feat(data): add BIDS discovery and reshaping

.gitignore

@@ -151,3 +151,6 @@ data/
 # Model artifacts
 *.pth
 brain_mask_extraction_model/
+
+# Claude Code review artifacts
+.claude/REVIEW.MD/

CLAUDE.md

@@ -128,3 +128,33 @@ Every plan MUST include a Constitution Check evaluated before research and after
 ```
 
 If speckit is not installed, follow the principles above manually.
+
+## Dependency constraint
+
+Do not add external dependencies unless they are already declared in `setup.cfg` or `pyproject.toml`. Current core deps include: torch, monai, nibabel, numpy, click, zarr, fsspec, joblib, scikit-image, psutil. If you think a new dep is justified, flag it explicitly — do not silently add it.
+
+## MONAI-first rule
+
+Before implementing any data loading, transform, loss, metric, or inference utility, check whether MONAI already provides it. Prefer wrapping MONAI over reimplementing. Check `monai.transforms`, `monai.data`, `monai.inferers`, `monai.losses`, `monai.metrics` before writing new code. If MONAI's version is insufficient, document why in a code comment.
+
+## Read before write
+
+Before modifying any existing module:
+1. Read the module completely. List its implicit assumptions.
+2. Check what MONAI provides that overlaps.
+3. Do not guess function signatures — read the source.
+4. If the module has existing tests, read those too to understand expected behavior.
+
+## Testing philosophy
+
+Every test should fail if the feature is removed. If a test passes regardless of whether your code exists, it's not testing your code.
+
+## Review process
+
+A reviewer subagent runs automatically (via stop hook) when you finish a task. It executes tests, inspects the diff, and writes a verdict to `.claude/reviews/`. If the reviewer blocks:
+1. Read the review file it points to.
+2. Fix every CRITICAL finding.
+3. Address WARNING findings where reasonable.
+4. The hook will re-run on your next stop.
+
+Do not commit until the reviewer passes.

conftest.py

@@ -6,6 +6,28 @@
 import torch
 
 
+def _mps_supports_conv3d() -> bool:
+    """Return True if Conv3D works on the MPS backend."""
+    if not (hasattr(torch.backends, "mps") and torch.backends.mps.is_available()):
+        return False
+    try:
+        m = torch.nn.Conv3d(1, 1, 1).to("mps")
+        x = torch.randn(1, 1, 2, 2, 2, device="mps")
+        m(x)
+        return True
+    except RuntimeError:
+        return False
+
+
+# Disable MPS auto-detection when Conv3D is unsupported (PyTorch < 2.3 on
+# Apple Silicon).  Without this, ``nobrainer.gpu.get_device()`` returns MPS
+# and every 3D convolution raises ``RuntimeError: Conv3D is not supported on
+# MPS``.
+if hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+    if not _mps_supports_conv3d():
+        torch.backends.mps.is_available = lambda: False  # type: ignore[assignment]
+
+
 def pytest_collection_modifyitems(config, items):
     """Skip tests marked with @pytest.mark.gpu when CUDA is not available."""
     if torch.cuda.is_available():