Fix flaky test_pdf_argpercentile_gaussian with an absolute tolerance

[roytsmart]

Summary

test_pdf_argpercentile_gaussian intermittently fails in CI (e.g. on the runs for #186), for the random mean = na.NormalUncertainScalarArray(0, width=1) parametrization.

Root cause

The test reads the 25th percentile off a discretized grid x = linspace(mean - 5*std, mean + 5*std, num=100001) and compares it to the analytic Gaussian percentile mean + std*sqrt(2)*erfinv(2p - 1).

The grid value carries an absolute error of up to one grid spacing, 10 * std / (num - 1) ≈ 1e-4 (confirmed empirically: worst observed error 1.0e-4). The comparison used np.allclose(..., rtol=1e-1) with no absolute tolerance. For the random uncertain mean, whenever a distribution sample lands such that the expected percentile falls near zero (|expected| ≲ 1e-3), the bounded ~1e-4 grid error exceeds rtol * |expected| and the assertion fails. This is rare per run but shows up across many CI runs.

Fix

Add atol=1e-2 to the np.allclose call. This is ~100× the worst-case grid error yet far below the order-1 scale of the values, so it absorbs the near-zero discretization error without weakening the test for normal-magnitude values (where rtol=1e-1 still governs).

Demonstration of the failure mode and the fix:

expected = 5e-4; got = expected + 1e-4   # one grid cell of error near zero
np.allclose(got, expected, rtol=1e-1)              # False  (old)
np.allclose(got, expected, rtol=1e-1, atol=1e-2)   # True   (new)

Tests

pytest named_arrays/tests/test_pdf.py::test_pdf_argpercentile_gaussian passes (16 cases).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 97.24%. Comparing base (82d7639) to head (6f5faf5).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #187   +/-   ##
=======================================
  Coverage   97.24%   97.24%           
=======================================
  Files         105      105           
  Lines       14210    14210           
=======================================
  Hits        13819    13819           
  Misses        391      391           

Flag	Coverage Δ
unittests	97.24% <100.00%> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.