feat: add Cursor Agent CLI as first-class native agent

[jdistler]

Summary

Adds Cursor Agent CLI as a first-class native agent in Happy, backed by the Agent Client Protocol (ACP) over stdio.

Closes #248

Local testing showing cursor cli model selection (cursor pro gives you access to the default Grok model, then all the other frontier lab models as well) -

Screencast-trimmed.mp4

---

Background & Architecture

Cursor CLI (agent / cursor) includes headless ACP support via agent acp (stdio JSON-RPC 2.0). This integration maps Cursor into Happy's multi-agent architecture alongside Claude Code, Codex, Antigravity (agy), and Gemini.

Key Details

1. CLI Execution:
   • happy cursor spawns agent acp under AcpBackend.
   • Daemon remote session spawning correctly invokes happy cursor --happy-starting-mode remote --started-by daemon.
   • POSIX and Windows CLI availability detection checks for agent and cursor in PATH.
2. Protocol & Handshake:
   • Spawns over stdio with ACP protocolVersion: 1.
   • Automatic local credentials re-use from agent login (no extra auth required).
   • Dynamic model discovery from Cursor's session/new response (~130 models, including Grok 4.6, Claude Opus 5, GPT-5.6 Sol, Gemini 3.7 Flash, Composer 2.5).
3. App & Mobile UI:
   • Added Cursor harness to harnessCatalog.ts (HARNESS_NAMES, HARNESS_ORDER).
   • Added Cursor default configurations: agent permission mode (Full tool access), default[] (Auto / Cursor Router).
   • Curated pre-session model catalog in modelModeOptions.ts with fallback handling.
   • Added official Cursor brand assets (icon-cursor.png, @2x, @3x) and avatar badge geometry.

---

Test Coverage

• happy-cli: All 87 test files (820 tests) pass.
• happy-app: All 106 test files (1,119 tests) pass.
• Build: pnpm run build succeeds with clean TypeScript types across the monorepo.
• Live Verification: Verified agent acp initialize & session/new handshake against official Cursor CLI v2026.08.25.

[jdistler]

📝 Curated these to appear right away in model selector, but the true large list loads quickly as they are dynamically discovered (see video demo in PR description)