Version 0.3.1

Added

• Added official support for Python 3.14 across project metadata and CI pipelines.
• Added universal-edit-distance (v0.4.3) as an additional WER baseline in benchmarking scripts.
• Integrated ued.word_error_rate(refs, hyps) into benchmarks/speed_comparison_librispeech_data.py for runtime comparison.
• Integrated universal-edit-distance into benchmarks/memory_comparison_synthetic_data.py for peak memory usage comparison.
• Added an explicit [profile.dev] configuration to align optimization settings with release builds while retaining debug symbols.

Changed

• Optimized standard WER calculation by replacing full O(m×n) matrix allocation with space-efficient rolling window approach, reducing memory usage from O(m×n) to O(n) and improving performance.
• Updated pyproject.toml, CI.yml, and ci-check.yml to test and build against Python 3.14.
• Updated Rust dependencies: pyo3 to 0.27.2 and rayon to 1.11.0.
• Updated PyO3 usage in wer_analysis.rs to replace deprecated PyObject with Py<PyAny>.
• Replaced deprecated Python::with_gil with Python::attach to align with PyO3 0.27.x APIs.
• Enabled symbol stripping in [profile.release] to reduce PyPI wheel size.
• Aligned LTO, codegen units, and optimization level across dev and release profiles for consistent performance characteristics.
• Updated development dependencies to latest versions, including maturin, pytest, mypy, and ruff.
• Updated dataframe dependencies to latest versions of polars and pandas.
• Updated benchmark dependencies to latest versions, including jiwer, evaluate, datasets, and related tooling.
• Temporarily disabled torchmetrics and universal-edit-distance benchmarks due to missing Python 3.14 wheels.
• Updated benchmark scripts to run cleanly on Python 3.14 without torch-based dependencies.
• Updated weighted WER test script to suppress static type checker errors for intentional invalid input test cases.
• Refactored align_and_stats function in wer_analysis.rs to return a named AlignmentStats struct instead of a complex 7-element tuple, improving code readability and resolving Clippy type complexity warnings with zero performance impact.
• Added speed_comparison_librispeech_full.py benchmark script to evaluate WERx performance across both LibriSpeech test-clean and test-other datasets (5,559 total utterances), providing comprehensive real-world performance metrics for README documentation.

Fixed

• Removed PyO3 deprecation warnings during Rust extension compilation.