City Walk/WissenZum StadtspaziergangViewer
ImpressumDatenschutz
Leitfaden
  • So funktioniert der Stadtspaziergang
  • Woher die Daten kommen
  • Quellen nach Stadt
  • Wie der Viewer zu einer Angabe kommt
  • Der Weg der Daten
  • Bedienung
  • Glossar
Guide (English)
  • How the city walker works
  • Where the data comes from
  • Sources by city
  • How the viewer comes by what it shows
  • From download to browser
  • Using the viewer
  • Glossary
Entwicklung (English)
  • docs/ — the project knowledge base
  • Rendering — how the data becomes pixels
  • Data pipeline — the bakes, the build step and the artifact contracts
  • Architecture decision records
  • Data → feature provenance
  • Transformation catalog
  • Portability — rendering any location, with whatever data it has
  • Writing a walk
  • Routines: the twin's scheduled agents
  • Implementation plans, backlog and audit history
  • What the data streams carry that the walker does not use yet
Contents
Leitfaden
  • So funktioniert der Stadtspaziergang
  • Woher die Daten kommen
  • Quellen nach Stadt
  • Wie der Viewer zu einer Angabe kommt
  • Der Weg der Daten
  • Bedienung
  • Glossar
Guide (English)
  • How the city walker works
  • Where the data comes from
  • Sources by city
  • How the viewer comes by what it shows
  • From download to browser
  • Using the viewer
  • Glossary
Entwicklung (English)
  • docs/ — the project knowledge base
  • Rendering — how the data becomes pixels
  • Data pipeline — the bakes, the build step and the artifact contracts
  • Architecture decision records
  • Data → feature provenance
  • Transformation catalog
  • Portability — rendering any location, with whatever data it has
  • Writing a walk
  • Routines: the twin's scheduled agents
  • Implementation plans, backlog and audit history
  • What the data streams carry that the walker does not use yet
Entwicklung (English)

Transformation catalog

The ledger of every data → look transformation we've built, tried, or planned — so a decision made in one chat thread is not re-litigated or re-discovered in the next. This is the in-repo, version-controlled successor to scattered notes; append here when you ship, shelve, or reject a transformation.

Status legend: ✅ active (shipped) · 🧪 experimental (prototype / behind a slider, not load-bearing) · 📋 planned (designed, not built) · 🗃️ discontinued (tried or considered and rejected — kept so we don't retry it blindly).

Method badges: each active or experimental entry says how the viewer comes by what it draws — ❝taken as the source publishes it, =computed by a fixed formula over measured values, ◎detected by pattern recognition (thresholds, matching, image analysis: it can be wrong), ≈assumed where nothing is known for the thing (a default, an average, a design choice). The badges are defined in lib/city/methods.ts, explained for readers in the guide's How the viewer comes by what it shows, and the inquiry card shows the same four per line. A new or changed entry carries its badge.

Each entry records inputs, what it does, source preference / fallback (what it degrades to when its best input is absent — see portability.md), and where it lives. The why behind the load-bearing choices is in the ADRs; the attribute → visual-variable codebook is in rendering.md.


✅ Active

Geometry & ground

  • Terrain TIN =computed (the fine level, every tile, ±0.15 m) — DGM1 read at its native 1 m (2000²) → the terraces lifted and the ground lowered under the stairs (below; the stairs ±0.15 m deeper, so no approximating triangle pierces a tread) → Delatin error-bounded TIN (every source grid point within the tolerance of the mesh) → the fine level's glTF (terrain_<tile>_l0.glb.gz, reordered for meshopt, extras.tin), baked by scripts/bake-terrain-tin.ts + scripts/bake-tiles.ts (tinTerrainMesh) inside prepare-data.ts from the committed GeoTIFF (no new committed input). The viewer indexes its triangles for heightAt (lib/city/terrain-tin.ts TinIndex, skirt skipped) and gives the water an up-facing normal twin (terrain-layer.ts tinHeightAt, waterGeometryOf). No wall conflation on the TIN: the earth-retaining ribbons baked beside it snap to the step the ground measures (lib/city/wall-snap.ts via lib/city/walls.ts: steepest metre within 6 m of the OSM line, running medians along the wall over the neighbours that found the same step (crests within 1.5 m, SAME_STEP_M; 2026-10-21: on a terraced slope a median over two terraces' steps stood one block from the lower foot to the upper crest, up to 15 m tall on walls tagged 3 m — Meißen's Burgberg, read from above as boulders on the vineyard), face just in front of the ramp foot + a coping cap back to where the ground reaches its level, per column, the ribbon broken between runs whose caps differ by more than 1 m (stepRuns) and each run's cap ends smoothed within it (cap triangles deeper than 6.5 m: Meißen 586 → 193, Dresden's 33410_5656 391 → 144, Leipzig's 33316_5688 1 323 → 641, whose faces over 10 m went 50 → 0) — a cap ending at the median crest left the DGM's steep ramp showing behind it as a row of jagged facets wherever the ramp ran wider; where it finds no shelf within 5 m it drops a back face to the ground, and the foot reaches the lowest ground along the face — ADR 0030). The coarse 512² level keeps the grid and the conflation; a DGM with NoData keeps the grid for its fine level too (ADR 0030). In the 3D Tiles bake (2026-09-25): 0.30 / 0.39 / 0.49 / 0.31 M triangles (33410_5656 / 33410_5658 / 33412_5658 / 33412_5656) for 1.70 / 1.68 / 1.99 / 1.61 MB gzipped — the 1024² grid it replaced was 2.1 M triangles and 2.30 / 2.15 / 2.45 / 2.19 MB (20–27 % more); about 8 s per tile to bake. The earlier prototype (a private header + delta-plane payload, ±0.25 m on the neighbours, before the move to glTF) is where the numbers below come from. Study (scripts/terrain-study/, reproducible from the committed DGM + the gitignored LSC LAZ; held-out = a seeded 10 % of the 34 M class-2 ground returns, 3.4 M points; 35 profiles across the 7 tallest OSM retaining/city walls; DGM1 was made from all returns, so its numbers are slightly optimistic):

    variantmain-step width (m, median)profile RMSE vs laser (m)RMSE / p95 tile (m)RMSE / p95 20 m wall band (m)trianglesgz
    laser ground (0.25 m bins)0.75—————
    V0 shipped 1024² grid2.850.790.095 / 0.1140.43 / 0.802.09 M1.08 MB
    V0 + client conflation (what the viewer shows)1.751.470.140 / 0.1220.76 / 1.622.09 M1.08 MB
    V1 DGM1 2000² grid1.700.290.048 / 0.0530.23 / 0.188.0 M3.7 MB
    V2 laser ground 0.5 m grid1.200.160.046 / 0.0460.22 / 0.1032 M12.8 MB
    TIN of V1 ±0.15 m (the prototype)1.300.230.066 / 0.1080.22 / 0.150.30 M1.04 MB
    TIN of V1 ±0.10 / ±0.25 m1.30 / 1.400.23 / 0.220.056 / 0.0910.22 / 0.220.55 / 0.14 M1.79 / 0.49 MB
    TIN of V2 ±0.25 / ±0.10 m1.00 / 1.100.17 / 0.160.089 / 0.0540.23 / 0.220.23 / 1.03 M0.85 / 3.72 MB
    TIN of V1 + conflation burned in ±0.10 m0.601.400.135 / 0.0840.78 / 1.520.55 M1.78 MB

