[FreakyAdy]

Closes #5

Summary

Adds OnnxRuntimeBackend for local transformer embeddings without requiring PyTorch. It accepts an explicit ONNX model path and a matching local tokenizer directory, loads tokenizer assets with local_files_only=True, and supports both sentence-level embeddings and token-level hidden states with attention-mask mean pooling.

SemanticDetector now supports an ordered fallback_backends chain. If a configured backend fails, it tries the next backend; if all configured backends fail, it logs the failures and uses the lightweight TF-IDF fallback. Existing Sentence Transformers behavior remains the default. The ONNX backend is exported from evalguard.contamination and uses the new optional onnx extra (onnxruntime + transformers, no explicit PyTorch dependency).

Coverage

• Attention-mask mean pooling ignores padding tokens
• Direct sentence-embedding outputs are accepted
• Empty batches avoid loading optional dependencies
• Missing local models fail with a clear path error
• Fallback order and TF-IDF final fallback
• Invalid thresholds and maximum lengths

Added docs/onnx_semantic_backend.md with offline configuration and model-interface requirements. CI is running; no model weights are downloaded in tests.