Residual plot tests are not actually exercised

[baogorek]

Summary

The TestResidualPlots tests in tests/test_plotting.py are not actually testing the residual plotting functionality because the fixture produces results where has_residuals is False.

Details

The tests have conditional guards like:

def test_plot_residuals_available(self, result_with_residuals):
    if result_with_residuals.has_residuals:
        plotter = Plotter(result_with_residuals)
        fig = plotter.plot_residuals()
        assert isinstance(fig, ControlChartFigure)

But the result_with_residuals fixture returns a result where has_residuals is False:

@pytest.fixture
def result_with_residuals(self):
    np.random.seed(42)
    df = pd.DataFrame({
        'value': np.random.normal(100, 5, 100),
        'subgroup': np.repeat(range(20), 5),
        'time': range(100)
    })
    pdf = ProcessDataFrame(df)
    study = pdf.formulate(response=pdf.columns.value, factors=[pdf.columns.subgroup])
    return study.analyze()  # has_residuals is False!

This means the tests pass by doing nothing - plot_residuals() is never called.

Affected tests

• test_plot_residuals_available
• test_plot_residuals_histogram
• test_plot_residuals_qq
• test_plot_residuals_sequence

Suggested fix

1. Fix the fixture to produce results that actually have residuals

2. Consider eliminating scipy entirely from plot_residuals() to keep the package lightweight. The current scipy usage is minimal:

   • norm.pdf() for histogram overlay - trivial to replace:

     (1 / (std * math.sqrt(2 * math.pi))) * math.exp(-0.5 * ((x - mean) / std) ** 2)

   • norm.ppf() for Q-Q plot quantiles - can use well-known approximations (e.g., Abramowitz & Stegun or Beasley-Springer-Moro algorithm)

   This would match the approach taken in PR fix: Resolve ruff linting errors and remove scipy dependency #47 where scipy.special.loggamma was replaced with math.lgamma in spc_constants.py.

🤖 Generated with Claude Code

[cnicholas]

Fixed in commit 31b87d4.

Test fixture fix:

• Created proper replicated design (4 factors × 5 times × 3 reps = 60 obs)
• This ensures SDS 1 detection which produces residuals
• Removed conditional guards so tests actually run
• Added assertions to verify fixture behavior

scipy removal:
Already completed in commit 2b87d50 - we implemented pure Python _normal_pdf() and _normal_ppf() (Acklam's algorithm with ~1e-9 accuracy) so plot_residuals() no longer needs scipy.

Thanks for catching this test coverage gap, Ben!