    Client cost (Bun, decode → mesh → normals, then the idle-time BVH): grid ~300 ms + 470 ms BVH → TIN ±0.15 m ~115 ms + 90 ms BVH; GPU geometry ~50 → ~9 MB, and the terrain, water and mist sheets each draw a seventh of the triangles. Findings: (1) most of the "wall smear" is our resample, not the data; (2) the laser-scan DTM beats DGM1 only at the sharpest quay walls (0.6–0.75 m vs 0.9–1.05 m) for 2–4× the triangles and a committed LSC-derived artifact — not worth it; (3) the breakline burn sharpens steps but triples the wall-band error: it flattens terraced walls (the Jungfernbastei climbs 108.6 → 117.4 → 119.7 → 121.1 m within 8 m and gets one cliff to 121.1 m) and puts the step on the OSM line, which misses the measured step by −0.5…+1.0 m (4 m at the bastion's south face). Pitfalls checked: LSC class 30 (under buildings) and class 8 (water) are the DGM1 values to the mm (synthetic fill, not measurements); LSC vs DGM1 datum bias 0.000 m; the TIN's seam step against the 512² neighbours is unchanged on average (0.14 m, max 3.1 vs 2.5 m where a wall crosses the edge; the skirt hides it). Visual (real GPU, the primary-only prototype): terraced Jungfernbastei reads correctly (terrace levels kept, trees on the first level visible), quay walls straight and clean from the air; residue: a few sub-metre ground spikes at wall feet and faint shading bands on snapped wall faces, faint facet streaks on the water next to bridge piers. Still open: constrained breaklines (a vertical wall is two vertices at one xy — Delatin cannot, a constrained Delaunay with the snapped wall line could, and would remove the spikes and the cap), NoData support (the bake refuses holes), crease-angle normals. Two independent TINs meet at a seam with different border vertices (T-junctions, each tile's own edge heights); the skirt hides the crack. Neighbour check (real GPU, full block): the Brühlsche Terrasse west of the seam (33410_5656), the Terrassenufer across the 33410/33412 seam, the Altstadt around the Frauenkirche from the air and the northern seam (33412_5658) show no ground step, crack or wall break at the tile edges; the terrace wall reads as one continuous face across the seam. Seam bands (fixed 2026-09-25, grids and TIN alike): a bright band ran along every seam, one border triangle wide — metres on a TIN's flat road — because the 30 m skirt shared the border vertices and dominated their area-weighted normals; the terrain's normals now come from its surface triangles alone (scripts/bake-tiles.ts terrainNormals). Across the Elbe the water and mist sheets drew the skirt as a wall of water; their geometry now leaves out every triangle without area in plan (terrain-layer.ts waterGeometryOf). Checked on a real GPU at the 33410/33412_5656 seam (heights there agree to 5 cm on average).

  • Terrain (glTF, two levels) =computed — DGM1 → a triangulated mesh + a 30 m edge skirt to hide inter-tile seams, at two levels per tile: L1 512² (coarse, geometric error 40 m) replaced near the camera by L0, the TIN above (a 1024² grid until 2026-09-25, and still for a DGM with NoData). The coarse GeoTIFF is resampled at build time (scripts/bake-tiles.ts readDgm, bilinear, NoData → NaN → the quads touching it are left out), the wall breaklines are burned in (below), normals computed, and the mesh written as glTF (KHR_meshopt_compression + KHR_mesh_quantization, pre-gzipped terrain_<tile>_l0|l1.glb.gz). The browser never decodes the DGM: it reads ground height back from the grid vertices (gridElevations), or from the TIN's triangles. scripts/bake-tiles.ts, lib/city/terrain-geometry.ts, terrain-layer.ts (ADR 0024).

  • Ground plate =computed (Bodenplatte) — the whole site's ground at a 62.5 m grid (33² vertices per tile, the edges shared with the neighbours, so it has no seams), each vertex the fine ground's height and the commonest land-cover class within half a cell around it, read across the tile edges (so colour and shading meet without a seam too), coloured by the palette at runtime. One small site file (plate.bin.gz, 49 kB for Dresden's fifteen tiles, 21 kB gzipped) fetched with the tileset; the viewer draws a tile's plate (its inner vertices 1.5 m under their heights, its edges on them, so it meets a landed neighbour without a step) wherever that tile has no terrain level on screen yet, so a tile that never streamed, or that the cache let go of, shows ground instead of the sky on a flight or a turn. lib/city/ground-plate.ts, scripts/prepare/plate.ts bakePlate, ground-plate-layer.ts (plan 076).

  • Far buildings =computed (Gebäude in der Ferne) — each tile's buildings a second time, the same glTF with every vertex its shape does not need dropped (meshopt's simplifier over the written glTF, at most 0.5 m: 1.85 M → 0.29 M triangles and 5.0 → 2.6 MB gzipped on Dresden's spawn tile; 5 cm kept 62 % of them, 20 cm 32 %), shown beyond ≈ 260 m of the tile and from the air, the baked LoD2 replacing it near (FAR_CITY_ERROR 4 m). The shape is the same but for small roof detail; the windows lose the per-vertex ground samples along the walls and the doors, which nothing at that distance resolves. Chosen over LoD1 blocks from the footprints, the plan's first idea: blocks would have cost about as many triangles (≈ 30 a building) and lost the roofs, a jump at every switch. scripts/tile-glb.ts simplifyGlb, lib/city/tileset.ts (plan 076).

  • Squares and islands from OSM ❝taken (Plätze, Inseln) — the DLM draws many squares as one road area: the Albertplatz's pedestrian island, its lawns and fountains were grey carriageway with no kerb. The land-cover bake carves them back out, over road texels only: OSM highway=pedestrian / area:highway (footway, pedestrian, traffic island) areas and fountain basins become built-up (4), leisure=park|garden and landuse=grass|village_green|meadow|flowerbed meadow (1), the lawn winning inside a pedestrian area. Idempotent, so bun run bake <site> --step islands applies it to the committed raster without the raw DLM (2026-09-25: ≈ 0.2–0.7 M texels per tile, Prager Straße and the Altmarkt among them). The legend then carries the OSM credit. pipeline/bake/landcover.py carve_islands.

