Changes

0.1.0 (unreleased)

Initial release.

• embeddings.core: load-model / with-model / close over any sentence-transformers-style ONNX export (model.onnx + tokenizer.json), embed / embed-batch / dimension. Session inputs are introspected (input_ids / attention_mask / token_type_ids fed only when the model declares them); batches are padded attention-mask-aware; rank-3 token outputs are pooled, rank-2 pre-pooled outputs used directly.
• embeddings.pooling: mask-aware :mean / :cls / :max.
• embeddings.math: dot, norm, l2-normalize, cosine-similarity on primitive float[].
• Validated against real all-MiniLM-L6-v2: parity with the Python sentence-transformers reference within 2e-2 per component, batch results identical to single embeds.