Test data¶

Records live under testdata/. They are inputs to the code generator and notes for the oracle commands. The unit tests build their arrays in memory. They do not load a binary file from disk.

Schema¶

testdata/schema/series.json is the code-generation input:

{"name":"Series","dataset":"values","dtype":"float64","rank":1}

pixi run generate writes tests/generated/Series.mojo. pixi run check-generated diffs a fresh run against that file. tests/test_generated.mojo encodes two float64 values and decodes them.

What the unit tests cover¶

Test	What it checks
`tests/test_hdf5.mojo`	Superblock version byte 0. Scalar float32, 1-D float64, 2-D int32, chunked gzip float64, partial uint8 chunks, partial 2-D int16 chunks, group and dataset attributes. A second write matches the first file's bytes.
`tests/test_netcdf.mojo`	CDF-1 magic and a round trip of one dimension, one float64 variable, a global text attribute, and a variable text attribute. The same document through NetCDF-4.
`tests/test_bp5.mojo`	One BP5 directory: variable `v`, values `1`, `2.5`, `-3.25`, attribute `note=hi`.
`tests/test_deflate.mojo`	Adler-32 of an empty buffer, a stored-block round trip, and a multi-block buffer larger than 65535 bytes.

Oracles¶

scripts/oracle.sh writes /tmp/gld-hdf5-oracle-f64.h5, /tmp/gld-hdf5-oracle-classic.nc, /tmp/gld-hdf5-oracle-nc4.h5, and /tmp/gld-hdf5-oracle-step.bp. When the tools are installed:

h5dump prints the float64 dataset 1, 2.5, -3.25.
ncdump prints the classic file, and ncdump -k reports classic.
bpls does not open the BP5 directory. The script does not treat that as a crash of the library. The gap is documented in Techniques.

These tools are not installed in the GitHub Actions job. The job runs pixi run test, which does not need them.