feat: workbench is the base page — top-bar surfaces, warm engine sessions, live mid-run geometry

[devin-ai-integration]

Summary

Three changes that make the Engineer workbench the project page:

1. Base page. /overview is now a pure redirect to /engineer, and every link that used to land on overview (project create, breadcrumb, shared-projects nav, file tree, folder browser, invite accept, unauth next=) goes straight to the workbench. The create-flow AI kickoff (PipelineKickoffListener) moved from the overview page into the stage page, so "prompt → new project → copilot bootstraps it" still fires — now on the workbench:

// [stage]/page.tsx
const brief = await prisma.projectBrief.findUnique({ where: { projectId_branchId: ... } });
<PipelineKickoffListener hasBrief={Boolean(brief?.prompt || brief?.intendedUse)} />

2. Top-bar surfaces instead of tab churn. Single-instance surfaces (Assembly, PCB, Checks, Repository, Ideate, Verify, Launch, Renders) are permanent buttons in the top bar — no open/close dance. Only multi-component surfaces (CAD, Schematic) remain closable tabs via +. Fixed panes mount lazily on first visit and stay mounted (CSS visibility swap), so revisiting is instant.

3. Faster viewport, live agent geometry.

• CadViewport gains a warm-session pool: on clean unmount the connected Zoo WebRTC session (rtc + video element) is parked for 120s keyed on token|baseUrl|headless|scenery; the next viewport with the same options adopts it and only re-executes its script — skipping the multi-second WebRTC handshake that made reopening a CAD tab feel dead. Live viewports never share a session; only released ones are pooled.
• text_to_cad now streams each part into the workspace the moment it generates (per-part mutateModel3dDoc inside runJob) and emits a new saved CAD-progress phase. The copilot provider reacts to saved chunks by invalidating design/engineer queries mid-run, so open viewports render parts while the rest of the batch is still generating — no more waiting for the whole tool to finish. Refetch is debounced (200ms) so bursts coalesce, and the viewport keeps the last render on screen while re-executing.

Behavior notes: old /overview bookmarks redirect; ?view= deep links unchanged; mid-run saves reuse the existing locked read-modify-write so parallel parts can't clobber each other.

Link to Devin session: https://app.devin.ai/sessions/f2de7850ec9640f7b7bdd99b34fc3348
Requested by: @alanshen27

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring