[TommySquared]

Summary

GeoTIFF overlays were drawn in the wrong place whenever the raster was not already in Web Mercator. This warps each rendered image onto a Web Mercator grid before it is sent to the browser, so overlays line up with the basemap, with vector layers, and with each other.

The problem

process_geotiff_rgb / process_geotiff_singleband rendered the PNG on the file's native grid and reported its lon/lat bounding box (transform_bounds). The browser places an image source by those four corners and stretches it linearly in Web Mercator. For a UTM or equal-area raster this rotates and stretches the image:

Raster (UTM)	Worst misplacement	Typical
180 km wide at 68°N	9.5 km	3.1 km
43 km wide at 40°N	680 m	340 m
~10 km wide	140 m	75 m

Two rasters in different projections (e.g. UTM and EPSG:6933) are misplaced differently, so they visibly fail to overlap. Vector polygons are unaffected, so imagery appears to slide against them.

The fix

_to_web_mercator() takes the finished RGBA from _process_rgb / _process_singleband and reprojects it with nearest-neighbour resampling onto a Web Mercator grid. Pixels outside the source footprint get alpha 0. It returns exact lon/lat bounds, which form an axis-aligned box in both CRSs, so four-corner placement is exact.

• Nodata, mask and alpha handling still runs on the native grid, unchanged (fix: honor mask/alpha bands and mask RGB nodata only when all bands match #16 behaviour preserved).
• EPSG:3857 input is passed through unresampled.
• Rasters reaching past ±85.05° keep the previous placement, with a warning.

Tests

server/tests/test_geotiff_placement.py:

• Renders an 80 × 60 km UTM raster at 68–69°N through the public API, places the PNG the way the browser does, and checks three marker pixels: all land within one 100 m source pixel (36–75 m). The previous placement put the worst 1.3 km off; a companion test measures that failure on the same raster.
• The RGB footprint is transparent outside and opaque inside.
• Web Mercator input is not resampled.

The rest of the suite is unchanged relative to main (same pre-existing failures/errors in the MCP and screenshot harness).

🤖 Generated with Claude Code