  • Surface colours ❝taken =computed — Basis-DLM land-cover → a 4096² class-id raster (8-bit, ids 0–9 — 9 arable since 2026-10-21, below —, burned lowest priority first so water wins; pipeline/bake/landcover.py → landcover_<tile>.png + legend). The colours are not baked: lib/city/landcover.ts holds the one pastel palette, and landcover-splat.ts paints it on the GPU once per tile at load (RGB = palette, A = water coverage from a 3×3 tent over class 8) into an sRGB, mipmapped, anisotropy-16 target — what the baked PNG used to be sampled with. A desktop's fine level paints its 4096² raster at half size (2026-10-10, plan 070 step 4, splatScaleFor): a texel is the mean of the 2×2 class colours under it and the water the four tents' mean — the full-size splat's first mip, so only the near field changes, where a boundary now blends over 1 m instead of 0.5 m and the 0.5 m staircase of a diagonal edge is gone. 64 MiB less per fine tile (85 → 21 MiB with mips); SwiftShader plates of the Elbe shore, a road's meadow edge at eye level and the Großer Garten from 50 m showed nothing blockier. The minimap and the site's map picture (site-map.webp: the start page's city card, the /wissen hero) use the same table; a colour change is a look change, not a re-bake (ADR 0010). terrain-layer.ts. Fallback: OpenStreetMap where the provider publishes no Basis-DLM in the Shape profile (Hamburg, Berlin): pipeline/bake/landcover_osm.py writes the same class raster, legend and hedge / tree-row lines from landuse/natural/leisure areas, buildings and amenity areas (as settlement, burned first because OSM nests the green inside it), buffered highways, waterways and rail. On Leipzig's centre tile 71 % of texels agree with the DLM raster — water 95 %, settlement 78 %, roads 55 % (ADR 0037). The OSM land cover (2026-10-21, B15–B26): a one-way main road (motorway to tertiary) at half its class's width (at least 2 m a side: a dual carriageway's two ways had each burned a whole road's width), light rail, subway, narrow gauge and a tram on its own bed (not embedded) burned as railway (off bridges), allotments and cemeteries as meadow, OSM farmland as arable. Hamburg's road land 44.9 → 43.8 ha, its railway 4.7 → 5.1 ha on one tile. Road widths from OSM ❝taken ≈assumed (2026-10-21, pipeline/bake/road_widths.py): a DLM axis without BRF (71–77 % of the road length in Grimma, Meißen and Unna, 95 % in Dresden) took one width per road class (5 m for every Gemeindestraße); now, piece by piece (20 m), it takes the carriageway of the OSM car way running with it (≤ 3 m, ≤ 20°): width:carriageway/width, else lanes (else the DLM's FSZ) × 3 m + 2 m per side of parking on the carriageway (1 m for bays half on the kerb, none for bays on it), else the class's width plus that parking; 1.5–12 m a side. Mean error against BRF: 0.98 / 1.58 / 2.34 m in Grimma / Meißen / Unna, against 1.64 / 1.72 / 2.96 m for the class default (on Dresden's main roads with BRF it would be worse, 2.15 against 1.67 — they keep their BRF). Road land against main, measured on the re-baked tiles with this change's other road rules (tunnels, below): Dresden 415.6 → 518.7 ha over 15 tiles (+25 %), Leipzig +23 %, München +15 %, Grimma +13 %, Unna +9 %, Meißen +5 %; Hamburg −5 % (its land cover is OSM's, no DLM widths: only the tunnels). Arable ❝taken =computed (class 9, 2026-10-21): the Basis-DLM's arable land (veg01 VEG 1010/1011/1013, arable_areas) and OSM farmland are no longer meadow; painted a pale loam (ARABLE_CLASS, [204, 186, 154]), a crop's green where the NDVI sees one (bias 2.5, smoothstep(0.12, 0.42); about 40 % of it reads as crop), faint 0.75 m furrows along the field's long axis (the paving raster's bearing on arable land, surface.py field_frames; terrain-layer.ts arableGround). It counts as lawn for the kerbs and as green in the site's statistics. Grimma 215 ha, Unna 129 ha, Meißen 1.5 ha (measured on the re-baked class rasters). Nothing in a tunnel is burned (2026-10-20): a DLM road axis loses its stretch along a DLM tunnel (ver06 BWF=1870, the rail layer's rule: > 15 m within 2 m, common.above_ground), and every OSM way a terrain bake reads (osm.read_ways: the OSM land cover, paving, road markings, lamps, furniture, allotments) skips tunnel=* and location=underground — a culvert too. Before, the road of a tunnel was painted on the ground above it, with a kerb along both sides: measured against OSM's tunnels, Dresden painted 36 % of its tunnels' 2.2 ha off the surface roads as road with 2.1 km of kerb (Tunnel Bramschstraße, Wiener Platz), Meißen 0.8 ha over the Schottenbergtunnel's hillside (336 m of kerb), Hamburg 57 % of its tunnels (Wallring-, Deichtortunnel), Munich 30 % (Altstadtringtunnel, Paul-Heyse-Unterführung). Nothing piped is burned either (2026-10-21, landcover.py stream_buffers): not a DLM stream axis flagged HDU_X (hatDirektUnten), nor its stretch in a ver06 culvert (BWF 2010 Durchlass, 2011 Rohrdurchlass, 2012 Düker, 2 m round its line or area; a bridge, 1800, does not cut it) — on the six DLM sites every flagged axis lies in such a structure but 46 m in Grimma, and the orthophoto shows no water there; and on an OSM land cover (Hamburg, Berlin) no water area or waterway tagged covered, in a tunnel, or underground or indoors (landcover_osm.hidden_water; read_ways already skipped the culverted lines). It was a straight blue line through Unna's station (the Kortelbach, 1.0 km) and a blue stripe across every road over a culvert: 8.5 km of stream axis in Dresden (23 % of them, 1.70 ha), 8.6 km in München (88 %), 4.9 km in Leipzig (59 %; the Pleißemühlgraben under the square by the Neues Rathaus), 2.0 km in Unna (57 %) and Grimma (30 %; the Thostgrundbach under its street), 1.8 km in Meißen (43 %). OSM's own culverts are not used on the DLM sites: they cut an open stream in Unna.

  • Meadow NDVI tint =computed (Wiesenfärbung) — on class-1 farmland/meadow only, the DOP greenness (ndvi_<tile>.png, read at its mip 2 — 8 m a texel — through a cubic B-spline, bsplineSample; the plain biased mip only past ~6–16 m a pixel; the bilinear mip it replaced drew ~10 m squares that read as a chequerboard from afar) shifts the pastel sage lush deep-green↔dry hay. In the terrain's colour node (meadowNdvi, gated by the meadow class of the raster), at Wiesenfärbung's fixed strength (MEADOW_NDVI, 0.6). The higher-variance NDVI canvas the analysis flagged (meadow carries 1.46× the crown NDVI variance). Absent raster → no-op.

