feat: instant draft-first worktree creation and multi-run launcher redesign (#741)

## Summary

- **Instant worktree creation from chat draft**: selecting "+ New worktree" in the draft branch selector immediately creates a session draft and bootstraps the worktree in the background — no modal interruption
- **Redesigned multi-run launcher**: compact 2-column grid layout in a right-sized dialog with scroll shadow, sticky footer, tooltips replacing verbose descriptions, and project icons in the selector
- **Branch selector aligned across surfaces**: multi-run and agent manager branch pickers now use the shared git store and match NewWorktreeDialog behavior (same default resolution cascade, no synthetic HEAD option, all branches shown)
- **Opaque model multi-select dropdown**: fixes text bleed-through on translucent backgrounds by compositing `--surface-elevated` over `--surface-background`
- **"+ New" inline button in sidebar worktree headers** for faster worktree creation

## Why

Worktree creation was behind modal flow that interrupted the user's train of thought. The draft-first approach lets users start typing immediately while the worktree bootstraps. The multi-run launcher had an oversized form layout with redundant explanations, and its branch picker behaved differently from the main worktree dialog - causing confusion about which branches were available and what the default was.

packages/ui/src/lib/opencode/client.ts

@@ -15,6 +15,7 @@ import type {
 } from "@opencode-ai/sdk/v2";
 import type { PermissionRequest } from "@/types/permission";
 import type { QuestionRequest } from "@/types/question";
+import { waitForWorktreeBootstrap } from "@/lib/worktrees/worktreeBootstrap";
 type StreamEvent<TData> = {
   data: TData;
   event?: string;
@@ -701,6 +702,10 @@ class OpencodeService {
       throw new Error('Message must have at least one part (text or file)');
     }
 
+    if (this.currentDirectory) {
+      await waitForWorktreeBootstrap(this.currentDirectory);
+    }
+
     // Use async prompt endpoint so the client doesn't block waiting
     // for model work (SSE will deliver output/status).
     // This avoids 504s from proxy timeouts on long-running turns.