Laya4j: pure-Java 17+ SDK for Laya on ONNX Runtime

# Laya4j

Java SDK for [Laya](https://github.com/NandhaKishorM/laya) System 1 decision models —
fast, non-autoregressive, calibrated decisions in a single forward pass via ONNX Runtime.

```java
LayaPredictor p = LayaPredictor.single(
    Path.of("models/laya-decision-multilingual-mmbert-base-v0.3.21-e4e9ddf.onnx"),
    Path.of("models/tokenizer"),
    "multilingual");

Map<String, Decision> r = p.predict(state, Presets.triage());
r.forEach((k, v) -> System.out.println(k + ": " + v));
```

---

## Install

```xml
<dependency>
    <groupId>com.laya4j</groupId>
    <artifactId>laya4j-core</artifactId>
    <version>0.2.0</version>
</dependency>
```

Requires **Java 17+**. Publishing flow: `mvn clean deploy -P release` (see `## Publish to Maven Central`).

---

## Quick Start

```bash
# 1. Clone
git clone https://github.com/githubMJ/Laya4j.git
cd Laya4j

# 2. Build & test
mvn test

# 3. Run end-to-end demo (5 demos, includes Python alignment check)
mvn -Pdemo test-compile exec:java
```

The demo includes a Python-alignment check (4 fixed cases, expected diff = 0.0000).

---

## How It Works

```
Java app
   │
   ▼  LayaPredictor.predict(state, questions)
┌─────────────────────────────────────────┐
│ ScriptDetector  ─►  LayaRouter         │
│   (detect script)    (pick checkpoint) │
│                      english / multiling │
│                      typed-decisions      │
└─────────────────────────────────────────┘
   │
   ▼  LayaOnnxModel.predict
   build_sequence  (SequenceBuilder)
        │  same byte-ids as Python
        ▼
   ONNX Runtime Java  ──►  decision logits
        │
        ▼  softmax + temperature
   Decision (ChoiceDecision | ScoreDecision | NoulDecision)
```

**Three primitives**: `choice` (multi-class), `score` (ordinal), `noul` (yes/no probability).

---