  • Kerbs and lawn edges =computed ≈assumed (Bordsteine, Rasenkanten) — the DLM road class (7) is the surveyed carriageway, so its edge is the kerb line. pipeline/bake/edges.py measures, from the committed class raster, the signed distance to the road edge and to the meadow edge (urban green included) out to ±6 m, box-smoothed so the isolines run straight along a diagonal instead of following the 0.5 m staircase → edges_<tile>.png (2048², two bytes per texel). From the road's 0-isoline it also writes the kerb lines (kerbs_<tile>.geojson, road on the left) where the far side is ground (not water or railway) and the road is not one out of town: none where the nearest OSM carriageway (within 20 m) is a motorway or its link, lies in a speed zone tagged DE:rural/DE:motorway, or posts 70 km/h or more without an urban zone (edges.out_of_town). The road edge is the carriageway's, not the parcel's (2026-10-21): green road land past it (a verge, Urban green below) is meadow in both distances, so its kerb stands at the carriageway — in town only. Every road edge had one; Mapillary at 13 such kerbs on three sites (2026-10-20) showed an edge line, a verge and reflector posts at 12, the 13th a junction's island (Unna). It takes 8.1 km of stone off Unna (5.2 % of its kerbs), 6.6 km off Dresden (0.6 %), 5.2 km off Grimma (4.9 %), under a kilometre off Hamburg, Meißen and Leipzig, none off München. The painted band and its sun shadow (below) follow the same lines (2026-10-21): markings.py sets bit 8 of the road-markings raster's G channel (KERB, LANE_BITS.kerb) within 2 m and a texel's diagonal of every kerb line (less its lowered stretches, below: full_height), and the fine level's shader draws the band only there (ground-detail.ts kerbLines, bilinear over the four texels, so the band fades out over a texel where a kerb ends); before, it ran on along every road out of town, beside water and rails and over a tunnel, for the shader did not know the kerb lines. The coarse level keeps the class rule. (The road-markings PNG grows: Dresden's 33408_5658 50 → 100 kB, the phone's 16.5 → 36 kB. Where a kept kerb ends against rail or water the band may overrun it by up to 3 m.) The terrain bake stands a kerb stone on them in the fine terrain glTF (lib/city/kerbs.ts, 12 cm above the road, 24 cm wide, face toward the road, its top falling to the pavement's level where the stone ends (a verge or bank below the road too), densified where the ground bends between its 2.5 m columns, its face's foot down to a gutter below the road (ADR 0030) — a stone level across stood a second step on the pavement side, whose smoothed DGM lies only a few centimetres above the road; kerb-layer.ts, casts shadow) — the DGM1 smooths the step away and no terrain mesh built from it can hold it. In the fragment pass (ground-detail.ts) the distance draws a pale stone band on the pavement side and a darker gutter on the road side (the coarse level's kerb), a lawn lip with a normal kink, the parking lanes and the paving rows along the kerb. The first cut drew the kerb as a shading-normal step from the class texels alone; on the raster's staircase it read as dashes and odd shadow flecks (replaced 2026-09-25). The stone's own shadow on the road is drawn from the sun: when the sun stands behind the kerb, the strip out to 12 cm · cot(elevation) is shaded — a shadow the shadow map cannot hold (7 cm texels spread by the soft PCF, less the depth bias). On Bodendetail (GROUND_DETAIL). Plan 023. Lowered kerbs (pipeline/bake/kerb_drops.py, 2026-10-21) — 3 cm over the road (flush for kerb=flush|no) ±1.5 m at a crossing node's kerbs either side, where a crossing way meets a kerb, at kerb=lowered|rolled nodes (snapped within 3 m), and across driveways and untagged service ways (half their width, ≥ 1.5 m); kerb=raised|regular keeps it; a 0.5 m ramp at each end (low on the kerb lines, kerbHeightAt). 0.8–3.4 % of the kerb per tile, 0.6–3.8 km. The painted band and its shadow leave the lowered stretches out; the ramps keep them.

  • Path edges =computed ≈assumed (Wegränder, 2026-11-02) — a path (class 6) beside anything but itself, near the camera. The splat paints a 1 m colour texel (a phone's 2 m) as the mean of the class texels under it and the ground read it bilinear, and the paving kinds, the sand, the grey of sealed ground and the urban green switched per class or paving texel: a narrow path at a slant — a forest path, or a park path two texels wide in built-up land, which no baked edge distance covers (the edge raster knows roads and meadows) — came out as a mosaic of diamonds and teeth. Where the 4×4 class texels round a fragment hold a path, ground-detail.ts smoothPaths weights them by a cubic B-spline (a C2-smooth coverage at the class raster's own resolution, whatever the splat's), cuts the edge on its iso line (0.45), wandering ±0.12 m by a world-space noise and frayed over 0.25 m, and paints the palette colours of the path and of the nearest texel beside it across it, the splat's own colour coming back as the coverage falls to nothing (no seam where the 4×4 ends). Beside a path the baked lawn reaches on 1.2 m to the path's smooth edge — on the lawn's side only (the lawn's distance falls toward the path), never across a narrow path onto the paving beyond it. The same weight (GroundColour.path) then decides the sand, the slabs' share and the urban green's, and a path texel off the smooth edge is paved as the land beside it. The paving raster's walk kinds are read as their shares of the 4×4 texels round the fragment under a cubic B-spline, cut at one half and faded into the fragment's own texel between 0.65 and 1 m a pixel (walkShares; bilinear shares of the 1 m texels drew a slanted edge as teeth), (on the ground a walk kind paints — built-up, path, unclassified, and a path's smooth edge — as before), so a stray sealed texel among a path's sanded ones is a faint grey, not a block. Since 2026-11-03 the bake carries a tagged way's own kind on along its path, up to 4 texels past the way's buffer (surface.py along_paths; only a walkway's own buffer seeds it, never an area, a sidewalk band or a footway=sidewalk way), where the DLM's BEF says nothing, before the town's default (asphalt) fills the rest: the class raster's path is often wider than the buffer, and a sand path came out grey at its sides. Out to 1.2 m a pixel (the shares to 1 m), then the splat as before. No bake, no raster.

  • Paving materials ❝taken ≈assumed (Beläge) — OSM surface=* on the highways (plus sidewalk:*:surface bands beside the roads, footway:surface, pedestrian squares, parking lots) → a 2048² four-byte raster per tile (pipeline/bake/surface.py → surface_<tile>.png, greyscale, the bytes interleaved so the viewer's own PNG decoder reads it exactly): R packs the carriageway's and the pavement's material (asphalt, concrete, slabs, sett, unpaved, grass pavers; park · 64 + walk · 8 + road), G the way's direction (on arable land the field's long axis), B and A's low five bits the 13-bit offset along the street, A's bits 5–6 the street parking's place and bit 7 whether its bays are painted (Parking below). The shader reads road on class 7 and walk elsewhere; unknown falls back to asphalt / slabs (class 4) / sand (class 6). A path OSM gives no surface (2026-10-21, path_defaults) takes the Basis-DLM's BEF, else asphalt inside the settlement and unpaved outside it — texels of an unknown path: Dresden 12 638 → 1 748, Unna 22 779 → 562, Grimma 24 919 → 1 462, München 4 752 → 203. A street's sidewalk:*:surface band reaches from the kerb to the nearest building its normal meets (1–8 m, band_depths), else 3 m. Patterns in the street's own frame — slabs in running bond (low-contrast joints), concrete plates, gravel and asphalt mottles, grass pavers; sett as a darker, warmer tone with a fine direction-free grain (the drawn stone grid with pillow shading read as busy and seamed where two streets' frames met, and was abstracted away on review) — fade out by fwidth before they alias; the material's tint stays at any distance. Coverage (Dresden, 2026-09-19 extract): ~87 % of the highway ways carry surface, ~68 % of the DLM carriageway texels get a material. Fine terrain level only; absent raster → the class defaults. On Bodendetail. ground-detail.ts, terrain-layer.ts.

