# Role
You are an Expert Full-Stack Developer specializing in Rust (Axum), modern JavaScript (Riot.js), and AI API Integration.

# Goal
Create a complete "Phase 1" implementation plan and the core file structure for a Real-Time Group Chat Application powered by OpenRouter.

# Project Context
This application is a group chat where multiple humans can talk to each other and a single AI assistant.
*   **Phase 1 (Current Scope):** Humans + 1 System AI in a shared chatroom.
*   **Phase 2 (Future):** Multiple AI Agents added to the chat.
*   **Critical Requirement:** The architecture must be event-driven to support the future addition of autonomous agents.

# Tech Stack Requirements
*   **Backend:** Rust (using **Axum** framework).
*   **AI Engine:** **OpenRouter API** (Support for user-selectable models like Claude 3.7, Llama 3, GPT-4, etc.).
*   **Real-time:** WebSockets (using `tokio-tungstenite` or Axum's built-in `ws` extractor).
*   **Frontend:** Vite with **Riot.js (v9+)**.
*   **Database:** SQLite (managed via `sqlx` for async Rust) - keeping it simple for Phase 1 but strongly typed.
*   **Styling:** **NO FRAMEWORKS.** Use Custom CSS with CSS Variables (Custom Properties) for theming. Leverage Riot.js scoped styles.

# Functional Requirements & Specification

1.  **User System:**
    *   Simple email/password registration (Argon2 hashing).
    *   JWT-based authentication for HTTP and WebSocket upgrades.
    *   Users can "Invite" others via email (generate a unique link).

2.  **Chat Engine & OpenRouter Integration:**
    *   **Room Creation:** When a user starts a chat, they must be able to select which **OpenRouter Model ID** (e.g., `anthropic/claude-3.7-sonnet`, `meta-llama/llama-3-70b-instruct`) drives that specific room.
    *   **Group Context:** The AI must understand it is in a "Room" with multiple people.
    *   **Message Structure:**
        *   `sender_id` (UUID)
        *   `sender_name` (Display Name)
        *   `timestamp` (UTC)
        *   `content` (Markdown string)
        *   `mentions` (Array of User IDs tagged)
    *   **AI Trigger:** When the AI is mentioned or (optionally) on every message, the backend sends the recent chat history to OpenRouter.

3.  **Frontend (Riot.js):**
    *   Use `.riot` components.
    *   Implement a Markdown renderer (e.g., `markdown-it`) with syntax highlighting (e.g., `highlight.js`).
    *   **Model Selector:** A dropdown in the "Create Chat" modal to pick the OpenRouter model.
    *   **Design System:**
        *   Create a `global.css` for root variables (Dark Mode colors, spacing).
        *   Use Riot's `<style>` tag for component-level, scoped CSS.
        *   The design should be professional, clean, and "Dark Mode" by default.

# Deliverables Needed From You

1.  **Directory Structure:** A comprehensive ASCII tree of the project structure (separating `server/` and `client/`).
2.  **Rust `Cargo.toml` dependencies:** The specific crates needed (`reqwest` for OpenRouter API, `axum`, `tokio`, `serde`, `sqlx`, etc.).
3.  **Database Schema:** The SQL migration code for:
    *   `users`
    *   `rooms` (Must include a `model_id` column to store the chosen OpenRouter model string)
    *   `messages`
    *   `invites`
4.  **Core Backend Code:**
    *   The WebSocket handler in Axum.
    *   The **OpenRouter Service Function**: A Rust function that takes `(room_history, model_id)` and sends the request to `https://openrouter.ai/api/v1/chat/completions`.
    *   The `Broadcast` loop that sends AI responses back to the room.
5.  **Core Frontend Code:**
    *   The `vite.config.js` setup for Riot.
    *   The `global.css` file defining the color palette (Professional Dark Mode).
    *   The `chat-room.riot` component demonstrating the message loop, Markdown rendering, and **scoped CSS**.

Please ignore strict error handling for edge cases to keep the code readable, but ensure concurrency safety in the Rust WebSocket manager. Start by outlining the folder structure.