native-scene: don't drop world meshes whose VB size isn't a stride multiple

BuildItemFromMesh required vb_bytes to be an exact multiple of the vertex
stride and discarded the entire mesh otherwise. Community custom maps
(ArenaBuilder-built worlds) ship vertex buffers with a few trailing bytes
past the last vertex - e.g. 43968 bytes at stride 36 leaves 12 over - so
their terrain was rejected wholesale and the world rendered with no ground.

Measured on one such map: 608-656 "world item build FAILED" lines per
session (rate-limited, so the real count is higher) against 0 on stock
content.

Relax the requirement to "there is room for at least one vertex". Every
consumer already derives the vertex count as vb_bytes / stride with integer
division - BuildMeshBuffers does exactly that and sizes the GPU buffer from
it - so the trailing bytes were always ignored. The exact-multiple test was
redundant with handling that already existed downstream.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/skate3_native_scene.cpp

@@ -2107,8 +2107,16 @@ bool BuildItemFromMesh(uint8_t* base, uint32_t mesh, DrawItem& item) {
   item.vb_bytes = BSwap32(vb_words[2]);
   item.ib_addr = BSwap32(ib_words[0]) & 0xFFFFFFFC;
   item.ib_count = BSwap32(ib_words[2]);
+  // vb_bytes need only COVER a whole number of vertices, not equal one
+  // exactly. Community custom maps (ArenaBuilder-built worlds) ship vertex
+  // buffers with a few trailing bytes past the last vertex - e.g. 43968
+  // bytes at stride 36 leaves 12 over - and requiring an exact multiple
+  // rejected the entire mesh, so the world rendered with no ground at all.
+  // Every consumer already derives the vertex count as vb_bytes / stride
+  // (integer division), so the trailing bytes are simply unused; all that
+  // is really required is room for at least one vertex.
   if (item.vb_addr == 0 || item.ib_addr == 0 || item.vb_bytes == 0 ||
-      item.ib_count == 0 || item.vb_bytes % item.stride != 0) {
+      item.ib_count == 0 || item.stride == 0 || item.vb_bytes < item.stride) {
     g_rej_geom.fetch_add(1, std::memory_order_relaxed);
     return false;
   }