  • Parking ❝taken ≈assumed (Parkplätze) — OSM street parking (parking:{left,right,both} = lane / yes / street_side / on_kerb …, with :orientation) and car parks on the ground (amenity=parking, amenity=parking_space; the service=parking_aisle driveways cleared) → the paving raster's top two bits. On the carriageway a parking lane from the kerb distance — 2 m with bays every 5.5 m (parallel) or 5 m with bays every 2.5 m (perpendicular / diagonal) —, its lines only where OSM says the bays are marked (2026-10-21, parking:<side>:markings, the old marked / painted_area_only; A's bit 7), else a 3 % tint and no line; half_on_kerb straddles the kerb by 1 m, on_kerb lies on the pavement (A's bits 5–6, ground-detail.ts parkPlace, parkPainted); in a car park bay lines every 2.5 m across the aisle (or, without a mapped aisle, the lot's long axis), painted unless markings=no or the lot is unpaved. Pale painted lines, faded out past ~15 cm/px; a car park without surface is asphalt. Painted street bays: Dresden 310 → 7.0 km (225 825 → 10 777 m² on a tile), Leipzig 162 → 4.7 km, Hamburg 47.7 → 0.4 km, München 169 → 0 km, Meißen's 6 598 m² → 0. Dresden: ~600 surface car parks, ~530 mapped bays, ~780 aisles, street parking on ~1 000 roads. No cars (not in any dataset). On Bodendetail. ground-detail.ts.

  • Sports grounds ❝taken ≈assumed (Sportplätze) — OSM leisure=pitch / track (areas, and tracks mapped as a line), nothing indoor, covered or on a roof (playgrounds are street furniture, below) → pipeline/bake/sport.py: a table of grounds (sport_<tile>.json, one row each: centre, long axis, size, surface, line scheme, shape) and a 2048² index raster (sport_<tile>.png, four bytes per texel: the row on top, a second row reaching it, the exact outline bit). The surface is the surface tag, else the one the orthophoto shows (2026-10-20, below), else the sport's usual one (football grass, tennis clay, basketball hard court, athletics tartan, beach volleyball sand); the lines are the sport's (the first of a ; list): football, tennis, basketball, volleyball, a handball / multi- sport court, running lanes, a chess board. The shape is the minimum rotated rectangle (area ≥ 85 % of it), a capsule band for an oval track (the band's width solved from the mapped area), else the mapped outline. In the terrain fragment pass (sport-ground.ts) the four texels around a fragment name up to eight candidate rows; each row's analytic shape decides inside or out, so edges are exact rather than the raster's metre staircase — the first cut read one row per texel and sawed teeth into a track's inner edge where its capsule model strayed past the mapped outline into the pitch's texels (fixed with the second row). Pastel surfaces (lib/city/sport.ts), mown stripes on grass, fine grain, and the standard line dimensions (FIFA, ITF, FIBA, FIVB) scaled down to fit a smaller ground; lines box-filtered over the pixel footprint (a 12 cm line is a steady hairline from afar); blue tape on sand. An untagged surface is read off the orthophoto (photo_surface): the football default had drawn every untagged pitch as grass, and in Dresden one of the twelve is — four are artificial turf, six the red rubber of a small all-weather pitch, one cinder. The photo is taken to Dresden's colour balance (dop_balance.py) and read on the outline shrunk by 2 m, in its sunlit texels, when at least half of them are in sun and no more than a quarter stands 2 m over the ground (crowns, roofs): a calibrated NDVI ≥ 0.2 is grass; then in linear RGB a b/g over 1 (not red) a blue or teal coated hard court, an r/g ≤ 0.9 artificial turf (green but not growing), an r/g ≥ 1.4 tartan when its b/g is ≥ 0.8 (pink-red) and otherwise clay on a tennis court or cinder, anything else earth below a b/g of 0.88 (clay on a tennis court, else cinder) or a hard court. A running track is not read (its band is narrow, and its mapped area often holds the infield: the first pass read football grass into tracks), and a sand ground — a riding arena, beach volleyball — only turns grass (sand reads as earth or a pale hard court). A flight without infrared (Bavaria's GLI) cannot tell turf from grass, so a green ground keeps its default there (🗃️ A visible-band turf test, below). On the tagged grounds of the cities with infrared the NDVI tells grass from turf on 63 of 69 (five tagged grass show nothing growing — converted since, perhaps); a tagged surface is never changed (the photo is one day's). A pitch of no known sport, drawn only with a tagged surface until then, is drawn with the photo's too (no lines). On the fine level the dressing stands goals, basketball posts and nets (sport-fixtures.ts, sportFixtures): pale-clay bars, the nets a lavender-grey veil in the street furniture's palette and matte material, one merged mesh each per tile, in the tile that owns the ground's centre. The lines on a desktop's fine level only; a phone's pitches show their land-cover class, its goals and nets stay (🗃️ Sports-ground lines beyond a desktop's fine level); absent files → the land-cover class. Dresden (2026-09-19 extract): 431 grounds over the fifteen tiles (175 on the four first; a table row per tile a ground reaches). Textures scale with Bodendetail; colours and lines stay. A block of tennis courts (2026-10-21, W20): a tennis ground over 1000 m² is round(area / 650) courts — its rectangle cut into the grid nearest a court's 2 : 1 (one row on a tie), each cell its own row with its lines and net (Dresden 53 → 56 rows, München 6 → 8).

