fix: index the domain-transfer LUT by label value, not by position

[Hendrik-code]

The defect

Per-class weights came from

class_masks = seg_region_masks(seg, max_regions=NC).float()

which, for a single-channel label map, orders its channels by the values that happen to be present, ascending. _accumulate then indexes the bank's class axis by that position:

lut_c = lut_used[c if per_class else 0, c]

The bank's class axis is a fixed anatomical taxonomy, so which anatomy a class's transfer curve landed on depended on what else was in the patch — and changed from batch to batch.

Measured

On a synthetic bank whose class c maps every intensity to the constant c / NC, so the class each voxel received reads straight off the output:

patch holding {0, 3}        -> label 3 reached LUT class 1
patch holding {0, 1, 3, 2}  -> label 3 reached LUT class 3

Same label, same anatomy, different transfer curve, decided by its neighbours.

The fix

Expand a single-channel label map per value (seg.round() == arange(NC)) so channel c means label c everywhere. One-hot callers keep seg_region_masks and are untouched.

Not run against the real bank — stated plainly

The LUT bank is hundreds of megabytes, built offline and deliberately not shipped; SMAUGLAB_DOMAIN_BANK is unset in this environment, so I could not exercise this against real data. The test writes a synthetic .npz in the bank's own format instead.

What the test pins is stability — a label reaches the same class whatever else is present — which holds whatever the bank's taxonomy turns out to be. If the bank's class c is not label c, the mapping still needs an explicit table from whoever owns the bank; this PR makes the indexing deterministic, which it was not.

Tests

unit_tests/test_domain_transfer_classes.py: a label must reach the same class regardless of which other labels are present (fails on the old code with 1.0 vs 3.0); each label maps to its own class when all are present (a control that passes either way, since position equals value there); output stays finite and correctly shaped.

Verified failing before the fix and passing after. Full suite unchanged against baseline; ruff check . clean.

🤖 Generated with Claude Code