[davidackerman]

This is part of the batch, #137. Merging #137 with a merge commit marks this PR merged too.

AI-annotate asks the model to leave a thin black line between objects that touch. On Accept, the mask was shrunk to the annotation voxels first and cut into objects after. When the image's pixels are finer than the voxels, each voxel takes only one pixel in every two, so the line could fall between the pixels taken and two objects were written as one id:

image (8 nm):   ■ ■ ■ □ ■ ■ ■ ■     □ = the black line
pixels taken:     ■   ■   ■   ■     the line falls between them
saved (16 nm):    ■   ■   ■   ■     one object instead of two

The mask is now cut into objects at the image's resolution, where the model drew the line, and then the object ids are shrunk onto the voxels. Two objects stay two ids even where no black voxel survives between them. The fix is in both the server and the browser.

Who sees a change

• Server models whose input voxels are finer than their output: touching objects are no longer merged. A thin neck of one object that the shrink skips leaves pieces sharing one id, since they're one object at the image's resolution.
• Everyone else, including every browser-catalog model: the image's pixels are the same size as the voxels or coarser, and the labels are exactly what they were. The same voxels are written in every case; only the ids can differ, and only when shrinking.

Code

• pipeline.mask_to_write_shape now returns the mask's objects on the write box (int32, 0 for background), numbered in the box's raster order. labels_for_box takes those objects instead of a bool mask.
• The staged mask_write.npy holds the objects. The preview's overlay and the mask fraction read them as > 0.
• The browser's maskToWriteShape and labelsForBox mirror the Python. The parity fixture gains a case with a one-pixel line on coarser voxels.

Checks

• New test: two objects split by a one-pixel line, at 4 nm pixels on 8 nm voxels, come out as two ids. It failed before the fix.
• tests/ai_annotate: 278 passed. The full vitest suite passes and the typecheck is clean.
• The old and new pipelines gave identical labels on 108 random cases where the image is the voxels' size or coarser, with painted voxels and the planes beside.