Fix MLT polygon winding artifacts in maplibre-gl-js

[trasch]

Problem

maplibre-gl-js renders triangular fill artifacts on polygons-with-holes (e.g. oceans) when displaying PostGIS-sourced MLT tiles at low zooms (visible at z7).

Root cause

Two independent issues combine to produce the artifacts:

1. Inconsistent ring winding in tile space (MLTEncoder)

maplibre-gl-js's classifyRings regroups a flat ring list into polygons by using the first ring's signed area to decide exterior-vs-hole. PostGIS ST_ClipByBox2D / ST_MakeValid can leave rings with inconsistent winding in tile space, so the first ring of a polygon-with-holes may end up with the wrong sign, gets classified as a hole, and the polygon is split — producing triangular fill artifacts.

2. Morton/Hilbert dictionary vertex reordering (C++ MLT encoder)

With real PostGIS-sourced tiles (many features per layer, mixed geometries, vertices spanning the buffer zone with negative coordinates), the upstream C++ MLT encoder's space-filling-curve dictionary path produces tiles whose decoded ring0 starts at a dictionary-sorted anchor rather than the encoded first vertex, and whose winding is flipped — defeating the winding fix above.

Reproduced on a real served tile (z7/x65/y47, oceans layer, 67-ring MultiPolygon):

• Encoder input ring0 first coord: (0, 2735)
• Decoded ring0 first coord: (-405, 2730), sign + (wrong)

Fix

1. normalizePolygonTilespace in MLTEncoder.swift

Reverses each ring so the exterior has negative area and holes have positive area (maplibre's y-down convention), consistently across every polygon before encoding.

2. Disable Morton/Hilbert in Bridge.cpp

Forces the plain zigzag-delta vertex path that preserves vertex order and winding. Disabling enableMortonEncoding alone is not sufficient (the Hilbert path is gated only by validCoordinateRange), so geometryEncodingOption is set to PLAIN to take the early-return plain branch in encodeGeometryColumn. Cost: slightly larger tiles.

Upstream issue: https://github.com/maplibre/maplibre-tile-spec

Also

• Updates the vendored maplibre-tile-spec submodule to the latest main (87878a90, +28 commits — none touch the encoder).
• Adds a regression test multiPolygonManyRingsWithNegativeCoordsRoundtrip.

Verification

• All 261 existing tests pass.
• The real served tile (z7/x65/y47) now decodes correctly: oceans 67-ring MultiPolygon ring0 = [0, 2735] sign - (was [-405, 2730] sign +).
• Artifacts gone in the maplibre-gl-js debug map.