  • Road markings ❝taken ≈assumed (plan 026) — OSM highway=crossing nodes, highway=traffic_signals nodes with a traffic_signals:direction (or direction), and the roads' lanes, oneway, lane_markings and cycleway[:right|:left|:both]=lane → pipeline/bake/markings.py: a table of rotated rectangles (markings_<tile>.json: centre, the axis across the road, half extents, kind) and a 2048² raster (markings_<tile>.png, four bytes per texel: the row reaching it, 16-bit; lane bits; the signed distance to the carriageway's middle). A crossing is painted when OSM says it is marked (2026-10-21, B18–B26, crossing_kind): a zebra only from crossing:markings zebra/ladder, crossing_ref=zebra, crossing=zebra or the zebra's sign DE:350; a Furt (two broken lines) for crossing:markings lines/dashes/dots/yes or a signalled crossing (German signalled crossings are a Furt: 97 % of Dresden's and Leipzig's that say so); marked or uncontrolled alone paints nothing (in the orthophotos they are mostly unmarked, or cross a cycle track — only about 30 % of the zebras they drew were there), nor do unmarked and markings=no. Zebras: Dresden 172 → 35, Leipzig 88 → 34, Hamburg 149 → 81, München 100 → 33, Unna 63 → 26, Meißen 11 → 7, Grimma 3. Its axis is the normal of the nearest carriageway way at the node, its length the contiguous DLM road texels along that normal — measured also 5 and 10 m along the road, the narrowest valid run winning, because at a junction the normal through the node runs down the crossing street. A stop line lies 3 m before a directed signal, across the right half of the approach (the whole road on a oneway). Cycle lanes (2026-10-21, cycle_sides, cycle_lane_bits): a Radfahrstreifen (:lane=exclusive) a solid 25 cm line at 1.85 m from the kerb, a Schutzstreifen 12 cm dashes at 1.5 m, 2.25 m where it is tagged ≥ 2.05 m wide; moved out past a parking lane on that side; red where surface:colour=red (and the whole track of a red cycleway mapped on its own, RED); none for pictograms only — on the side the raster says (resolved in the bake per texel — the paving raster's bearing is only known modulo 180°; lib/city/markings.ts cycleLineFromKerb). Solid / dashed: Dresden 51.8 / 27.0 km, Leipzig 40.9 / 10.4, Hamburg 27.8 / 6.0 (+ 1.5 km red), München 19.7 / 5.0, Unna 0.4 / 8.8. Centre lines: 12 cm dashes (3 m in 8.25 m) where the signed distance to the middle crosses zero, on two-way primary/secondary/tertiary/trunk/unclassified roads with lanes ≥ 2 and a carriageway ≥ 5.5 m, clear of junctions by 12 m — residential streets tagged lanes=2 (≈300 ways) are left unmarked, as they mostly are in Dresden; solid where overtaking=no or the divider is a solid line, none on a physical divider or within 2 m of a tram track (the tracks run on the axis: Dresden 33.1 km, Leipzig 9.0, München 4.7 dropped); with four lanes or more a solid middle and dashed lane lines halfway to each kerb (centre_bits, lane_fields). Painted in the terrain's fragment pass (road-markings.ts), box-filtered (a zebra far off averages to a pale band), clipped to the carriageway, worn, fine level only; scales with Bodendetail. Baked from the BBBike extract of 2026-09-19: over the fifteen tiles (2026-09-26) 837 crossings (127 zebra, 710 Furt) and 583 stop lines, 169 crossings and 93 signals left unpainted; the first bake (2026-09-25, four tiles) had 364 crossings (59 zebra, 305 Furt) and 214 stop lines, 60 crossings and 32 signals unpainted (no DLM road within 3 m of the node — mostly service roads the DLM does not carry — or no carriageway narrower than 30 m nearby). Across a seam (2026-09-26): rows are measured on the class rasters of the tile and its neighbours, so a carriageway is no longer cut at the tile edge (8 rows had been, one down to a 0.12 m half-length), and a neighbour's row whose paint reaches into the tile is painted there too (13 such rows; before, 4 rows reached past an edge that no neighbour painted). Checked on the four first tiles against the crossing footway through each node: the painted axis is within 20° of it for 278 of 301 (median 2°), and 18 of 364 rectangles lie less than 70 % on the DLM carriageway — under the plan's 1-in-10 stop. ≈12 s per tile, 77–107 KB raster + 5–11 KB table. Overlapping rows (review, 2026-09-25): the raster names one row per texel, last wins, so a crossing OSM maps twice (a node per way, a zebra node beside a signal node) clipped its twin — 62 of the 578 rows lost paint, 6 of them all of it (e.g. a Furt under a zebra on 33410_5658). The bake now merges rows of one family on one axis (within 30°) whose outlines come within 1.5 m — a zebra wins over a Furt, the union across the road, along it only when parallel — and gives each texel within 0.75 m of a rectangle to the nearest one before any margin: 516 rows (52 zebra, 251 Furt, 213 stop lines), none losing paint; 9 lose a sliver only where another row paints over it (paint_lost, logged and tested). Phones read a 1024² twin (markings_low_<tile>.png, 25–36 KB; a 1.45 m core so the rows still resolve — none loses paint): 4 MiB of GPU memory per fine tile instead of 16; its lane bits are coarser (≈90 % of the cycle-lane and centre-line texels agree with the 2048² raster). Two carriageways in line (the other sites, 2026-09-26): on Leipzig's ring and Hamburg's wide streets the crossings of both carriageways touch across the median and merged into one row painted over it (44 m across on Leipzig's ring); a merge that would reach further across than one carriageway plus a metre (MAX_MERGED_HALF_M, 16 m half-length) now leaves them two rows. Dresden's largest merged row is 15.18 m, so its files are unchanged. Not yet judged on a GPU.

  • Urban green ◎detected (Stadtgrün) — the DLM's built-up class (4) covers courtyards, front gardens and parks inside the settlement alike. Where the DOP NDVI (upsampled, blurred) passes 0.3 on classes 0 and 4 and OSM does not call the ground paved, edges.py counts it as meadow; the shader paints it exactly as meadow — its colour, mottle, NDVI tint and lawn edge. The railway class (5) too, off every track bed (ways.track_beds + 0.5 m; 2026-10-20): the DLM's parcel holds the embankments and verges, which the ballast grey painted grey — off the traffic, the calibrated index passes 0.3 on 5 % of it in Dresden, 8 % in Leipzig, 23 % in Hamburg and 29 % in Unna. The road class (7) too, beyond everything paved (2026-10-21, PR #173 B12, edges.py verge_green): the DLM's parcel holds the embankments, verges, junction islands and the woods between a dual carriageway's lanes, which the class raster painted as asphalt. Clear of the carriageways and track beds (ways.traffic), OSM's highways at their width (+ 3 m with a sidewalk tag, each + 0.5 m) and its paved areas (car parks, squares, highway areas but the traffic islands, buildings), what is green low ground (NDVI > 0.3, DOM1 under 1 m over the DGM) or a wood (a crowned patch at least 8 m across, ≥ 20 % of it green: under a crown the NDVI may be the crown's own green over a pavement, and a leaf-off wood reads green in patches only) counts; closed over 2 m, holes under a crown filled (the tree on an island), patches ≥ 25 m². It repaints 0.43–1.57 % of a site's road land (Dresden 2.8 ha, Leipzig 1.7, Unna 0.9, München 0.7, Meißen 0.3, Grimma 0.15; Hamburg none: its road class is OSM's buffers, which the paved ways cover). Without DOM1 none (a street tree's crown reads green over a pavement), nor without the DLM's road axes on a DLM site. The first cut blended toward the meadow colour at 0.85 × slider and read as barely there. Stadtgrün at full strength (URBAN_GREEN, fixed). ground-detail.ts urbanGreen.

