fix: read a background-inclusive one-hot correctly in _foreground

[Hendrik-code]

The defect

The package carries two one-hot conventions:

helper	convention
collapse_onehot_to_index (gpu/fromSeg.py)	background implicit — channel c is label c + 1, background voxels are all-zero
seg_region_masks (gpu/fromSeg.py)	background included — one channel per distinct value, its own docstring says "background included"

_foreground (gpu/contrast.py) is written for the first. Given the second, the background channel covers exactly the voxels the foreground channels do not, so amax > 0 is True at every voxel:

background-inclusive one-hot: fraction foreground = 1.0     <- wrong
background-implicit one-hot : fraction foreground = 0.5
single-channel label map    : fraction foreground = 0.5

With an all-True foreground, in_seg applies the transform to the whole patch and out_seg to none of it — which is word for word the failure _foreground's own docstring says it was written to fix, in the one layout it was not checked against. unit_tests/test_seg_layout.py builds exactly this shape (for v in range(5)), and seg_region_masks produces it for every single-channel label map.

The fix

A background-inclusive one-hot is recognisable without guessing at label values: it is multi-channel and leaves no voxel unset, because the background channel covers whatever the foreground channels do not. A background-implicit one-hot always has all-zero background voxels unless the patch is labelled edge to edge, which a real segmentation patch is not. Single-channel input is never second-guessed.

Tests

unit_tests/test_foreground_layouts.py asserts all four layouts give the same answer for the same anatomy — label map, background-implicit one-hot, background-inclusive one-hot, and the actual output of seg_region_masks, since one helper feeds the other's layout. Plus an empty mask (no foreground) and a fully labelled single-channel map (all foreground, not second-guessed).

Verified failing before the fix (2 failed — the two background-inclusive cases) and passing after. Full suite unchanged against baseline; ruff check . clean.

🤖 Generated with Claude Code