fix: clamp Web Mercator mesh for south-up affines

[kylebarron]

This is a follow up to #574. I was working with @gadomski on a Zarr visualization with bounds [-180, -90, 180, 90].

I expected it to be solved by #574, but it turned out it wasn't touched by that because that PR expected an origin in the top-left (i.e. with a negative e in the affine), whereas this dataset has an origin in the bottom-left.

Before:

After:

---

Problem

createInitialWebMercatorTriangulation (added in #574 to fix degenerate near-pole triangles for global EPSG:4326 imagery, issues #182 / #351) assumes the tile is north-up — it takes north = topLeft; south = bottomLeft and bails via the north - south <= LAT_EPSILON guard.

Grids with a positive-e affine are stored south-up — common for GRIB/IFS-derived reanalysis data. For these, projectedTileCorners puts the southern edge at topLeft, so north - south is negative on every tile, the guard trips, and the clamp is silently skipped. The pole then gets meshed: makeClampedForwardTo3857 collapses the polar vertices onto a single Y, producing degenerate triangles that never converge:

RasterReprojector: mesh refinement did not converge after 10001 iterations
  (maxError=0.125, currentError=19.79...)

…along with visible pole misalignment (high-latitude pixels drift poleward of the basemap).

Fix

Derive the band orientation-agnostically. Latitude varies linearly along the seed's v axis as lat(v) = top + v * (bottom - top), so intersect that segment with [-maxLat, maxLat] by solving for v at each limit and taking the overlapping interval via min/max. This keeps v anchored to the texture's top row (topLeft), so the band is correct whether latitude increases or decreases with v.

North-up behavior is unchanged (verified — all pre-existing tests pass without modification).

Notes

• The previous "south-up tile" path returned undefined (full mesh) by design; that's exactly the case that needed handling, not skipping.
• A naive "normalize corners to north-up, keep the formula" approach does not work: the seed's v-axis is pinned to the texture's top row, not to north, so swapping corner values mirrors the band vertically unless the output is also flipped (v → 1-v). Deriving the band directly avoids that footgun.

Tests

Adds two regression tests:

• a global south-up tile (topLeft = -90, bottomLeft = 90) → clamped band, previously undefined;
• a south-up tile clamped on only the south edge (topLeft = -90, bottomLeft = 80) — asymmetric, so it would catch a mirrored band.

vitest run on the suite: 7/7 pass. tsc --noEmit clean.

🤖 Generated with Claude Code