Test the pure computational functions

[gouwens]

Second of three. Stacked on #17 — review that first; the diff here is only this branch's own commits.

Covers every publicly documented name in coordinates, linestring3d, processing, dataset, metrics, morphology and angle, plus _matching_voxel_indices. These need no fixtures beyond small arrays.

Expectations come from the domain, not from the current output

A 3-4-5 triangle for segment lengths. A streamline perpendicular to a plane at 90°, one lying in it at 0°, one at 45° at 45°. One voxel per layer giving 10 µm layers at 10 µm resolution. Scale invariance where doubling both inputs must not move the answer.

No golden-master values. #11 is a deliberate behaviour change whose results match neither architecture's previous answer, so golden values captured from the current branch would cause a correct contribution to fail the suite — the precise opposite of what the suite is for.

Large-allocation helpers (transform_*_to_volume, upscale_ish_volume) are always called with small shapes; their defaults allocate against the full 1320×800×1140 atlas.

_matching_voxel_indices gets the most attention — it is pure, takes small arrays, and is the common core beneath the closest-streamline search, surface-voxel collapsing and 2-D coordinate calculation. Both open contributions reduce to its behaviour.

Defects recorded (xfail, strict)

Each marker names its tracker issue, so an unexpected pass leads straight to the report.

• ccf_streamlines.morphology issue with load_swc_as_dataframe with different delimitier when processed with MorphIO v3.3.3 #8 — load_swc_as_dataframe uses sep=" " and cannot read tab-separated or column-aligned SWC. It doesn't fail: every numeric column comes back NaN, silently, in a dataframe of the right shape.
• find_closest_streamline accepts resolution but does not use it to voxelise #23 — find_closest_streamline accepts resolution and uses it to scale the returned coordinates, but never forwards it to coordinates_to_voxels, so at any resolution other than (10, 10, 10) it voxelises the input against the wrong grid and looks up the wrong streamline.

Documented, not asserted

Two behaviours are characterization tests rather than xfails, because neither is knowably wrong:

• LineString3D.rotation_to_vector divides by 1 + dot, so it is undefined for an antiparallel target and returns non-finite values rather than raising. Cortical streamlines never hit this; a direct caller can.
• _matching_voxel_indices silently returns wrong answers for an unsorted lookup with no sorter, because np.searchsorted doesn't check. Nothing in the package validates the reference files' ordering.

Verification

156 passed, 5 xfailed, ~5s. Green on all four matrix entries.

🤖 Generated with Claude Code