fix: exclude VASP data generator from pytest - #387

The rename is right and I verified it fixes a real failure. Blocking only on the docstrings, which now advertise a regeneration workflow that corrupts the fixtures.

The fix works

With the old filename restored, pytest --collect-only -q tests/fp/data.vasp.kp.gf gives

ERROR tests/fp/data.vasp.kp.gf/make_kp_test.py - FileNotFoundError: [Errno 2] No such file or directory: 'POSCAR'
!!!!!!!!!!!!!!!!!!!! Interrupted: 1 error during collection !!!!!!!!!!!!!!!!!!!!

At this head it reports "no tests collected", clean. The repository configures no python_files, so pytest's default ["test_*.py", "*_test.py"] applies and the old name matched the second pattern. Nothing anywhere in the tree references the generator by either name, so the rename breaks nothing.

Two corrections to the record while I am here, neither of which changes the verdict on the rename.

The AttributeError: module 'ase' has no attribute 'geometry' quoted in #361 does not actually fire. A bare import ase really does hide ase.geometry in ase 3.28, but import dpdata binds it as a side effect, and the old file imported dpdata. With the old import order the attribute access succeeds, which is why the collection error above is a FileNotFoundError raised on the line after cellpar_to_cell. Importing the function directly is still worth doing, since relying on a third-party package's import graph is fragile, but it is hardening rather than a fix for an observed error.

And the blast radius is smaller than the issue suggests: CI runs coverage run ... -m unittest, whose default discovery pattern test*.py never matched this file. This is a local-developer fix. Note that #388 does not change that either, it only declares the dependency.

Blocking: main() regenerates half of each fixture

The PR adds a module docstring saying the script regenerates the k-point test data, a main() documented as regenerating all fixture directories, and a __main__ guard. Together these turn a script that previously only ran by accident into one a maintainer is invited to run. But make_one writes only POSCAR. Each test.NNN/ also holds a kp.ref, which tests/fp/test_vasp.py loads and compares against; nothing in the repository writes those.

So following the new docstring desynchronizes the two halves:

$ cd tests/fp/data.vasp.kp.gf && python make_kp_data.py     # exit 0, all 30 POSCARs replaced
$ pytest tests/fp/test_vasp.py -k make_kp
FAILED tests/fp/test_vasp.py::TestVASPInputs::test_make_kp - AssertionError: False is not true

Recomputing make_kspacing_kpoints(cell, 0.16, False) for each directory afterwards, all 30 disagree with their committed kp.ref.

Either is fine by me: have make_one also write kp.ref alongside the POSCAR, or reword the docstrings so they do not promise a complete regeneration and say the references must be redone by hand. What should not ship is a documented entry point that quietly breaks the suite.

Not blocking

make_one reads a bare "POSCAR" and main writes bare test.NNN/, so the new entry point only works when the current directory is exactly the fixture directory; from the repository root it exits 1 with the same FileNotFoundError that caused the original bug. Anchoring both to Path(__file__).parent would make the __main__ guard robust rather than merely lucky, and the file is already open.

Separately, and definitely not for this PR: the angle sampling has been wrong since the file was created in 2022. # [1, 179) sits above np.random.random(3) * (178 / 180) + 1, which yields angles in [1, 1.99) degrees, so every fixture is a nearly collinear cell. Measured across the 30 committed POSCARs, all 90 angles fall between 1.0007 and 1.9848 degrees, and test.000/kp.ref is 3012 1844 2485. The test therefore never exercises ordinary cells. I mention it because you are rewriting this file, but if anyone picks it up: the obvious correction * 178 + 1 is not safe. Uniformly random angle triples are usually not realizable as a parallelepiped, and 143 of 200 sampled triples tripped ase's cz_sqr >= 0 assertion. It needs rejection sampling, and regenerating kp.ref along with it.

On merge order: this conflicts with #405, which modifies make_kp_test.py while this PR deletes it (CONFLICT (modify/delete)). #382, #383, #384 and #386 all merge clean.