fix(deck.gl-geotiff): render band-separate (PlanarConfiguration=2) images

[thatbudakguy]

What

getTileData in createUnormPipeline throws for any COG stored with PlanarConfiguration=2 (GDAL's INTERLEAVE=BAND):

if (array.layout === "band-separate") {
  throw new Error("Band-separate images not yet implemented.");
}

Because it throws per tile inside the tile loader, a COGLayer over such a file draws nothing at all and the only sign is onTileError firing once per tile.

This interleaves the tile instead. toPixelInterleaved in @developmentseed/geotiff already does exactly this — the array helpers just weren't exported from that package's index, although their types were.

Why this came up

A librarian at UMN reported a scanned map that wouldn't display in a viewer I'm working on that uses this library. The file is a 4-band RGBA JPEG COG with an alpha collar, written with -of COG -co COMPRESS=JPEG -co INTERLEAVE=BAND, and the reader handles it fine — only the render pipeline refuses it.

Changes

• packages/geotiff — export toPixelInterleaved, toBandSeparate, reorderBands, packBandsToRGBA and PackBandsToRGBAOptions from the index. The RasterArray* types were already exported; the functions that produce them weren't.
• render-pipeline.ts — interleave a band-separate tile before the RGB padding step, rather than throwing. Doing it first means the padding and texture-format logic below see the same shape they would for a pixel-interleaved image, so 3-band INTERLEAVE=BAND is fixed alongside 4-band.
• geotiff.ts — addAlphaChannel carried a second copy of the same throw, and unlike the pipeline it's part of the public API. It now interleaves too. Its return type narrows from RasterArray to RasterArrayPixelInterleaved, which every path through it has always satisfied; that's what lets the narrowing in getTileData survive the call.
• examples/vermont-cog-comparison — the narrowed return type makes that example's layout === "band-separate" guard statically unreachable, so it's removed.

Tests

• tests/render-pipeline.test.ts — getTileData interleaves a 4-band tile into one rgba8unorm texture, and interleaves and pads a 3-band one. The tile is handed in directly rather than read from a fixture: the only band-separate fixture is int8_3band_zstd_block64, and inferRenderPipeline doesn't build a pipeline for non-unsigned samples. Happy to add a uint8 band-separate fixture to geotiff-test-data as a follow-up if you'd like the coverage to be end-to-end.
• tests/geotiff.test.ts (new) — addAlphaChannel over band-separate RGB and RGBA input, plus the existing pixel-interleaved path.

I also checked the interleave against real data outside the suite: for a band-separate DEFLATE COG built from the reported file, toPixelInterleaved output is byte-identical to geotiff.js's independent readRasters({ interleave: true }) across four full 256×256 tiles (~1M samples), with the alpha channel spanning 0–255 rather than a constant. And with this change the reported COG renders correctly in ogm-viewer, transparent collar included.

pnpm build, pnpm -r --filter='!docs' typecheck, biome ci . and the geotiff / deck.gl-geotiff / deck.gl-raster suites all pass locally. The 17 failures in packages/geotiff/tests/integration-rasterio.test.ts are the .npy reference fixtures that CI generates with pixi run generate-npy; I don't have pixi here, and they fail the same way on a clean checkout.

Note on approach

Uploading one texture per band and sampling a sampler2DArray — the idea in #159 — would avoid the per-tile copy this adds, and is the better long-term shape. It needs the render pipeline to know how many textures a tile has, which is a larger change. This is the fallback path packBandsToRGBA's docstring already describes ("useful as a fallback path when a single 4-channel texture upload is preferred over one texture per band"), and it makes these files render today. Glad to rework it if you'd rather go straight to the texture-array version.

🤖 Generated with Claude Code

[thatbudakguy]

Apologies for the largely AI-generated PR; this came up in some work I was doing and seemed like a straightforward fix.

Let me know if these changes aren't desirable!