Version 0.2.0

Added

• Added support for Python 3.14
• New cosine_similarity_matrix API for pre-built NumPy candidate matrices, achieving speedup over list-based API by eliminating Python object iteration overhead. Optimized for production RAG pipelines and vector search with large candidate sets.
• Comprehensive test suite for cosine_similarity_matrix with 13 test cases covering edge cases, type conversions, batching behavior, and equivalence with existing API.
• Benchmark script matrix_vs_list_api.py demonstrating performance comparison between matrix and list-based APIs across various dataset sizes.
• Comprehensive benchmark script matrix_vs_numpy_sklearn.py comparing SymRank against multiple NumPy baseline strategies (naive, precomputed norms, normalized candidates) and scikit-learn, demonstrating large speedup across N=100-10,000.

Changed

• Optimized top-k selection in Rust core by replacing redundant sort_unstable_by with iterator reversal. The heap's into_sorted_vec() already returns sorted results, so reversing the iterator achieves descending order without an additional O(k log k) sort operation.
• Updated all Rust dependencies to latest stable versions:
  • numpy 0.25.0 → 0.27.1
  • pyo3 0.25.0 → 0.27.2
  • ordered-float 5.0.0 → 5.1.0
  • rayon 1.10.0 → 1.11.0
  • rand 0.9.1 → 0.9.2
  • wide 0.7.32 → 1.1.1
• Refactored speed_comparisons_streaming_10k.py to include matrix API comparison and align NumPy baseline with production-ready normalized candidates pattern.
• Optimized cosine similarity scoring loop with explicit zero-norm handling for query and candidate vectors.
• Stabilized SIMD-based cosine similarity implementation while preserving existing performance characteristics.

Fixed

• Prevented division-by-zero in cosine similarity when encountering zero-norm candidate vectors.
• Resolved duplicate ndarray dependency versions by aligning all crates to ndarray 0.17.1.

Removed

• Archived deprecated benchmark scripts:
  • speed_comparison.py
  • speed_comparisons_with_batch.py
  • speed_standard_vs_batch.py
  • results_comparisons.py
  • speed_comparisons_streaming_10k_batch.py
  • simd_vs_standard.py
• These scripts are superseded by the comprehensive matrix_vs_numpy_sklearn.py benchmark or test internal implementation details not relevant to users.