surtgis v0.14.3

Patch release. Closes ROADMAP item G ("Martini RTIN terrain
mesher", backlog with pickup-trigger "paper figure has flat regions
wasting triangles"). After this release the original 2026-06-06
survey backlog is down to one item: I (3D Tiles export), still
backlog pending a Cesium-user trigger.

Added

• relief_3d::martini — new module. Port of the Mapbox Martini
  algorithm (Agafonkin 2019, https://github.com/mapbox/martini).
  Adaptive triangulation of heightmaps by curvature: more triangles
  where the surface bends, fewer on flat regions. Three building
  blocks:

  • Martini::new(grid_size) — precompute the per-triangle
    coordinate table for a (2^n + 1) × (2^n + 1) grid. O(N²).
    Reuse across many tiles of the same grid size.
  • Martini::tile(heights) — build the error pyramid in O(N²),
    bottom-up.
  • Tile::mesh(max_error) — extract a triangle mesh whose
    worst-case vertical error is bounded by max_error. Smaller
    error → more triangles; larger → fewer.

• relief_3d::lod::MartiniMeshParams + QuadtreeMesh::from_dem_martini
  — drop-in alternative to from_dem that emits one Martini-meshed
  ChunkLod per chunk instead of stride-decimated LOD levels. Same
  skirts, vertex compression (VertexC), and LodPool lazy upload
  as the quadtree path; the render pipeline is unchanged. lod_levels
  is set to 1 on the returned mesh (Martini's error metric is
  already curvature-adaptive — distance-band LOD on top of it would
  be redundant).

  Default MartiniMeshParams { chunk_cells: 64, max_error: 0.002 }
  — same chunk size as the quadtree default, error tuned for DEMs
  normalised to [0, 0.45] scene units (≈ 0.5 % of the vertical
  range).

Changed

• QuadtreeMesh::select() now clamps the chosen LOD index to
  chunk.lod_data.len() - 1. The quadtree path was unaffected
  because every chunk had the same lod_levels; the Martini path
  needs the clamp so a select_lod choice of band 2 or 3 doesn't
  panic when the chunk only owns LOD 0.

Verified

• cargo test --release -p surtgis-relief-3d: 22/22 green (was 15
  pre-sprint; +5 Martini core + 2 integration).
• cargo build --release -p surtgis-relief-3d clean.
• Flat 129² DEM, 4 chunks: quadtree LOD 0 = 32 768 triangles vs
  Martini = ≤ 3 200 triangles. >10× reduction on flat regions
  — exactly the use case documented in the ROADMAP pickup trigger.
• Bumpy DEM at max_error = 0: Martini reproduces the quadtree
  LOD 0 grid within ±50 %.

Constraints / known limits

• chunk_cells must be a power of two (Martini RTIN requirement).
  Default 64 is the natural fit.
• DEMs whose side isn't a multiple of chunk_cells have their rim
  chunks fall back to the stride-1 quadtree surface mesh. Internal
  chunks Martini-adaptive.
• Build time slower than from_dem — the error pyramid + tree walk
  is O(N²) per chunk. Acceptable for static-camera paper figures
  and headless renders; not ideal for very large DEMs where stream-
  load time dominates.

ROADMAP status after v0.14.3

• G. Martini RTIN — done with this release.
• H. TopoToolbox channel network depth — done in v0.14.2.
• I. 3D Tiles export — backlog, pickup trigger: a user with a
  Cesium pipeline asks for it.