Motivation

• Provide a small, focused utility to reproduce the inspection-store downsampling used by the Z-alignment pipeline without running alignment.
• The tool must downsample only the Y and X axes while leaving Z unchanged so downstream Z-based processing is unaffected.
• Preserve and adapt common spatial metadata (resolution/voxel_size/voxel_offset) so the downsampled store is compatible with existing pipelines.

Description

• Added a new module emalign/downsample_stack_xy.py that implements a programmatic function downsample_stack_xy and a CLI entrypoint (python -m emalign.downsample_stack_xy).
• Downsampling is performed per-slice using scipy.ndimage.zoom so the output shape is [z, y//factor, x//factor] and the Z axis is copied slice-for-slice untouched.
• Spatial attributes resolution and voxel_size are multiplied by the factor while voxel_offset is integer-divided for XY, and .zattrs/.zgroup files are written to maintain Zarr compatibility.
• The CLI supports --factor, --interpolation (linear or nearest) and --overwrite and defaults to a 10x downsample and sensible chunk sizes.

Testing

• Compiled the module with python -m py_compile emalign/downsample_stack_xy.py and the compilation passed.
• Ran a smoke test that created a temporary 3x20x30 TensorStore Zarr input and executed python -m emalign.downsample_stack_xy <src> <dst> -f 10 --interpolation nearest, and verified the output shape remained 3 in Z and became 2x3 in YX.
• Verified that .zattrs on the output contains scaled resolution/voxel_size and adjusted voxel_offset; all automated checks succeeded.

Codex Task