output real numbers for Eccentricity, MajorAxisLength, and MinorAxisLength

[joefutrelle]

This PR fixes a bug where complex numbers (with 0 imaginary part) were being written out for Eccentricity, MajorAxisLength, and MinorAxisLength.

This requires minor changes to how the features are computed, and has an impact on CSV output formatting.

• replace the call to np.linalg.eig with a call to np.linalg.eigh (the symmetric matrix solver, which outputs real numbers)
• guard against small negative values from rounding being fed into sqrt and returning NaN
• write floating-point real numbers to CSV output, making precision uniform and removing 0j notation

"Golden" test output is corrected for the new output format; the numbers are the same.

[johnwaalsh]

In tests/test_regression.py I see these functions:

def _to_complex(value):
    """Parse one CSV cell into a complex number.

    Some columns (Eccentricity, MajorAxisLength, MinorAxisLength) are currently
    written as complex literals such as '(0.888+0j)'. Parsing everything as
    complex makes the comparison indifferent to that representation, so if those
    columns are later changed to emit plain floats the values still compare
    equal rather than failing spuriously.
    """
    if value is None or (isinstance(value, float) and np.isnan(value)):
        return complex(np.nan, 0)
    return complex(str(value).strip())


def _read_features(path):
    """Read a features CSV as {column: complex128 array}, preserving order."""
    df = pd.read_csv(path, dtype=str)
    return df.columns.tolist(), {
        column: np.array([_to_complex(v) for v in df[column]], dtype=np.complex128)
        for column in df.columns
    }

Should we no longer have the _to_complex function because we're no longer outputting complex values?

[joefutrelle]

Yes--this is left over from trying to get the tests to work properly before the fix, and it should be corrected.

[johnwaalsh]

Confirmed that this works as intended on my end. Looks good to me