In order to give AI agents more flexible and powerful control over the
tldraw canvas, this PR replaces the old server-side shape CRUD tools
(create_shapes, update_shapes, delete_shapes) with an exec-based
architecture where AI writes JavaScript that runs directly in the
browser widget against a focused editor proxy.

Reimplemented from
[max/mcp-code-mode](https://github.com/tldraw/tldraw/tree/max/mcp-code-mode)
with a clean commit history.

Relates to #6832

**Key architectural changes:**

- **Exec-based model**: AI writes JavaScript code that executes in the
browser via a sandboxed `executeCode()` function, with access to the
full Editor API through a focused editor proxy
- **Focused editor proxy**: ES Proxy that transparently converts between
AI-friendly focused shapes (simple string IDs, named colors, pixel font
sizes) and tldraw's internal representation
- **Search tool**: AI can query the Editor API spec (extracted at build
time from TypeScript declarations) to discover methods, shape types, and
helpers
- **Server-widget callback bridge**: `PendingRequests` system enables
async request-response between the Cloudflare Worker and browser widget
- **Build-time API extraction**: TypeScript compiler API script reflects
on tldraw packages to generate searchable API metadata and method
classification maps

### Change type

- [x] `improvement`

### Test plan

1. Deploy the MCP app and connect from Claude Desktop
2. Ask AI to create shapes — should use exec tool with focused editor
proxy
3. Ask AI to search for Editor methods — should use search tool
4. Verify checkpoints persist across reconnections
5. Test in Cursor, VS Code, and ChatGPT hosts

### Release notes

- Migrate MCP app from shape CRUD tools to exec-based architecture for
more flexible AI canvas control

### Code changes

| Section        | LOC change       |
| -------------- | ---------------- |
| Apps           | +4504 / -2375    |
| Documentation  | +23 / -39        |
| Config/tooling | +9 / -134        |