v0.12.4

What's New

• Pure TypeScript neural inference: Full Mamba forward pass implemented in TS. Loads 7MB binary model from MLX-trained weights. No ONNX, no daemon, no Python at runtime. Sub-second inference.
• ONNX export pipeline: MLX -> PyTorch -> ONNX -> onnxruntime-node. Alternative inference path for environments with ONNX support.
• Retrained TME v3 model: 99.38% eval accuracy on balanced 1,600-sample dataset. All 8 attack classes at F1>=0.96.
• P2 UX fixes: check command explains expected format, scan timeout shows port math, help examples use real packages.
• HuggingFace updated: v3 model published (npz, ONNX, binary formats).

Inference Tiers

1. Pure TS neural (7MB model, sub-1s)
2. ONNX via onnxruntime-node (if installed)
3. Vocabulary scorer (sub-10ms, no model needed)
4. NanoMind daemon (full LLM, if running)

1,301 tests passing.