Redesign the web workspace; rooms work as a group-chat bot crew

[Robert0622]

Stacked on #46. Redesigns the LoopForge web workspace and turns rooms into a group-chat bot crew, following the Grok-bot model.

Positioning

LoopForge is your AI delivery team:

• Home is the PM. You describe the work. Once the requirement is clear, it suggests who should take it:
  • an SOP: a fixed process with gated steps, for work with a strict, known procedure;
  • the room's bot crew: open-ended work where bots split it up and hand off.
• Rooms are where bots work in public. Ask @LoopForge a question or hand it work. The captain pulls in role bots, each posts its result and hands off to the next, and people can step in at any time.
• SOPs are the safety net. A crew that worked well can be saved as a draft SOP and rerun as a fixed process.

What changed

Home

• The launch card offers "交给 SOP" or "交给群聊 Bot", with the assistant's recommendation and reason (assistantLaunch.mode / modeReason).
• The SOP pickers now list the platform templates (通用模板) in their own group.
• Recent runs collapse a crew's step runs into one entry.

Rooms (web/rooms.js, internal/httpapi/room_crew.go, room_ask.go)

• @LoopForge <question> gets an answer built from the room's history, the quoted message and the status of the room's tasks. Without a model key it still answers status questions and explains what it can do.
• @LoopForge <task> starts a crew mission:
  • The planner uses AI, or falls back to name-weighted rules.
  • The plan is proposed and can be edited (改分工) before 开工.
  • Each step runs as a hidden one-node workflow through the normal dispatchIssue path, so runs, logs and approvals are unchanged.
  • The issue is reopened between steps, because each successful run closes delivery.
• The bot only speaks when spoken to:
  • Replying to a bot, or @某个 Bot, adds a step for that bot.
  • Replying to a mission adds a requirement for the next bot.
  • When a step is stuck, any message unblocks it.
  • Plain chatter is ignored.
• Messages can quote an earlier one: hover, click 回复, and click the quote to jump back. Bot replies show a typing state while they are being written. The process pane (执行过程) shows each step's brief, output and handoff.
• A finished crew can be saved as a draft SOP (sop-<mission>).

Flow designer

• Drag blank canvas to pan (middle button works anywhere, shift+wheel scrolls sideways).
• The canvas leaves room for the floating properties panel, so the last node is reachable.

Everywhere else

• New sidebar hubs: 首页 / 交付 / 群聊 / 构建 / 洞察 / 平台 / 设置.
• A new design layer in web/design.css (tokens --lf-*, light and dark), plus zh/en (web/i18n.js, web/i18n-en.js).

Status

• Passing locally: go vet ./..., go test ./internal/... ./web/... and the API e2e (go test -tags e2e ./e2e).
• Two new unit tests: TestCrewPlanRulesSplitsRolesByName and TestCrewWantsAnswerSplitsQuestionsFromWork.
• Playwright suite is not updated. Locally 25 of 40 tests fail and 4 do not run. They target the old layout: old sidebar entries such as .nav[data-view="runs"], old page controls and the old onboarding tour. Updating them to the new hubs is the first follow-up.

Follow-ups for whoever picks this up

1. Update e2e/tests/*.spec.js to the new hubs and pages (VIEW_TABSETS in web/app.js maps hubs to views).
2. Grok-bot parity still missing:
   • streamed answers, instead of a pending message that is filled in at the end;
   • individual bots answering questions when @-mentioned outside a mission;
   • reactions;
   • image and file understanding in rooms;
   • a periodic room digest.
3. Product polish not yet redone: the flow inspector content, plus the projects, metrics, observability, runners and settings pages.
4. Clean up the old CSS layers (product-v3.css, product-v4.css, platform.css) that design.css currently overrides with !important.

To try it: run in demo mode, open 群聊, and type @LoopForge 调研一下需求列表加负责人筛选的实现方案. The two demo bots (需求分析, 开发实现) run on the mock executor, so no model key is needed.