  • Water ❝taken =computed ≈assumed — the painted splat's alpha (water coverage, smoothsteped shoreline) + the terrain geometry + the wind's waves on a level normal (+ the drifting mist); the water and mist sheets hang next to their terrain mesh and leave with the tile, each over only the triangles within its reach of the class raster's water (100 m, the mist 600 m; since 2026-10-24, lib/city/water-mask.ts — they draped the whole tile before). water-layer.ts. Missing class raster → no splat, no water: the tile's ground falls back to the flat sage. Waves ≈assumed (since 2026-10-23, water-waves.ts; drawn, not simulated, since 2026-10-25): one wave field for the river sheet and the fountains' basins, a function of the place and the clock alone, so two tiles meet without an edge — two layers of gradient noise stretched 2.6× along the crest (12 m and 4.4 m across on the river, 1.2 m in a basin), each drifting downwind, the finer bent by the coarser, under gusts 80 m across (5 m in a basin) that leave calm patches beside rough ones. Each layer fades where its wavelength spans fewer than a few pixels. The river is matte (the maintainer's choice, 2026-10-25): its body's tint and depth tone, lit by the scene, the waves' normal tilting that light a little so the water moves — no sky mirrored in it. The sun's reflection: the material's own specular (roughness 0.6) a soft path toward the sun, shadowed by houses and decks, and emissive on it tight glints and, on the crests' lee side (where the field crosses one level, stroke), dashes of light. Three noise reads a pixel; the five-octave field before it took eight. It replaced crossing sine waves (⛔ below). Since 2026-10-21 the land-cover bake also splits the water class into bodies (pipeline/bake/water.py, the legend's water): 4-connected 1 m texels, cut only where OSM maps waterway=weir|lock_gate|dam across them. One surface per body: a DGM fill of at least 50 m² whose edge lies mostly more than 0.5 m below a larger fill of the same body is lifted to it, smallest first; nothing is lowered. A measured sloping surface keeps its slope (so a seam never steps), and so does a step at a weir. The terrain bake raises lifted cells on both levels (raiseWater). Hamburg 32564_5932: 16 629 → 1 701 step cells, 42 027 m² raised; its four tiles 35 087 → 5 036. The tide is not modelled: the Elbe stays at its DGM fill, 2.00 m, about mean high water. Shore ◎detected: water texels more than 0.3 m above the body's local level (20th percentile per 5 m block, median over 45 m) become meadow where the DOP is green (NDVI ≥ 0.15, GLI ≥ 0.05), the neighbouring built, rail or road class within 3 m, else sand (class 6). Streams under 8 m wide are left alone. On a quay the dry band is the DGM's ramp up the wall: the measured quay stands over it (Walls below). Dry area before → after: Dresden 33412_5656 12 140 → 468 m²; München 18 824 → 2 423 m²; Meißen 7 415 → 1 009 m². Tint =computed: each body's open water (≥ 5 m from its bank) is read in the cast-corrected DOP as a Lab median and pooled across the tiles it spans; drawn as its hue at the sheet's lightness, back to the palette where the read is grey, dark or close to it. Hamburg's Elbe is ochre and the Isar green; the Alster, the Fleete and Saxony's rivers keep the palette colour. Foam ◎detected ≈assumed: a white band along OSM weirs (the DGM's drop there, else 0.8 m) and along drops of more than 0.5 m within 20 m of open water with water on both sides; at most 16 per tile, fine terrain level only. München 784 m (up to 4.6 m), Dresden's Weißeritz 0.9 and 1.9 m. Not yet judged on a real GPU (the foam and the tint).

  • Landing stages, groynes, ferries ❝taken =computed ≈assumed (plan 031) — OSM man_made=pier (51 ways, buffered by width, else 3 m, flat ends; 5 areas), man_made=groyne (1) and route=ferry ways (3: the Johannstadt ferry, two paddle-steamer routes) (ODbL) → pipeline/bake/riverside.py → riverside_<tile>.geojson (44 piers, 43 pontoons, 1 groyne, 9 ferry stretches over the fifteen tiles, 2026-09-26; the OSM counts above and 25 / 31 / 1 / 4 are the four first tiles'). A pier's deck from DOM1 (2026-10-21, V27/V28): DOM1's median over its part on the water or the low ground beside it (≤ 1 m over the water), read off the cells 0.3–2.5 m over the water or the ground; the bank's DGM at its landward end (its highest dry sample)

    • 0.4 m only without DOM1 (that rule had a Dresden pier 5.7 m over the river). A pier DOM1 has under the DGM's river (Hamburg: the DGM's river at 2.0 m, the scan at −0.6 m with the tide) floats as a pontoon. A pontoon's freeboard fb = DOM1 over its hull − DOM1's open water 3–12 m out, clamped 0.4–1.5 m (0.9 m without DOM1); a fixed pier on the water touching a pontoon is its finger (finger: 1: its fb, no hut, no gangway). A stage is drawn only where a quarter of its cells stand ≥ 0.3 m over the water — Dresden's 145 m pontoon at 417284,5654960 lies under the river. A groyne is cut to where the DGM along it stands ≥ 0.3 m over the water beside it (w, the water's level 3–15 m out), its crest on the DGM — Dresden's one groyne lies under the water and is not drawn. Piers / pontoons: Dresden 44 / 43 → 16 / 48 (15 fingers, 23 not drawn; fb median 0.67 m, 18 of 48 at the 1.5 m cap), Hamburg 38 / 106 → 23 / 109 (4 floated by their deck; fb median 1.5 m, 61 of 109 at the cap — the cap may be too low for Hamburg), Leipzig's piers 7 → 5, Meißen's 2 → 1. A pontoon (floating=yes, or a pier reaching > 25 m out on the water class) is cut to its part on the water class — OSM draws many up the bank, and the DGM's river surface is flat (103.75 m, 104.3 m, 105.05 m by stretch) while the bank rises 2–5 m, so the uncut outlines spread 2.7–5.2 m of DGM under a hull; cut, the median spread is 0.22 m (max 2.27 m, a shoreline cell) — the plan's STOP asked for the water surface the way water-layer.ts draws it: the water sheet is the terrain there, so the runtime floats the hull on the lowest ground under it (waterLevel), never on a baked DGM value. bank is the nearest dry point (≤ 30 m) for its gangway, unless a fixed pier reaches it (28 of 31 have one). Ferry routes are cut to their stretches over the water class. Runtime riverside-layer.ts: a pier is a timber deck 0.3 m thick on instanced piles every 4 m round its edge where it stands over the water, with a railing (top rail + posts every 2 m) along the edges over the water (plan 029's fence panel is not merged here); a pontoon a soft slate hull from 0.3 m under the water to its deck's underside, under a pale deck 0.15 m thick at the water + fb (until 2026-10-21 −0.3 → +0.35 m under a deck at +0.5 m), a clay ticket hut with a slate roof on one longer than 15 m, a 1.4 m gangway with hand rails to its bank point (a finger pier the same without hut or gangway); a groyne a low stone ridge along the stretch the DGM shows, its crest 0.12 m over the DGM (which holds its stones), flanks 2.5 m out down to 0.3 m under the water beside it (w), else 1 m down; a ferry a faint dashed wake (1.2 m, 14 m dashes, alpha 0.55) on the water — drawn as its own ribbon over the water sheet rather than inside the water shader (no line table; same look, one draw), and, the STOP's conservative fallback taken without a GPU plate, only from the air: it fades in with the camera's height over the ground from 25 to 60 m (map-overlay.ts). The paddle steamers themselves are not drawn: no dataset has them. Look unverified on a real GPU.
  • Streaming site (3D Tiles) =computed — scripts/prepare-data.ts bakes the site into an OGC 3D Tiles 1.1 tileset (lib/city/tileset.ts) under public/data/<site>/, one per built site, each streamed at its own route (/dresden; scripts/prepare-sites.ts, ADR 0037): per site tile the buildings (loaded whenever the tile is in view, far off simplified, near as baked — REPLACE, plan 076) beside the two terrain levels (REPLACE). 3DTilesRendererJS loads and unloads by screen-space error (16 px) from the view camera and the sun's shadow camera, so a tile casting into the view stays loaded. Only the fine level is dressed (vegetation, lamps, monuments, rails; its stairs and walls are baked into it); distance, not a "primary" role, decides which tile is detailed, and collision, demolish, focus and double-tap work on every visible tile. Everything a tile adds leaves with it (tile-stream.ts, processTileModel / disposeTile) (ADR 0024). The lite e2e profile streams tileset-spawn.json, the spawn tile alone.

