Move the moment equations into FORTRAN 77 and build with meson-python

[saudzahirr]

The Cox (1979) moment equations are a quadruple loop over the input variables evaluated for every uncertain result, and in Python they dominated the cost of the library. They are now FORTRAN 77 subroutines in src/soerp/_soerp.f, compiled into an extension module by f2py, which returns the numerical core to the language the original SOERP was written in. Python keeps the automatic differentiation, the distribution moments and the public API.

Measured against the pure-Python implementation this replaces, on the same machine and inputs:

rawmoment(k=4), equation (A-9)     soerp_numeric(), the full call
n    python    fortran  speedup    python    fortran  speedup
2      59.9us    3.2us    18.8x     179.1us   12.5us    14.3x
3     203.8us    4.2us    48.8x     462.1us   13.8us    33.4x
5    1195.4us    4.4us   271.5x    2131.3us   16.9us   125.9x

10 18598.0us 13.0us 1434.9x 19636.0us 41.1us 477.3x

Cox's own three-part assembly now takes 16us end to end rather than 495us, so a result that used to cost half a millisecond is comfortably inside one. The whole test suite runs in half the time it did.

The kernel is verified, not assumed: it reproduces the exact expansion of E[y**k] for n = 1..5 across every combination of linear, pure-quadratic and cross-product terms, under both standardized and general input moments, and still reproduces Table 2 of the paper.

Build system, following the f2py-tpl template:

• meson-python replaces hatchling, with meson.build at the root, one per package level, and the version derived from git by bin/get_version.py.
• bin/build.py wraps install / wheel / clean.
• Windows links the gfortran runtime statically so a wheel imports without MinGW present. Each flag is probed rather than assumed, since -static-libquadmath only exists from GCC 9 onwards and an unsupported flag is a hard link error.
• VCCP is zeroed before use: it is intent(out) but only the i < j triangle is written, so the rest would otherwise be uninitialised memory.
• VARCMP takes its scratch arrays from the caller via intent(cache,hide) rather than using automatic arrays, which keeps the source inside FORTRAN 77 and leaves the number of input variables unbounded.

The package moves under src/. Without that, a local soerp/ directory shadows the installed package, so pip install . followed by pytest from the repository root imports the source tree, which has no compiled extension, and every test errors on cannot import name '_soerp'.

CI has to change shape because the wheel is no longer pure Python. A single artifact can no longer be installed across the test matrix, so the test job builds from source on each runner and installs gfortran where the image lacks it; wheel building moves to cibuildwheel across Linux, Windows and macOS, with the results collected into the artifact the publish workflow already expects.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!