# Task T2 — Game-ify in 3D (v2)

You are given a small vanilla-JS canvas project: a self-driving-car neural-network
demo (2D top-down road, sensor rays, generational learning). Your job is to turn it
into a **complete, polished, playable arcade game in 3D** — something a person would
genuinely enjoy playing for ten minutes.

This is the same brief as the 2D version of this task, with one difference: the game
must be **truly 3D** — a perspective camera on a 3D road with 3D vehicles (chase cam,
cockpit, cinematic — your call, justify it). Not a 2D game with tilted graphics.

## Requirements

- **Player-controlled car** (keyboard). The game, not the simulation, is the product.
- **Score system** — meaningful scoring the player understands at a glance.
- **Progression** — levels, waves, or escalating difficulty; the game must ramp up.
- **Full game loop** — start screen, playing state, game over, restart. No dead ends.
- **Polish** — visual and game-feel polish is explicitly part of the task: lighting,
  motion, camera work, feedback, HUD, sound if you can do it without external assets.
  The point of 3D is spectacle; deliver some.

## The neural network

The existing neural network is **yours to decide about**: remove it entirely if it
serves no purpose, or repurpose it (AI opponents, adaptive traffic, a rival racer…).
Justify your choice in RESULT.md. Dead code left behind counts against you.

## Constraints

- **Zero build step.** The game must run by serving this directory statically and
  opening `index.html`. No bundlers, no npm install.
- **Three.js via CDN is allowed** (a single `<script>`/importmap from a public CDN).
  Everything else self-contained. Raw WebGL is also acceptable if you prefer.
- The game will also be embedded in a **sandboxed iframe** where `localStorage` /
  `sessionStorage` / cookies **throw**. Treat storage as best-effort: wrap access in
  try/catch and degrade gracefully (e.g. high score just doesn't persist). The game
  must never crash because storage is unavailable.
- This is a **single attempt**. There is no follow-up prompt and no human to ask.
  Make decisions and commit to them. If you have browser tooling available, use it
  to verify the game actually runs before you finish.

## Deliverable

Besides the working game, write **`RESULT.md`** in the project root:

1. **What I built** — one-paragraph pitch of the game.
2. **Features** — bullet list of everything added/changed.
3. **How to play** — controls and goal.
4. **Camera & 3D approach** — what you chose and why.
5. **The neural network** — what you did with it and why.
6. **Design decisions** — the interesting trade-offs you made.
7. **Next** — what you would build with one more attempt.