Feat/export cli

[EliHei2]

feat(export): segger export CLI — get segmentations into the tools people actually use

Motivation : A segger run produces a segger_segmentation.parquet — but that would need post processing to get the transcirpt metadata and make cell table, make polygons for vis etc. this is to get that parquet file into Xenium
Explorer, anndata, +metadata .parquet, or spaitaldata (sopa specs).

Specifics: segger export reads a segmentation result and emits it in four formats:

• xenium (default) — Baysor-style segmentation.csv + viz polygons / cell GeoJSON for 10x's xeniumranger
  import-segmentation, and prints the exact command to run next.
• anndata — cell × gene .h5ad
• merged — original transcripts joined with segger assignments (+ fragment annotations).
• spatialdata — SOPA-compatible Zarr (optional segger[spatialdata] extra).

Vis: multi-core Delaunay cell-boundary generation from assigned transcripts (or reuse the platform's own
boundaries), and export-time re-thresholding so you can tune the assignment cutoff without re-running
segmentation.

Usage

`segger export -s <segmentation.parquet> -i -o [--format ...]

`
  # Xenium Explorer (default) — writes import-segmentation inputs + prints the xeniumranger command
  segger export -s seg.parquet -i data/xenium_run/ -o export/

  # AnnData for scanpy
  segger export -s seg.parquet -i data/xenium_run/ -o export/ --format anndata

  # Merged transcripts table
  segger export -s seg.parquet -i data/xenium_run/ -o export/ --format merged

  # SpatialData / SOPA Zarr (needs: pip install 'segger[spatialdata]')
  segger export -s seg.parquet -i data/xenium_run/ -o export/ --format spatialdata -n 8

  # Re-threshold at export time — no re-segmentation
  segger export -s seg.parquet -i data/xenium_run/ -o export/ --min-similarity 0.5

Key flags:

--format {xenium,merged,anndata,spatialdata} · --xenium-mode {transcript_assignment,geojson,both} ·
--min-similarity / --min-similarity-shift · --boundary-method {delaunay,input} · -n/--num-workers. 

Cell-ID column
auto-resolves across segger_cell_id/seg_cell_id/cell_id; spatialdata degrades gracefully if the extra isn't
installed.

Implementation

7 focused commits, ~one per module, reviewable in order:

1. refactor(utils) — utils.py → package + optional-dependency helpers
2. lazy export package + output-format/writer registry (OutputFormat, get_writer, register_writer)
3. Delaunay boundary generation --> vs. convex hull!
4. Xenium Explorer via xeniumranger import-segmentation
5. merged-transcripts writer
6. AnnData writer
7. segger export CLI

Heavy/optional deps (spatialdata, anndata) are imported lazily so the base install stays light. Covered by
tests/test_export_cli.py and tests/test_export_xenium_import.py.