tool Result and image rewrites as pending work tracked in #4283. That issue closed as completed today; #4350 and #4348 landed it on main (model-projection-transition-ledger.ts), and the paragraph above the new line still says the placeholder only changes current messages, which is also no longer true. As written the line is false the moment this merges. Drop it, or state that #4350 landed it and refresh the paragraph.

Three places keep the checklist in both the document and the new issue: the mcp draft's five backlog items against #4329, the six under side-conversation.md against #4331, and the five adapters in web-search-provider-capability.md against #4330. The PR's own premise is that progress in prose is unqueryable; adding a pointer without moving the list leaves two authorities, and one has already drifted: the mcp item still lists OAuth metadata, PKCE and dynamic registration as pending, which #2919 and #2920 shipped and #4329 itself ticks, leaving only platform secret storage (tokens sit in credentials.json today). Keep the rationale in the doc, let the issue own the list. That turns this from +63/−7 into a real consolidation, roughly +35/−35, and answers what this change made redundant.

Smaller: #4325 is repeated in eight files in slightly different sentences while gitoxide-repository-admission-capability-v1.zh-CN.md, which the issue names as a design source, has no backlink; one owning document and cross-references would do. Five link phrasings appear (后续路线：, Delivery is tracked in, tracked in [#N], 由 [Issue #N] 跟踪) where the tree already uses 跟踪：[title #N](url); pick that one. And work-board-contract.md still says Status: Phase 0 one line above the new tracker while Phase 1 shipped in #3135.

Before merge, separate from the above: 57 behind main, and bot-onboarding-runtime.zh-CN.md and both compaction documents changed there in the meantime, so a rebase and a re-read of those three.

Evidence boundary: static read of cd1ce44e against main 61224f51; every referenced issue's state checked; claims in the touched paragraphs spot-checked against source; Biome and the ASF header check run locally.

AI-assisted review: drafted with Maka; I verified #4283's state, the mcp OAuth items and the duplicated checklists myself.

简体中文

方向对，八个 tracking issue 我逐个核过，准确。合并前两件事：compaction 文档新加的那行把 #4283 写成待办，而它今天已 closed as completed，#4350/#4348 已落地，合并即为假陈述；三处（mcp、side-conversation、web-search）把清单同时留在文档和 issue 里，两个权威且已漂移（OAuth 三项已由 #2919/#2920 落地），建议文档留理由、清单交给 issue。小项：#4325 在八个文件重复且设计源文档反而没链接；链接措辞五种并存；work-board 状态行仍是 Phase 0。落后 main 57 个 commit，需 rebase。

[Phoenix500526]

The direction is right: docs/README.md already says issues and PRs hold plans and progress, and the eight tracking issues you opened are accurate, I read each against the document it points at. Lint, format and ASF headers pass on this head, merge-tree against main is clean. Two things before this merges, and a few smaller ones.

The line added to llm-compaction-events-log-projection-draft.md (and its zh-CN twin) describes durable Tool Result and image rewrites as pending work tracked in #4283. That issue closed as completed today; #4350 and #4348 landed it on main (model-projection-transition-ledger.ts), and the paragraph above the new line still says the placeholder only changes current messages, which is also no longer true. As written the line is false the moment this merges. Drop it, or state that #4350 landed it and refresh the paragraph.

Three places keep the checklist in both the document and the new issue: the mcp draft's five backlog items against #4329, the six under side-conversation.md against #4331, and the five adapters in web-search-provider-capability.md against #4330. The PR's own premise is that progress in prose is unqueryable; adding a pointer without moving the list leaves two authorities, and one has already drifted: the mcp item still lists OAuth metadata, PKCE and dynamic registration as pending, which #2919 and #2920 shipped and #4329 itself ticks, leaving only platform secret storage (tokens sit in credentials.json today). Keep the rationale in the doc, let the issue own the list. That turns this from +63/−7 into a real consolidation, roughly +35/−35, and answers what this change made redundant.

Smaller: #4325 is repeated in eight files in slightly different sentences while gitoxide-repository-admission-capability-v1.zh-CN.md, which the issue names as a design source, has no backlink; one owning document and cross-references would do. Five link phrasings appear (后续路线：, Delivery is tracked in, tracked in [#N], 由 [Issue #N] 跟踪) where the tree already uses 跟踪：[title #N](url); pick that one. And work-board-contract.md still says Status: Phase 0 one line above the new tracker while Phase 1 shipped in #3135.

