feat: Initial Globe view support

[kylebarron]

Screen.Recording.2026-05-22.at.12.02.20.PM.mov

---

Written by Claude on behalf of @kylebarron

What I am changing

Adds prototype globe-view support to deck.gl-raster. RasterTileLayer-based
layers (COG/Zarr) can now render on a 3D globe — deck.gl _GlobeView, or
MapLibre projection: 'globe' with deck.gl interleaved — not just the flat Web
Mercator map. Includes a new cog-globe example.

Deliberately scoped as a prototype: it renders real data end-to-end and
exercises the full pipeline, with the harder accuracy work called out and
deferred (see Limitations). Design + deferred work:
dev-docs/specs/2026-05-21-globe-view-design.md.

How I did it

• Globe tile selection. Implemented the globe computeBoundingVolume
  (was assert(false)): sample reference points (REF_POINTS_11), reproject to
  WGS84, and project onto the sphere via viewport.projectPosition for the
  frustum-culling oriented bounding box; keep a Web-Mercator-world AABB for the
  dataset pre-filter and the LOD latitude. The bounding-volume cache is rebuilt
  when the viewport projection mode switches (globe↔mercator).
• lng/lat-direct rendering. Globe renders in coordinateSystem: 'lnglat'
  (mesh vertices in WGS84 via projectTo4326) — no manual common-space mapping.
  Collapsed SimpleMeshLayer's composeModelMatrix branch in the
  MeshTextureLayer vertex shader to a single direct-projection path (the old
  branch projected lng/lat degrees as if they were meters).
• Z-fighting with the basemap. The raster mesh is coplanar with MapLibre's
  globe sphere in the shared interleaved depth buffer. A depth bias doesn't help
  with maplibre's globe depth encoding; instead depthCompare: 'always' +
  back-face culling (cullMode: 'back' for this grid's winding). See
  visgl/deck.gl#9592.
• LOD fix. The meters-per-pixel latitude must come from the mercator-world
  bounds, not the 3D OBB center (globe-common → worldToLngLat returns ~-89°
  near the Mercator singularity → meters/px far too small → the traversal loads
  the finest tiles across the whole globe when zoomed out).
• Anti-faceting scaffold (throwaway). Globe builds a uniform per-tile grid
  instead of the adaptive Delatin mesh, whose reprojection-error metric is blind
  to sphere curvature (an EPSG:4326 source yields 2 triangles that chord through
  the sphere).
• Merged main, reconciling with fix: traverse tiles across world copies (#517) #518 (world copies); world-copy passes are
  gated on viewport.subViewports, so they don't run on a globe.
• Documented the globe path in dev-docs/coordinate-systems.md.

How you can test it

pnpm install
pnpm --filter @developmentseed/deck.gl-raster build
pnpm --filter @developmentseed/deck.gl-geotiff build
pnpm --filter deck.gl-cog-globe-example dev

Confirm: the basemap renders as a 3D globe, the COG drapes onto the sphere and
lines up with the basemap (source selector to try several COGs), no z-fighting,
back-of-globe tiles are occluded, and LOD tracks zoom (coarse when zoomed out,
finer when zoomed in).

Unit tests: pnpm --filter @developmentseed/deck.gl-raster test — globe
bounding volume, cache rebuild on projection switch, and a globe LOD regression
test.

Limitations (deferred — tracked as follow-up issues)

• Spherical reprojection correctness (Globe view: curvature-aware spherical reprojection (remove uniform-grid scaffold) #566) — the adaptive reprojector's
  error metric is blind to sphere curvature; the uniform-grid scaffold is a
  stopgap.
• LOD limb foreshortening (Globe view: screen-space-error LOD (fix limb-foreshortening over-fetch) #567) — the mercator meters-per-pixel model
  ignores globe foreshortening, so tiles near the limb slightly over-fetch
  (never under). A screen-space-error metric would be exact.
• zarr-globe example (Add zarr-globe example #568) — not included yet.
• Cutline on globe is disabled; expected to come via ClipExtension
  (Remove CutlineBbox module in favor of upstream ClipExtension? #561).

Related Issues

• Part of Support for rendering in globe view #82 (globe view)
• Builds on fix: traverse tiles across world copies (#517) #518 (multi-world-copy tile traversal)
• Follow-ups: Globe view: curvature-aware spherical reprojection (remove uniform-grid scaffold) #566 (spherical reprojection), Globe view: screen-space-error LOD (fix limb-foreshortening over-fetch) #567 (screen-space LOD),
  Add zarr-globe example #568 (zarr-globe); cutline-on-globe depends on Remove CutlineBbox module in favor of upstream ClipExtension? #561 (ClipExtension)