laya-zig: a from-scratch, dependency-free Zig CPU runtime for Laya

# laya-zig — a pure-Zig CPU runtime for the Laya decision model

**English** | [简体中文](README.zh-CN.md)

> ### Status: work in progress
>
> This is about **two hours of work**, from "can this model run outside Python?"
> to a working port. It runs end to end and the forward pass is checked against an
> independent reference implementation, and it has had one serious performance
> pass (see *Measured* — a Snake decision costs ~200 ms on a laptop CPU), but
> treat it as a **half-finished project**: expect rough edges and missing
> features. It has not been used in anything real.
>
> Issues, corrections and pull requests are welcome.

A from-scratch Zig runtime for the [Laya](https://github.com/NandhaKishorM/laya)
System-1 decision model (`convaiinnovations/laya-multilingual`, 322M). No Python,
no torch, no BLAS: it reads `model.safetensors` directly and runs the whole forward
pass — tokenizer included — on the CPU, in a single process.

It ships with a CLI, a local browser UI, and a Snake probe that measures what a
zero-shot decision model actually does on a task it was never trained for.

## What this is

Laya is a **System-1 decision model**: one forward pass, non-autoregressive, no text
generation. You give it a state (text, JSON, or a conversation) and typed questions,
and it returns option probabilities directly — nothing to parse, nothing to
hallucinate. The reference implementation is Python (`pip install laya`); the model
and its training come from
[Convai Innovations](https://github.com/NandhaKishorM/laya).

This repository is an independent **Zig re-implementation of the same weights**,
aiming at a single dependency-free binary that reads 614 MB of weights and starts
answering in about half a second, offline.

## Highlights

- **No dependencies** — no torch, nump