[davidackerman]

Moves chunkmirage to worldarray 0.3 and removes code worldarray now covers, on the server and in the browser engine.

Server

• Pin. worldarray>=0.3,<1. uv.lock changes only worldarray, from 0.1.0 to 0.3.0, with no new dependencies.
• Precomputed bug fixed. A precomputed scale with a non-zero voxel_offset returned 404 for every chunk. The server parsed chunk keys from 0, while Neuroglancer and tensorstore count them from the offset (3-19_2-18_1-9 for offset [3, 2, 1]). test_precomputed_voxel_offset_becomes_translation now checks both keys, and fails on main.
• Chunk keys come from worldarray. Every format's keys go through wa.chunk_key and wa.parse_chunk_key, called from one Frontend.chunk_request. It replaces four regex parsers. zarr v2 still accepts either separator.
• OME multiscales come from worldarray. frontends/_ome.py now calls wa.ome_multiscales and drops its own unit and axis-type tables. What changes in the served metadata:
  • a translation is written even when it is 0;
  • OME 0.5 has version only beside multiscales, no longer inside each multiscale too;
  • an axis whose name is neither spatial, nor c, nor time is left untyped;
  • the multiscale no longer carries "type": "chunkmirage" or a description.
• N5 levels with no offset now have their corner at the origin at every level, worldarray's rule since 0.1.1. The one test that pinned the BigDataViewer placement follows.
• Unchanged. N5 attributes and precomputed info stay chunkmirage's own, since worldarray writes neither.

Browser engine

• web/src/placement.ts ports worldarray's placement rules to TypeScript. It replaces the separate placement code in openImage, the N5 reader, the xarray reader and the precomputed reader. Fixes that come with it:

  • an OME multiscale's own coordinateTransformations are applied;
  • OME 0.6's intrinsic coordinate system is read, where it threw before;
  • N5's offset is applied, and an N5 level with no offset has its corner at the origin;
  • funlib's voxel_size and a COSEM transform on an array are read;
  • pixelResolution or resolution on an N5 group is read;
  • a zarr array outside an OME group opens (funlib, COSEM, or xarray, now of three axes or more), and a single N5 array opens.

• web/src/layout.ts types axes as worldarray does. A channel axis was typed space, and so was a t axis without a time unit.

• web/test/metadata.test.ts runs worldarray 0.3.0's two fixtures, copied unchanged (I checked them against the PyPI sdist):

  • 23 metadata layouts, each placed where worldarray places it;
  • 14 OME documents, each written as worldarray writes it. The 15th case is left out because the browser writes neither a multiscale type nor its metadata.

  Breaking either placement or typing on purpose makes the test fail.

Docs

• formats.md: the placement table, the N5 rule, precomputed keys and the browser reader.
• design.md: the worldarray note.
• The CHANGELOG.

Not done

• neuroglancer.py's unit tables stay, because worldarray's unit table has no kilometre, and a test covers km.
• The parts of ngff.py that read scene:// stay. They handle OME 0.6 drafts that worldarray doesn't.