feat: [dsm-rgb-landcover-tirol] DSM geometry + orthophoto/land-cover albedo for the Tirol demo

[JustCreature]

What

The Tirol demo view now renders a photo-real world: trees and buildings as actual geometry (BEV ALS DSM drives the fine tier), the 20 cm BEV orthophoto draped as albedo near/mid range, and the land-cover raster as a material layer (water gets a flat tint + sun glint). Everything is config-driven via config.toml — no UI changes — and every new layer degrades gracefully when its file is missing, so the demo keeps working on machines without the mosaics.

Source	File	Role
ALS DSM, 1 m, EPSG:3035 (single IFD, LZW 64×64, NoData = +3.4028235e38)	big_size/ALS_DSM_CRS3035RES50000mN2650000E4450000.tif	fine-tier surface geometry
Orthophoto RGB, 0.2 m, EPSG:31255 (ModernJPEG YCbCr, mask IFDs, 12 overviews)	color/2019470_Mosaik_RGB.tif	albedo
Land cover, 0.2 m, EPSG:31255 (Gray(4) nibble-packed, 10 overviews)	color/2022470_Mosaik_LC.tif	material codes (water/veg/buildings)

How

dem_io

• NoData generalization (bug fix): the DSM's GDAL_NODATA = +3.4028235e38 sailed through the old NaN / < -1000 / == 0 predicate and parsed as a valid 3.4e38 m elevation, poisoning max_terrain_h and the march's sky exit. extract_window, parse_geotiff_auto and the overview-cache builder now share is_nodata_value(), which reads the GDAL_NODATA ASCII tag (42113) and rejects the float-max family.
• composite_surface_over(canvas, overlay, feather_px): paints valid DSM cells onto the fine-tier DTM with a 6 px BFS feather at the ALS coverage edge. The DTM stays the canvas so fine-tier coverage never shrinks where the DSM window is clipped — deliberately not fill_nodata_from_base with swapped roles, whose 30 px outward blend would sand real canopy down to bare earth along every coverage edge.
• extract_color_window (color.rs): camera-centred window from the JPEG-YCbCr ortho (the tiff crate's zune path returns raw YCbCr; BT.601→RGB happens on the worker) + land cover nearest-resampled onto the same grid. Output is one RGBA8 buffer: RGB = albedo, A = material code ladder (water 255 / high veg 192 / med veg 128 / building 64 / other 0 — codes ≥ 64 apart so bilinear boundary mixes degrade into the neighbouring band instead of aliasing to an unrelated material).
• ifd_overview_levels: NewSubfileType-filtered IFD walk. The RGB mosaic interleaves per-dataset transparency-mask IFDs with its overview chain, so a raw scale index (à la ifd_scales) would mis-pair scales with IFDs.
• The LC mosaic turned out to be Gray(4) — 4-bit nibble-packed chunks, unpacked in extract_u8_window. Class codes are undocumented; they were pinned empirically with the new inspect_color example against single-class spots: Achensee deep centre = 100 % class 5 (water), Hintertux glacier = 100 % class 2 (ground/ice), forest slope = 96 % class 1 (high veg), valley meadow = 88 % class 6 (low veg), Mayrhofen centre = 44 % class 4 (buildings).

render_gpu

• Two mipmapped Rgba8Unorm ortho windows join the canonical bind group as entries 20–23, each with its own origin/extent/cos_rot/sin_rot uniform block (CameraUniforms 224 → 288 bytes; the std140 layout test pins the new offsets). Deliberately non-sRGB — the whole shader works in display-space 0–255.
• upload_ortho_fine/close follow the established drop-first reload cycle (placeholder swap → rebuild_bind_group → poll(Wait) → tracked alloc) so reload peak stays max(old, new); set_ortho_*_inactive eagerly frees. All allocations go through the vram tracker.
• gen_rgba_mip_bytes box-filters RGB but takes nearest alpha — averaging the discrete material codes would fabricate classes at every mip.
• Shader: hit-point albedo = procedural ramp → close ortho → fine ortho, with the same 500 m edge smoothstep and per-window rotation as the height tiers, mip level from the existing distance heuristic. Where ortho wins, the brightness floor is lifted (mix(brightness, clamp(brightness, 0.55, 1.0), w·0.7)) — the photo carries the capture day's baked sun shading, and full diffuse on top double-shades north faces. Water (A ≥ ~0.9) blends toward a lake tint and adds a mirror-direction sun glint.

