ci: real-data fixture + matlab tests on github actions

[neuromechanist]

Summary

Adds a small (2.3 MB) BIDS-EEG fixture and a GitHub Actions workflow so the MATLAB test suite runs on every push and PR, not only locally on the maintainer's /Volumes mount. Implements issue #11 plus the practicum brief's Week-4 CI scope that we had been deferring.

Why now: PR #29 just landed the Phase 1 derivatives without CI coverage. Before Phase 2 (AMICA), the test infrastructure should be in place so new code is covered from day one.

Fixture

tests/fixtures/bids_mini/ — real 60 s slice of sub-NDARAA948VFH's ThePresent recording from the HBN R3 100 Hz local dataset. No synthesis, no resampling.

• Built by scripts/build_test_fixture.m (deterministic, idempotent; re-runs from the same /Volumes source produce byte-identical output for the JSON/TSV sidecars).
• Carries both subject-level and root-level BIDS sidecars. The root-level task-ThePresent_eeg.json is the key one: it has Manufacturer / ManufacturersModelName that EEGLAB uses to auto-look-up the GSN HydroCel 129 channel locations. Without it, clean_rawdata falls into a location-free path that errors at srate < 110 Hz on our 100 Hz data.
• License CC-BY-SA 4.0 inherited from HBN-EEG (attribution and license preserved via dataset_description.json + README).

CI workflow

.github/workflows/matlab-tests.yml:

• matlab-actions/setup-matlab@v2 on R2024a (same release as the existing checkcode job).
• EEGLAB pinned to tagged release 2026.0.0, cloned with --recurse-submodules so EEG-BIDS, clean_rawdata, ICLabel, dipfit, firfilt come along automatically. Cache keyed on the tag so a version bump invalidates cleanly.
• Biosig auto-installed via EEGLAB's plugin_askinstall on first run (not a submodule; needed for BDF I/O).
• Runs tests/matlab/run_all_tests.m against the fixture via HBN_BIDS_ROOT.

Test refactor

• tests/matlab/test_bids_root.m: new resolver helper. Tries HBN_BIDS_ROOT, then tests/fixtures/bids_mini, then /Volumes/.... Errors with a precise message if none resolve, so a CI failure surfaces "no fixture" rather than disguising itself as a phase1 error.
• test_phase1_no_subjects.m, test_phase1_smoke.m: drop the hardcoded /Volumes path; use the resolver.

Local CI

• scripts/local_ci.sh falls back to the fixture when BIDS_ROOT is unset, so a clean checkout works without the maintainer's /Volumes mount.
• Exports HBN_BIDS_ROOT so the shell and the in-MATLAB resolver agree.

Pre-commit changes

The fixture mirrors upstream HBN sidecars byte-for-byte. Several hooks were rewriting them or flagging upstream artifacts:

• mixed-line-ending: was treating the binary BDF as text and corrupting it. Now skips BDF/EDF/SET/FDT/MAT.
• end-of-file-fixer, trailing-whitespace: now skip tests/fixtures/.
• typos: now skips tests/fixtures/ so upstream spelling artifacts in participants.json don't fail CI.

Misc

• .github/workflows/lint.yml: drops the stale develop branch trigger (main-only now).
• logs/ added to .gitignore (transient MATLAB batch logs from local runs).

Local verification

All 4 tests pass against the fixture:

                name                 passed
    _____________________________    ______
    "test_list_eligible_subjects"    true
    "test_write_qa_channels_csv"     true
    "test_phase1_no_subjects"        true
    "test_phase1_smoke"              true

clean_rawdata on the 60 s slice removes 11 channels (vs 6 on the full 210 s recording for the same subject); different rejection statistics on different-length windows, both correct. The smoke test asserts that at least one channel is rejected and that the checkpoint round-trips with matching nbchan / srate, so a no-op rejector cannot silently pass.

Test plan

• Local: run_all_tests green against HBN_BIDS_ROOT=tests/fixtures/bids_mini.
• Local: run_all_tests green against the full /Volumes dataset (regression check).
• Fixture rebuild is deterministic and idempotent.
• Pre-commit hooks all pass without corrupting binaries.
• CI workflow runs to green on this PR (first invocation will be a cache miss for EEGLAB; expect ~10-15 min on first run, ~3-5 min thereafter).

Closes #11.
Refs #1.