Stop test runs writing derived weather models into the repo

[jlmaurer]

Problem

Running the test suite rewrites files inside the repository.

combine_weather_files writes the time-interpolated weather model next to its inputs:

https://github.com/dbekaert/RAiDER/blob/dev/tools/RAiDER/cli/raider.py#L845

weather_model_file = wfiles[0].parent / (...)
ds_out.to_netcdf(weather_model_file)

That is reasonable for a real weather-model cache, but the test fixtures point that cache at tracked directories, so every run with -interp center_time or -interp azimuth_time_grid deposits derived files into the working tree:

• test/gunw_test_data/weather_files/GMAO_*_timeInterp_*.nc — two tracked files overwritten, showing up as binary diffs in git status that are easy to commit by accident.
• test/gunw_azimuth_test_data/weather_files/HRRR_*_timeInterp*.nc — four files, ~41 MB each (~165 MB) written into the tree on every run and invisible to git status, because .gitignore:29 hides that directory.

Checksumming all 24 tracked .nc files before and after a full pytest -m "not long" run confirms exactly two tracked files change, both _timeInterp_; the other 22 are read-only.

The two overwritten GMAO files are pure byproducts:

• they are derived from the four tracked input files in the same directory;
• no test references them — conftest.py lists only the inputs;
• deleting both and running pytest test/test_GUNW.py regenerates them and passes.

They also churn without any change in the science. Comparing the committed copy against a regenerated one, the delay values are bit-identical; the whole 1.80 MB → 2.20 MB size difference is wet/hydro widening from float32 to float64 (2 × 394400 − 2 × 197200 = 394400 bytes, matching exactly), because the interpolation weights are numpy float64 scalars. A numpy promotion-behaviour difference alone produces a 400 KB binary diff.

Changes

1. Untrack the two derived GMAO _timeInterp_ files (git rm --cached; the files stay on disk) and add .gitignore rules for *_timeInterp_*.nc / *_timeInterpAziGrid_*.nc.

2. Stop the tests writing into the repo at all. A new _linked_weather_files helper in test/conftest.py symlinks the tracked input files into a session-scoped scratch directory and hands the tests those paths. Since the delay workflow never calls .resolve(), wfiles[0].parent is the scratch directory, so the derived output lands there instead. Symlinks mean nothing is copied — this matters for the 128 MB of HRRR inputs.

   Applied to all three weather-file fixtures: weather_model_dict_for_gunw_integration_test, weather_model_dict_for_azimuth_time_test, and weather_model_dict_for_center_time_test.

Verification

With every derived file deleted from both fixture directories beforehand:

pytest test/test_GUNW.py -m "not long" -k "not availability_check"

passes, and neither fixture directory receives any _timeInterp_ / _timeInterpAziGrid_ output afterwards. git status is clean.

(test_hrrr_availability_check_using_gunw_ids is deselected above only because it hits live HRRR servers and fails with a socket timeout on this network; it is unrelated to these changes.)

Not included

The four tracked 32 MB HRRR input files under test/gunw_azimuth_test_data/weather_files/ are left alone here. They are already matched by .gitignore:29 but tracked anyway, and unlike the derived products they cannot simply be removed — test_azimuth_timing_interp_against_center_time_interp needs them. Fixing that requires a way to fetch them, which is out of scope for this PR. Tracked separately in #815.