test: Add pytest infrastructure and regression tests for the sklearn Ridge decoder

[KenyaOtsuka]

Why

The repository had no automated tests, making changes to the sklearn Ridge
decoder difficult to validate.

This PR adds a regression test suite that pins the current behavior before a
follow-up refactor of the decoder. The decoding implementation itself is
unchanged.

What

• Add the missing scikit-learn dependency and pytest development setup.
• Add GitHub Actions to run the test suite on PRs and pushes.
• Add synthetic fixtures covering repeated stimuli, multi-file subjects,
  multiple ROIs, and 2-D / 3-D feature layers.
• Add golden outputs generated from the current implementation.
• Add 32 tests covering training, prediction, normalization, label alignment,
  chunking / Fortran-order handling, output layout, and edge cases.

Details for reviewers

• pytest uses pythonpath = ["."] so the repository's top-level analysis
  modules are importable without modifying sys.path.
• Synthetic data uses unbalanced repeat counts so per-trial and per-stimulus
  statistics are distinguishable.
• Training data is split across two .h5 files to exercise the existing
  multi-BData paths.
• The reference decoder independently reproduces the current numpy /
  scikit-learn Ridge semantics, including normalization, label expansion,
  chunking, order='F' reshaping, and float32 behavior.
• Golden comparisons use a tolerance because float32 Ridge results can differ
  slightly across BLAS implementations.
• The chunk-order test uses position-dependent perturbations so an incorrect
  C-vs-Fortran reshape cannot pass accidentally.
• The current failure on zero-variance feature units is explicitly pinned.
• The decoder artifact layout is tested because evaluation.py reads
  normalization files from that layout.
• Regenerate goldens only with python -m tests.generate_golden, and only when
  the expected numerical output intentionally changes.
• Each test runs in its own working directory because the scripts create a
  relative DistComp SQLite database under ./tmp/.

Test

uv sync --group dev
uv run pytest  # 32 passed

[Copilot]

🟡 Changes recommended

A couple of tests load .npz goldens via np.load(...) without closing the returned NpzFile, which can leak file descriptors across the test run.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

Adds a first automated test/CI baseline for the scikit-learn Ridge decoding pipeline, including synthetic fixtures and golden regression outputs, so future refactors can be validated without changing the current decoder behavior.

Changes:

• Add pytest + scikit-learn dependencies and configure pytest import behavior via pythonpath.
• Introduce a synthetic on-disk dataset generator plus a reference (legacy) Ridge implementation used to pin current semantics.
• Add GitHub Actions workflow to run the test suite on pushes and pull requests.

File summaries

File	Description
uv.lock	Locks new runtime/dev dependencies (scikit-learn, pytest and transitive deps).
pyproject.toml	Adds scikit-learn, defines dev dependency group, and sets pytest configuration.
.github/workflows/test.yml	Runs uv sync --group dev and pytest on CI (Python 3.11).
README.md	Documents how to run tests and how to regenerate golden fixtures.
.gitignore	Ignores Python bytecode caches and pytest cache.
tests/conftest.py	Adds shared fixtures and ensures each test runs in an isolated working directory.
tests/generate_golden.py	Script to regenerate the committed golden .npz outputs.
tests/helpers/synthetic.py	Generates small synthetic BData + Features trees to exercise decoder edge cases.
tests/helpers/pipeline.py	Calls the real training/prediction scripts and provides helpers to read outputs.
tests/helpers/legacy_ridge.py	Reference numpy/sklearn implementation to match current scripts and support regression tests.
tests/test_pipeline_golden.py	End-to-end golden regression test for predictions + normalization artifacts.
tests/test_legacy_reference.py	Validates the reference implementation against scripts and goldens.
tests/test_predict_feature.py	Pins decoded-feature output layout and behavior for predict_feature.
tests/test_train_decoder_sklearn_ridge.py	Pins decoder artifact layout, normalization definitions, and rerun-skipping behavior.
tests/test_chunking.py	Pins chunking behavior and Fortran-order reshape semantics.
tests/test_label_alignment.py	Pins label alignment and feature expansion semantics used during training.
tests/test_zero_variance_features.py	Pins current failure mode for zero-variance feature units.

Review details

• Files reviewed: 15/20 changed files
• Comments generated: 2
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.