[zhangzhenghao]

Summary

• Add installable CPU Python bindings under python/, using pybind11 and scikit-build-core (pip install ./python, import clap_cpp).
• Expose model loading, text/token and waveform embeddings, frame/CLaMP3 features, tokenizer, decoding and resampling with owning float32 NumPy results.
• Release the GIL for native work and serialize same-model calls; support explicit close and context management.
• Build a self-contained CPU extension with static ggml and bundled ICU, no GPU backends or external OpenMP runtime. Include pinned native sources and licenses in the source distribution.
• Preserve UTF-8 paths through Python and Windows native model/audio file loading; add real inference/decoding tests for non-ASCII paths.
• Add CPU Python CI for Linux x86_64/ARM64, macOS and Windows; build wheels from source distributions and test the installed package.

This targets the current main C++ API and does not depend on the unmerged C API PR #27. GPU bindings, publishing to PyPI and model downloads are out of scope.

Verification

• Installed CPU wheel: 58 pytest cases passed, including generated waveform-capable CLAP/MERT inference, independent NumPy text oracle, token parity, frames, decoding/resampling, lifetime, errors and concurrent same-model calls.
• Native CTest after core path changes: 26/26 passed.
• Direct checkout install and full source-distribution → wheel → install → pytest path passed.
• Wheel contents and Linux ldd checked: no CLI binaries or external ggml, ICU or OpenMP dependency.
• Independent review passed after the Unicode-path fix; actionlint and git diff --check passed.

Local verification is Linux CPU/Python 3.11. All 9 PR checks passed for this head: native CI on Linux x86_64/ARM64, macOS and Windows x86_64/ARM64; Python source-distribution/wheel/install/inference tests on Linux x86_64/ARM64, macOS and Windows x86_64, including Unicode paths. No GPU verification is claimed.