@ai-ecoverse/py-numpy

NumPy 2.5.4 for slicc WASIX CPython 3.15 (cp315-wasix_wasm32 side modules). Since the Python 3.15 flip (homescoop#761) this is latest; the cp314 builds stay on npm by exact version (release notes).

• Build: in homescoop's CI from the PyPI sdist, with Cython 3.3.0 and numpy's vendored meson, using wasixcc PIC side modules (sysroot-ehpic, wasix-sysroot 2025.9.30-38). The C++ headers are the sysroot's.
• OpenBLAS: @ai-ecoverse/wasix-openblas 0.3.34-1, linked statically into _multiarray_umath, _umath_linalg and lapack_lite (OpenBLAS with its C LAPACK, single-threaded). Since 2.5.4-3; before, a 0.3.28 build inside this package.
• Python range: >=3.15.0-1 <3.15.1.
• numpy.linalg: raises LinAlgError for a failed factorisation (homescoop#1082).
• Commands: numpy-config (--version, --cflags) and f2py are on PATH, as launchers for numpy's console scripts (homescoop#1405). f2py can't build an extension in slicc, since there is no Fortran compiler.
• Notices: THIRD-PARTY-NOTICES.md carries OpenBLAS's, LAPACK's and GotoBLAS2's licences and those of the code numpy vendors (libdivide, pocketfft, Highway, the numpy.random generators).

On wasm32

• The default integer is int32 (homescoop#1093). wasm32 is a 32-bit platform, so np.intp is 4 bytes and np.arange(3) or np.array([1]) are int32 (desktop: int64). Their sums wrap at 2**31, and a CSV round trip in pandas comes back as int64. Pass dtype=np.int64 where it matters.
• No floating-point error reporting (homescoop#1081). WebAssembly has no FP status flags, so 1.0/0, np.sqrt(-1) and overflows give the right inf/nan but no RuntimeWarning, and np.errstate(all="raise") doesn't raise FloatingPointError. np.linalg still raises LinAlgError.