The ARTS 2D path has no working automated coverage — three independent failures

[joglekara]

The 2D/ARTS integration tests do not test anything, and the CI check that appears to cover them is green for the wrong reason. Three separate problems stack up; fixing any one alone leaves the path still uncovered.

Surfaced while working on #120 / #122, which changes 2D numerical output and therefore wanted this safety net. Not caused by that PR.

1. gpu_runner exposes no GPU to JAX

.github/workflows/test.yaml defines test-tsadar-gpu with runs-on: "gpu_runner" and installs -e ".[gpu,test,hdf]", so the intent is clearly there. But both jobs on a recent run report identical results:

job	result
test-tsadar-cpu	24 passed, 5 skipped (172s)
test-tsadar-gpu	24 passed, 5 skipped (225s)

And the GPU job's own log shows the angular tests skipping:

tests/test_forward/test_angular_2v.py s
tests/test_inverse/test_angular_2v.py ss

Those skip on if not any(["gpu" == device.platform for device in devices()]), so JAX saw no GPU. Either the runner has no device attached or the [gpu] extra isn't bringing up CUDA. Worth a python -c "import jax; print(jax.devices())" step in the job so this is visible instead of silent.

2. The config files those tests open were deleted

Even with a working GPU, both tests raise FileNotFoundError:

test	opens	status
tests/test_forward/test_angular_2v.py:39	tests/configs/arts2v_test_defaults.yaml	exists
tests/test_forward/test_angular_2v.py:42	tests/configs/arts2v_test_inputs.yaml	missing
tests/test_inverse/test_angular_2v.py:134	tests/configs/arts2d_test_defaults.yaml	missing
tests/test_inverse/test_angular_2v.py:137	tests/configs/arts2d_test_inputs.yaml	exists

Both missing files were removed in b75dec9 (#95). Each test lost exactly one of its two config files, which is why this went unnoticed — the tests never got far enough to complain, because they were already skipping.

3. The surviving config has rotted independently

tests/configs/arts2d_test_inputs.yaml:81-84 sets:

flm_type: mora-yahi
LTx: 22500.
LTy: 40000.

but spherical_harmonics.py:229 reads dist_cfg["params"]["dtx"] on the mora-yahi branch. Restoring the deleted defaults file would surface a KeyError next.

Net effect

There is no automated check on the 2D/ARTS path at all — not on CPU (skipped), not on GPU (skipped, and would crash). #122 adds tests/test_forward/test_sinogram.py::test_calc_in_2D_matches_exact_rotation, which drives calc_in_2D end-to-end on CPU in a few seconds and gives that path some floor. But it is a numerical-equivalence check against the pre-existing implementation, not a physics regression test, and it does not replace the integration tests.

Suggested order

1. Add a device-report step to the GPU job so ported some code to jax #1 is never silent again.
2. Decide whether gpu_runner is expected to have a GPU. If not, the angular tests are dead weight and the job name is actively misleading.
3. Restore or regenerate the two missing config files, and fix arts2d_test_inputs.yaml's LTx/LTy -> dtx.
4. Consider making the ATS tests fail rather than skip when their fixtures are missing — a skip is indistinguishable from "no GPU here" and is what hid this.

[joglekara]

Follow-up ARTS2D correctness and inference issues discovered while auditing the unsuccessful fits:

• ARTS2D: bin-integrate unresolved EPW resonances and make the IRF area-preserving #134 — unresolved small-angle EPW resonance and non-area-preserving IRF
• ARTS2D: make 2-D susceptibility consistent with 1-D and correct drift geometry #135 — 1-D/2-D susceptibility sign and drift-geometry inconsistency
• ARTS2D: repair spherical-harmonic coordinates and make anisotropy trainable #136 — spherical-harmonic coordinates and nontrainable anisotropy
• ARTS2D: define a true reduced 2-V EDF marginal instead of a renormalized 3-D slice #137 — 3-D slice versus reduced 2-V marginal definition
• ARTS2D: make ratintn a complete, exact-node-safe principal-value quadrature #138 — incomplete and exact-node-unsafe principal-value quadrature
• ARTS2D: use a noise-aware, regularized objective with explicit gain nuisances #139 — noise-aware regularized objective and explicit gain nuisances
• Add a null-space-aware ARTS2D recovery benchmark with held-out angles #140 — null-space-aware recovery benchmark with held-out angles
• ARTS2D: keep optimizer best loss and checkpoint weights on the same iterate #141 — state/loss-consistent optimizer checkpoints

This issue remains the shared CI/fixture dependency. The new correctness issues each specify the CPU physics, gradient, or recovery regression that should be enabled as that coverage is restored.