Fetch the real-data test fixture from figshare instead of running fMRIPrep locally

[izpyon]

This targets ks_update_fmriprep so it lands inside KamitaniLab#116 rather than as a separate PR. It implements the real-data fixture workflow discussed there: the fixture is now downloaded from figshare instead of being regenerated by each developer.

I have uploaded the pre-processed data as offered in my earlier comment, and it is now published:

doi:10.6084/m9.figshare.32857559.v1 — ds006319_fmriprep-1.2.1_minimal.tar.gz, 1,769,324,197 bytes, md5 7f4ec67fb6e73239d3e5dc1066ef8f55, CC BY 4.0.

@kencan7749 If you have already started on this yourself, please feel free to close this and take only the parts that are useful. I did not want to hold you up with another round of questions.

What changes

• Add scripts/real/step_1_figshare_download.sh. Downloads the archive, verifies its md5 before extracting, and places it under tests/data/mri/. A truncated download therefore fails loudly instead of surfacing as a confusing golden-master mismatch. FIGSHARE_LOCAL_TARBALL=/path/to.tar.gz points it at a locally rebuilt archive instead of the published one.
• Rename scripts/real/step_5_run_test.sh to step_2_run_test.sh, so the real-data workflow is the same two steps as the mock one. Contents unchanged.
• Move the pipeline scripts to scripts/fixture_generation/ (step_1_download.sh, step_2_run_freesurfer.sh, step_3_run_fmriprep.sh, step_4_prepare_gm.sh, _fs_env.sh.example and its .gitignore). These are moves, not deletions: they are no longer part of the test workflow, but they are still needed to reproduce the fixture, and the README has to document the generation command.
• Add scripts/fixture_generation/step_5_build_minimal_bundle.sh plus clean_bdata_header.py and gen_label_mapper.py. Without these, what this repository can reproduce stops at the ~60 GB intermediate outputs, not the published archive. The build step also strips the BData header, because BData.save() records the call stack, including absolute paths, into /header, which must not be published.
• Update README.md and TEST_COVERAGE.md, and ignore the downloaded fixture.

Nothing in test_fmriprep_real.py or test_fmriprep_utils.py is modified. RealDatasetMixin already resolves tests/data/mri/ds006319, the golden master and the label mapper, and it still skips itself when the fixture is absent.

Addressing the review comments

@micchu's comment:

Request	Where
Host the fixture externally and download it explicitly only for real_data tests	figshare; step_1_figshare_download.sh is run manually. Nothing is fetched during a normal pytest run, so the default suite stays offline
Repository holds test code and download / verification scripts, not the binaries	Only scripts are added; ds006319/ and the label mapper are now gitignored
Document fixture version, checksum, source dataset, fMRIPrep / FreeSurfer versions, Docker image tag, generation command, and expected directory structure	Fixture table in tests/mri/fmriprep/README.md
Make the archive filename identify the fMRIPrep version	ds006319_fmriprep-1.2.1_minimal.tar.gz
Include *_events.tsv and *_bold.json	Included for runs 01–03
Single subject with multiple sessions / runs so exclude is exercised	sub-S1 with ses-SoundTest01 (fully excluded) and ses-SoundTest02 (run-01 excluded, run-02 and run-03 loaded), matching exclude={"session/run": [[1, 2, 3, 4, 5], [1]]}
Workflow should only download, extract and verify the checksum, with as little shell as possible	The real-data workflow went from five scripts to two, and the fMRIPrep execution phase is out of it

Two things I did not do, deliberately:

• The archive is tar.gz rather than zip. Happy to rebuild and republish as zip if you prefer.
• The real_data marker addopts point is left alone, since it belongs to this PR rather than to the fixture work.

If you would rather have the download logic in Python than in a shell script, say so and I will convert it.

Fixture contents

tests/data/mri/
  ds006319/
    sub-S1/ses-SoundTest02/func/          run-01..03 events.tsv and bold.json
    derivatives/fmriprep/fmriprep/sub-S1/
      ses-SoundTest01/func/               empty, holds session index 0
      ses-SoundTest02/func/               run-01 confounds; run-02/03 space-T1w_desc-preproc_bold and confounds
  golden_master/real/
    test_output_fmriprep_real_exclude.h5  expected output
    temp.tsv                              pinned stimulus_name label mapper

Only what the test reads is included: 478 MB of inputs and a 2.4 GB golden master, down from the ~60 GB the pipeline produces. The pinned label mapper is what makes this possible; without it RealDatasetMixin rebuilds the mapping at run time by scanning all 560 raw events.tsv files, which would mean shipping the whole dataset. The source data are derived from OpenNeuro ds006319 v1.0.1 (CC0), functional EPI only, with no anatomical images.

Verification

On the lab base environment (Python 3.8.5), against the published article rather than a local copy:

$ bash ./tests/mri/fmriprep/scripts/real/step_1_figshare_download.sh
md5 OK: 7f4ec67fb6e73239d3e5dc1066ef8f55

$ bash ./tests/mri/fmriprep/scripts/real/step_2_run_test.sh
tests/mri/fmriprep/test_fmriprep_real.py .                               [100%]
================== 1 passed, 43 warnings in 101.75s (0:01:41) ==================

$ python -m pytest tests/mri/fmriprep/ -m "not real_data" -q
36 passed, 1 deselected, 50 warnings in 3.84s

git status stays clean afterwards: the 478 MB dataset and the 2.4 GB golden master are both ignored.