Testing

The suite collects 798 items: 725 test functions in the 25 files of tests/ and 73 docstring examples from sys_mapping/. Eight tests carry the slow marker. Timing benchmarks live in the separate sys_mapping_benchmark repository, with the measurements in Benchmarks.

Running the tests

Install the package with the extras CI uses, then run pytest from the repository root:

pip install -e ".[dev,regression,glass]"

pytest                               # everything, docstring examples included
pytest -m "not slow"                 # the fast suite
pytest tests/test_contamination.py   # one module
pytest --cov=sys_mapping --cov-branch --cov-report=term-missing

pyproject.toml sets testpaths = ["tests", "sys_mapping"] and addopts = "-v --doctest-modules", so a bare pytest collects the docstring examples alongside tests/.

We ran the full suite with python -m pytest -q -p no:cacheprovider --cov=sys_mapping --cov-branch in the sys_map environment, with the LS10 and Gaia maps present and the Uchuu mocks absent: 776 passed, 7 skipped, 19 xfailed in 4 min 40 s, at 98.0 % branch coverage. Six of the skips are the test_glass_mocks.py tests that need the Uchuu mocks and one is a docstring example marked +SKIP; the 19 expected failures are the transformability entries described below.

Tests that need data outside the repository skip when it is absent: 28 of the 39 tests in test_real_templates.py and parts of test_maps.py and test_simulation.py need the LS10 and Gaia maps under ~/data/legacysurvey/dr10/systematics/0032/, six tests in test_glass_mocks.py need the Uchuu mocks, and the Corrfunc tests in test_utils.py need Corrfunc.

Test modules

Counts are from python -m pytest --collect-only -qqq.

File	Tests	Scope
`test_accuracy.py`	34	Numerical thresholds stated in the docstrings: contamination round trip, maps, likelihood, correction, utilities, model selection, bootstrap
`test_bootstrap.py`	7	`bootstrap`: spatial patches, block bootstrap, jackknife
`test_contamination.py`	19	`contamination`: parameter layout, pack/unpack, forward and inverse model, two-point correction
`test_correction.py`	30	`correction`: debiasing, PCA rotation, two-point correction and its covariance, cross-template terms, harmonic correction, over-correction warning
`test_covariance.py`	19	`covariance`: low-rank precision against a dense reference, mock sandwich, sample covariance, Hartlap factor
`test_des_y6_features.py`	19	Spatial cross-validation folds, inverse-variance pixel weights, template vetting, over-correction debias, method-marginalised covariance
`test_diagnostics.py`	50	`diagnostics`: null test, SNR ranking, footprint masking, ISD significance, residual correlation test, calibrated significance
`test_edge_cases.py`	53	Input checks, fallbacks, optional arguments and missing-dependency errors of the public API
`test_glass_mocks.py`	36	`glass_mocks`: \(n(z)\), full-sky mocks, Uchuu loading, matched spectra, spectrum choice, per-pixel null draws
`test_inference.py`	15	`inference`: log-probability, emcee sampler, chain summaries, `refine_to_mle`
`test_jax_acceleration.py`	15	JAX kernels against NumPy references: ISD \(\Delta\chi^2\), `isd_marginal_fit`, null test, polynomial-OLS backend, parallel GLASS mocks
`test_jax_transformability.py`	46	`jax.jit`, `jax.vmap` and `jax.grad` on 17 public numeric cases
`test_likelihood.py`	11	`likelihood`: Gaussian and skew-normal likelihoods, gradients, GLS precision
`test_ls10_script.py`	23	Helpers of `scripts/run_ls10_analysis.py`: parameter expansion, weight convention, skew flag, mandatory null spectrum, null fields
`test_maps.py`	34	`maps`: power spectra, synthetic maps, pixelisation, overdensity, template assignment, real-template loading
`test_mocks.py`	40	`mocks`: lognormal field, Galactic mask, mock catalogues and suites, pipeline integration
`test_model_selection.py`	39	`model_selection`: likelihood ratio test, forward selection, SNR pre-selection, `lrt_from_maxima`
`test_nonlinear_response.py`	39	Non-linear template responses: normalisation, injection, ISD-1 against ISD-3
`test_power_spectrum.py`	21	`power_spectrum`: harmonic bias, pseudo-\(C_\ell\), template subtraction, mode projection
`test_real_templates.py`	39	Every method on a mock built with the LS10 and Gaia templates (below); footprint standardisation; resolution from occupancy
`test_regression.py`	39	`regression`: weights, ElasticNet, polynomial OLS, ISD, method comparison, `run_decontamination`
`test_samplers.py`	20	Analytic additive posterior, NUTS, chain execution, sampler dispatch
`test_simulation.py`	21	`simulation`: contamination grid, injection, FITS round trip, systematic maps, footprint mask, \(w(\theta)\) recovery
`test_snr_preselection.py`	35	Two-stage SNR pre-selection on a GLASS mock with two injected templates
`test_utils.py`	23	`utils`: TreeCorr and Corrfunc two-point functions, KK correlations and covariance, template correlation matrix
`sys_mapping/*.py` (docstring examples)	73	Every `Examples` section, 1 to 11 per module

