docs(ai-autopilot): bootstrap capstone example — the whole flow offline (#124)

[suleimansh]

Part of #124. Delivers the offline capstone example; the live-verification half stays infra-gated (see below).

A new private examples/bootstrap-quickstart workspace that runs the entire AI-framework epic in one deterministic, no-API-key flow:

detect framework (preset)
  → Bootstrap: scope → architect → build → full-fledged loop → deploy
  → scale mode: CODE-OVERVIEW.md

AiFake scripts the model and FakeRunner is an in-memory sandbox, so it's fully offline and deterministic. Run it:

pnpm --filter @gemstack/example-bootstrap-quickstart start

What it exercises, composed for real (not stubbed orchestration):

• Presets (Web-app preset: Next.js + Vike targeting (persona/prompt/loop layer) #115) — Vike detected from the project deps; the build workers are that framework's personas + the shared neutral core.
• Bootstrap (Bootstrap mode — zero to a production-grade app (scope prompt + full-fledged loop + narration) #116) — one scope question, an architect that records choices to the decisions ledger, a build scaffolding files with persona workers inside a runner, the full-fledged loop (checklist blocks once on "no auth", clears after the fix → production-grade in 2 passes), and a deploy decision behind the DeployTarget seam.
• Surfaces (ai-autopilot: surfaces (terminal / in-page / background) #100/Generalize the surface stream to be event-type generic #120) — every phase streams over launchAutopilot<BootstrapEvent, BootstrapResult>.
• Scale mode (Scale mode — maintain CODE-OVERVIEW.md for large-codebase navigation #114) — CODE-OVERVIEW.md generated from the scaffold.

Verified: the smoke test (2/2) asserts the whole chain, and start prints the full narration end to end (scope → architect w/ 2 recorded decisions → 3 files scaffolded → checklist ✗ then ✓ → SSR/dockploy deploy → CODE-OVERVIEW.md).

Live verification (the infra-gated half of #124): the framework pieces verify offline here; the honest "zero to a running app" proof needs a real model + LocalRunner producing an app end to end. Swapping the two fakes is the only code change. Keeping #124 open to track that.

Private example (not published) — no changeset.