Lunar Laya: Atari-style lander with a CUDA-trained Laya checkpoint that lands 30/30 on MLX

# Lunar Laya

An Atari-inspired lunar landing experiment with **Laya typed decisions running
locally on Apple MLX**, deterministic arcade physics, and a self-contained
browser flight recorder. Watch rotation, engine power, fuel, terrain contact,
model probabilities, and guidance interventions on the same timeline.

**Now includes an additional CUDA-trained Laya checkpoint and a VanillaJS
mission-control SPA.** The trained model landed 30/30 held-out local MLX flights
without post-inference assistance, following explicit guidance requests.
The original checkpoint and modes remain unchanged. See
[training details and limitations](docs/training.md).



Actual FP16 trained-model recording, seed 3000, played at 4× simulation time.
See [GIF reproduction and provenance](docs/media.md) and
[low-memory native MLX deployment](docs/mlx-optimization.md).

This is a new simulation inspired by **Atari's 1979 vector arcade Lunar Lander**,
not an Atari ROM emulator, a Gymnasium LunarLander wrapper, or the LayaAir game
engine. Laya here means Convai Innovations' decision model. MLX runs its neural
inference; the small physics loop runs in ordinary Python.

## Documentation map

| Guide | Read it for |
|---|---|
| [User guide](docs/user-guide.md) | Installation choices, every command-line option, replay controls, telemetry, offline use, and troubleshooting |
| [Implementation](docs/implementation.md) | Data flow, equations, units, terrain, contact rules, guidance, model interface, and recording schema |
| [Developer and experiment guide](docs/development.md) | Python API examples, replay regeneration, result analysis, reproducibility, testing, and changing the controller |
| [Validation](docs/validation.md) | Actual model and baseline outcomes, latency measurement boundaries, reproductio