The slow tests are three calibration checks on clustered fields in test_diagnostics.py (residual test size and power, calibrated-significance false-positive rate), two comparisons of the analytic posterior and NUTS with emcee in test_samplers.py, and three checks of template-correlation support at the pixel scale in test_utils.py.

Coverage and CI

[tool.coverage.run] in pyproject.toml measures sys_mapping with branch = true; [tool.coverage.report] sets fail_under = 97 on the combined line and branch figure, which pytest-cov enforces whenever --cov is given. The measured coverage, per module, is in Test and JAX coverage: 98.0% combined, 98.8% of lines and 94.8% of branches.

.github/workflows/tests.yml runs on Python 3.11 and 3.12 with JAX_PLATFORMS=cpu and XLA_PYTHON_CLIENT_PREALLOCATE=false, after installing .[dev,regression,glass]. Pull requests run the fast suite (-m "not slow"); pushes to main, the nightly schedule and manual runs run the full suite. Both measure branch coverage and upload coverage.xml to Codecov with a flag per Python version (py3.11, py3.12). The LS10, Gaia and Uchuu data are absent there, so the tests that need them skip.

Docstring examples

Docstring examples run under --doctest-modules with the ELLIPSIS and NORMALIZE_WHITESPACE flags. NumPy 2 prints scalars as np.float64(0.5) and np.True_; the root conftest.py selects np.set_printoptions(legacy="1.25") whenever doctests are collected, so the examples print the plain form.

JAX tests

tests/test_jax_transformability.py wraps 17 public numeric cases as functions of one array and applies jax.jit, jax.vmap and, where a gradient is meaningful, jax.grad, 46 cases in all. A case passes when the transform runs on traced inputs and reproduces the eager result. The functions written against NumPy are listed in _NUMPY_ON_TRACER and their cases marked strict xfail: 8 functions and 19 cases. Porting one of them turns its cases into unexpected passes, which fail the suite until the entry is removed.

tests/test_jax_acceleration.py checks the JAX kernels that run the ISD, ranking and null-test statistics against NumPy reference implementations kept in the test file.

Real-template integration tests

tests/test_real_templates.py builds one mock at NSIDE 32 with four templates, synthetic families 0 and 1, GAIA_nstar_faint and LS10_GALDEPTH_Z, on the pixels where both real maps are finite and positive (the mask of load_real_templates()).

Parameter	Value
\(a_i^{\rm true}\)	\((0.08, -0.05, 0.06, -0.04)\)
\(b_i^{\rm true}\)	\((0.04, 0, -0.03, 0.05)\)
Clean field	lognormal, \(C_\ell^G \propto (\ell+1)^{-2}\), \(\sigma_G = 0.5\)
Galaxies per pixel	50 (Poisson)
Randoms	8 times the galaxy density, no noise
Seed	7

The tests check shapes, finiteness and a mean absolute amplitude error below a loose tolerance: 0.20 for OLS, 0.25 for ElasticNet (3 folds), ISD-1 and run_mcmc additive chains, and 0.30 for \(a_i\) and \(b_i\) of combined chains. ISD runs with a fixed chi2_68 = 50 in place of a mock calibration, and ISD-3 is checked for finite, positive weights and termination. The likelihood-ratio test must reject the additive null at 5%. The residual test must detect contamination on a template held out of the correction, against 30 null realisations put through the same correction. The SNR ranking must be non-negative with at least one entry above 0.01.

Adding tests

One test class per public function or function group.
@pytest.fixture(scope="class") or "module" for expensive shared inputs.
Test shapes, value ranges and known analytic limits.
Seed every random draw.
Mark long-running tests with @pytest.mark.slow; pull requests run without them.
Add the file to the table above.

Standalone scripts such as scripts/run_systematic_tests.py produce the result pages and are not collected by pytest.