viewer / launcher

• DemoViewConfig gains surface_tile_paths / ortho_tile_paths / landcover_tile_paths (serde-defaulted to the BEV mosaics, resolved against tiles_dir; old configs parse untouched).
• The fine worker extracts a DSM window with the same centre/radius as its DTM window and composites it on — normals, DDA self-shadowing and bicubic smoothing then all run on the composite, so trees and buildings get geometry and shadows with zero shader changes for this feature.
• StreamingTier is now generic over its payload (TierCentre trait), so two new ortho workers (fine ≈ 0.8 m/px, close ≈ 6.4 m/px; radii per the new ortho_radii(VramClass) presets) reuse all drift/trigger bookkeeping. Steady-state ortho cost: ≈ 170 MB (Mid), ≈ 330 MB (High), ≈ 17 MB (Low).
• Base reloads deactivate + invalidate both ortho windows (their origins are base-relative). The OOM ladder gains step 0: drop both ortho textures before touching any height tier — cosmetic-first degradation.
• T toggles the ortho drape; R (debug reload) now also invalidates the ortho tiers.

Post-review rendering fixes

• Buildings & trees no longer look like smooth mountains. The fine tier now carries the DSM, but the Catmull-Rom bicubic normal smoothing (built for bare-earth ridgelines within smooth_radius_m) was rounding every building wall and canopy edge into a hill. The smoothing is now gated by world-slope magnitude — it blends back to the raw per-texel Sobel normal where the gradient is wall-steep (smoothstep(1.0, 2.5) m/m), so man-made faces and canopy edges stay crisp while gentle terrain stays smooth. (Note: true hollow geometry under a tree canopy isn't representable in a single-valued heightfield — this makes trees/buildings read as distinct sharp objects, which was the visible problem.)
• Black spot where the ortho coverage ends. The BEV mosaic fills areas with no photo coverage with pure black (0,0,0), which draped onto the terrain as a black hole. Each ortho sample is now gated by luminance (smoothstep(2, 16)) so near-black no-data fades out and the procedural terrain colour shows through; real ortho pixels stay well above the threshold, making the coverage edge a clean one-texel fade.

Verification

• 172 workspace tests + 33 bin tests pass (incl. real-device GPU suites under lavapipe: ortho upload→inactive lifecycle, vram accounting returns to baseline), clippy --all-targets -D warnings clean.
• Gated tests against the real multi-GB tiles (skip cleanly elsewhere): mask-IFD filtering on the RGB pyramid, DSM float-max NoData mapping, lake-window water-material + YCbCr chroma checks, plus a full offscreen smoke render of the pipeline through the production placement math.
• Live run (M4 Max, Mid preset): fine ortho 5000² @ 0.8 m/px decodes in 1.3 s on its worker, close window 2978×4306 @ 6.4 m/px in 0.7 s; uploads are steady-state write_texture calls; zero wgpu validation errors.

Coverage

The color/ortho readers and placement math only ran end-to-end against the multi-GB mosaics, which CI lacks, so they showed uncovered and dropped patch coverage below the 60 % gate. Synthetic-fixture tests now exercise the same production paths without the real files: patch coverage 56 % → 81 % (measured under the CI condition, tiles absent).

File	Before	After
dem_io/src/color.rs	18 %	95 %
dem_io/src/geotiff.rs	40 %	98 %
viewer/tiers.rs	28 %	69 %

viewer/mod.rs stays uncovered by design — it's the winit/GPU event-loop glue the project's testing policy excludes.

Note: the default Hintertux camera sits ~200 m west of the DSM coverage edge (starts ~11.69°E) — fly east into the Tux/Ziller valleys (Mayrhofen is the showcase) to see buildings and forest canopy in relief.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 72.16804% with 371 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/viewer/tiers.rs	40.75%	221 Missing ⚠️
src/viewer/mod.rs	0.00%	118 Missing ⚠️
crates/dem_io/src/color.rs	89.63%	23 Missing ⚠️
crates/render_gpu/src/scene/tiers.rs	98.00%	4 Missing ⚠️
crates/dem_io/src/heightmap.rs	97.65%	3 Missing ⚠️
crates/dem_io/src/geotiff.rs	95.23%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!