[Hendrik-code]

Problem

# Uncrop again
if pad_size > 0:
    whole_vert_nii_uncropped.apply_pad(-pad_size)     # <- result dropped

NII.apply_pad has inplace=False, so the padding added at the top of predict_instance_mask was never
removed. On the shared test fixture:

pad_size=0 -> (68, 52, 67)
pad_size=4 -> (76, 60, 75)      # 2 * 4 larger per axis, and a shifted field of view

The docstring says the padding is "removed afterwards". predict_instance_mask is exported as
spineps.predict_instance_mask, so a caller using the documented parameter gets a mask that does not
line up with their input. The pipeline itself always passes the default 0, which is why this was never
noticed.

A second defect sat behind it: the padding was applied before the rescale to the model resolution, so
pad_size input voxels become a different number of voxels in the grid the un-pad operates on. Even a
correctly rebound apply_pad(-pad_size) would have removed the wrong amount whenever the input zoom
differs from the model's.

Fix

• Rebind the result of the un-pad.
• Pad after the rescale instead of before, which makes the removal an exact inverse at any input
  resolution.

pad_size=0 -- every call inside the pipeline -- is bit-for-bit unchanged: the existing
test_regression_golden digests still match.

Tests

unit_tests/test_instance_pad_size.py: pad_size=4 returns the same shape, affine and voxel content as
pad_size=0, and the same holds for a model whose resolution forces a rescale. All 3 fail on main.

Full suite: 226 passed, ruff clean.

🤖 Generated with Claude Code