Typed morton boundary

zagg's morton output coordinate is a real datatype, not an anonymous integer column. mortie defines the morton_index type once, and zagg carries it typed across every surface it crosses (issue #135):

Surface	Carrier	Form
pandas carrier (worker)	pandas.DataFrame	mortie.MortonIndexArray extension array (#71)
arro3 carrier (worker)	arro3.core.Table	morton_index Arrow extension type over the PyCapsule C Data Interface (added in mortie 0.8.4)
catalog / shardmap parquet	pyarrow.Table	morton_index pyarrow extension type (registered by mortie on import)
Zarr store (on disk)	Zarr v3 array	plain uint64 — the packed words, byte-for-byte

Who owns what

mortie owns the interchange type; zagg stores uint64. The extension metadata (ARROW:extension:name == "mortie.morton_index") lives at the interchange layer only. The write boundary (zagg.processing.write._iter_carrier_columns) extracts the packed uint64 words from a typed column — on either carrier — so the on-disk output is identical to a plain-uint64 write. Reads reconstruct the typed array via zagg.grids.morton.to_morton_array.

The adapter lives in zagg.grids.morton:

• morton_to_arrow(values) — MortonIndexArray (or raw words) → arro3.core.Array carrying the extension type, zero-copy via __arrow_c_array__;
• morton_from_arrow(col) — any Arrow C-Data source (array, chunked column, or capsule pair) → MortonIndexArray;
• is_morton_arrow(col) — detects the extension type from field metadata.

No pyarrow is needed on any worker path: the arro3 carrier and mortie's PyCapsule surface speak the Arrow C Data Interface directly (pyarrow stays in the off-Lambda catalog extra). Arrow nulls map to mortie's all-zero empty sentinel word in both directions, so missing values round-trip.

cell_ids: morton-only storage (D16)

Since the issue #304 flip, morton is the only stored cell coordinate: the packed uint64 words carry the cell identity (and its order, intrinsically), the store's dggs attrs declare grid name: "morton" + coordinate: "morton", and NESTED HEALPix ids are derived at read (the moczarr fabrication layer). The issue #135 cell_ids_encoding knob is retired — a config still carrying it fails validation with a migration pointer.

For transition stores (browser-direct demos until the gridlook morton decode lands), the escape hatch keeps writing the legacy NESTED array in addition to morton:

output:
  grid:
    type: healpix
    parent_order: 6
    child_order: 12
    emit_cell_ids: true   # default: false — morton only

A hatch store's declared coordinate stays morton; cell_ids is an extra legacy member. The grid's spatial_signature() (recorded in ShardMap manifests) is unchanged by the hatch — deliberately, so shard maps remain reusable — while the full signature() records emit_cell_ids (the hatch changes the leaf schema, so mixed states never co-aggregate). (The pre-existing output.grid.indexing_scheme config key is descriptive only and must stay nested.)

Shardmap parquet manifests

ShardMap.to_parquet / ShardMap.from_parquet are the Arrow-native siblings of the JSON manifest: shard_keys is a typed morton_index column, granule and AOI payloads ride as per-shard JSON strings, and provenance lives in the schema metadata (mirroring the Catalog geoparquet convention). This path requires pyarrow (pip install zagg[catalog]) and runs off-worker.