Before merge, separate from the above: 57 behind main, and bot-onboarding-runtime.zh-CN.md and both compaction documents changed there in the meantime, so a rebase and a re-read of those three.

Evidence boundary: static read of cd1ce44e against main 61224f51; every referenced issue's state checked; claims in the touched paragraphs spot-checked against source; Biome and the ASF header check run locally.

AI-assisted review: drafted with Maka; I verified #4283's state, the mcp OAuth items and the duplicated checklists myself.

简体中文
方向对，八个 tracking issue 我逐个核过，准确。合并前两件事：compaction 文档新加的那行把 #4283 写成待办，而它今天已 closed as completed，#4350/#4348 已落地，合并即为假陈述；三处（mcp、side-conversation、web-search）把清单同时留在文档和 issue 里，两个权威且已漂移（OAuth 三项已由 #2919/#2920 落地），建议文档留理由、清单交给 issue。小项：#4325 在八个文件重复且设计源文档反而没链接；链接措辞五种并存；work-board 状态行仍是 Phase 0。落后 main 57 个 commit，需 rebase。

Done

[Astro-Han]

Both blocking items from the last round are closed. I checked them against the files at this head rather than against the thread states.

The compaction paragraph no longer points at the closed #4283. What replaced it describes what packages/runtime/src/model-projection-transition-ledger.ts and packages/core/src/model-projection-transition.ts actually do, and "replay, budgeting and compaction" matches the comment on prepareContextBudgetPolicy almost word for word. The table row moved with it, so the stale "placeholder changes only current messages" claim is gone from both language versions.

The three duplicated checklists moved out of the documents. Section 8 of mcp-runtime-architecture-draft.zh-CN.md, Follow-up Work in side-conversation.md, and Follow-up adapters in web-search-provider-capability.md now keep the rationale and hand the list to #4329, #4331 and #4330. I read those three issue bodies against the removed lists: nothing was dropped in the move, and the stale MCP OAuth items are ticked in #4329 rather than repeated. #4325 now appears once, with gitoxide-repository-admission-capability-v1.zh-CN.md carrying a cross-reference to the owning design instead of a ninth copy. Link phrasing is one form throughout. work-board-contract.md states the Phase 1 status.

Links: all 16 issues and 2 PRs referenced resolve, and every tracker target is open (#615, #1625, #2560, #2596, #3731, #3909, #4311, #4319, #4324, #4325, #4326, #4327, #4328, #4329, #4330, #4331). #3135 is merged, which the Work Board files on main confirm. Biome format, Biome lint and the ASF header audit pass on this head. The branch is 26 behind main, but merge-tree is clean and the five touched documents that also changed on main changed in sections this PR does not touch, so a rebase is not needed on my account.

Approving. One item is worth a commit before merge and three are optional, all inline.

Evidence boundary: static read of 9b408dce against main e1e108a7; every referenced issue and PR state checked; the rewritten compaction claims checked against the projection-transition sources; Biome and the ASF header audit run locally. I did not run the test suites, and nothing here needed them.

AI-assisted review: drafted with Maka; I verified the issue states, the reducer call sites, and the three moved checklists myself.

简体中文

上一轮的两个阻塞项都关掉了，我是对着这个 head 的文件核的，没看 thread 状态。

compaction 那段不再指向已关闭的 #4283，新写法跟 model-projection-transition-ledger.ts 和 model-projection-transition.ts 对得上，「replay、budgeting、compaction」跟 prepareContextBudgetPolicy 上的注释几乎一致；表格那一行也一起改了，中英两版都不再有旧说法。

三处重复清单都搬走了，文档留理由、清单交给 #4329、#4331、#4330。我把三个 issue 正文跟被删的清单逐条比过，没有丢东西，MCP 的 OAuth 三项在 issue 里是已勾选状态。#4325 现在只出现一次，gitoxide 那篇改成指向设计文档。链接措辞统一了，work-board 的状态行也更新了。

链接核验：引用到的 16 个 issue 和 2 个 PR 全部有效，tracker 目标都是 open；#3135 已合并，main 上的 Work Board 代码可以印证。format、lint、ASF header 都过。落后 main 26 个 commit，但 merge-tree 干净，另外 5 个在 main 上也变过的文档改的是本 PR 没碰的段落，所以不用为我 rebase。

给通过。一条建议合并前顺手改掉，另外三条可选，都写在行内了。