  • Progressive first frame =computed — the spawn tile's buildings + any of its terrain levels render first; the heavy dressing waits behind a gate the HUD opens after the handover (startStreaming in create-app.ts) and is built one tile at a time, each landing re-rendering the shadow map. Until the site has first loaded the fog far plane is clamped to ~1.1 km so tiles still in flight read as haze (ADR 0024).

  • Rasters at 2048² =computed — scripts/prepare/side-bakes.ts downsamples the 4096² class raster NEAREST, one band (scripts/downsample-raster.ts downsampleClassRaster), to a quarter of the texture memory: for the fine level on phones (MOBILE_RASTER_PX, chosen by the client per device tier); a 1024² one for the coarse terrain level on every device (COARSE_RASTER_PX, since 2026-10-10 — 2048² before) and a 512² one for the minimap. No raster whose alpha is data goes through an image resize any more (the painted splat has the only such alpha, and it is made on the GPU). Cost: ~1 m instead of ~0.5 m class boundaries on a phone's fine level, ~2 m on the far terrain on every device (a pixel or less out there).

  • Buildings ❝taken — CityJSON LoD2 → build-time glTF per tile: scripts/bake-city-mesh.ts runs cityjson-threejs-loader, and its cityMesh welds it into one mesh with an EXT_mesh_features feature id per vertex (_FEATURE_ID_0) and a roof flag. The per-object style (tint, roof colour, base, eave and storey heights, glow, roughness) and the demolish tree ride as an EXT_structural_metadata property table (scripts/tile-glb.ts); the client packs it into an RGBA32F texture the clay node reads with textureLoad (lib/city/city-mesh.ts packObjectTexels), so the style lives once per building, not once per vertex. Footprints go, packed, to footprints_<tile>.fpr.gz (lib/city/footprint-pack.ts), and as ink into the site's one minimap picture (minimap.webp, scripts/bake-minimap.ts, cover per pixel from lib/city/footprint-figure.ts), which also holds the bridge decks from each tile's bridge file, in the colour of the class they carry — the class raster shows the river under them. Demolish = filter the building tree out of the index buffer + rebuild the BVH; BVH picking/collision, each tile's BVH built when a ray first reaches it (ensureCityBvh, plan 069, ADR 0024), not as it lands. city-layer.ts. The DOP roof LUT is folded in at bake time.

  • Small structures from the laser scan ◎detected (plan 034) — what the laser saw 2–6.5 m above ground and LoD2 does not carry: garden and allotment houses, sheds, carports, container buildings, pavilions, and parts of larger buildings the LoD2 lacks. pipeline/bake/small_buildings.py on the 0.5 m scan rasters (lsc.py): candidates outside the LoD2 (+1 m), mapped walls and bridge decks, off rail/road/water (classes 5/7/8); the core = cells with no multi-echo return in their 3 × 3 window (in the leaf-off flight vegetation splits the pulse, a roof does not — but a roof's edge does, so the plan's blob-wide "multi-echo ≈ 0" rule found 4 blobs on the spawn tile, the per-cell rule 130, none of them touching a crown); cores ≥ 4 m² grown back one cell, kept at 6–150 m² with a planar top (residual < 0.35 m, tilt ≤ 35°), filling > 60 % of their minimum rotated rectangle, ≥ 2 m wide, NDVI ≤ 0.25 (a clipped evergreen block is flat and single-echo too), not a roof-edge sliver (aspect ≥ 3 with ≥ 20 % of its rim on the LoD2), not vehicle-sized (2–2.8 × 4.5–18 m, < 4.2 m) unless an OSM building outline confirms it. What the flight saw that day only is left out through OSM: pedestrian areas and squares (the Christmas markets were up on 27–30 Nov 2024 — the Altmarkt alone held 101 blobs, the Neumarkt 39, the Stallhof 11, Prager Straße 28), within 6 m of a pedestrian street's line, marketplaces, building sites and surface car parks; monuments and stop shelters (other layers draw them) within 3 m. Each is its rectangle with z (lowest ground), h (the fitted top's median above it) and, tilted > 8°, hc (a pent roof's corner heights) → data/<site>/dlm/smallbuild_<tile>.geojson (the provider's credit). 6 625 on the fifteen tiles (703 pent roofs; the first bake shipped 6 783 of 8 168 before the OSM context), the allotment colonies most of them. bakeCityMesh appends each as a closed box (lib/city/small-buildings.ts, sunk 0.2 m) — its own object and root (demolish, picking, collision and the minimap work unchanged), building true, the wall tint hashed from its first corner, its DOP roof colour where in sun, else the flat-roof slate palette, no glow, no storey band, source = 1 in the property table (0 = LoD2). +3.3 % city glTF. Review fixes (2026-09-26): the bake reads 40 m of the neighbours' scan and layers across each seam and keeps a structure on the tile holding its centroid (27 now straddle a seam; before, a blob touching the edge was lost on both sides); z is the lowest ground under the whole rectangle and the DGM1 at its corners, and a box over ground falling > 1.5 m is dropped (21 boxes had floated

    0.5 m at a corner, the worst 1.74 m; now 1, 0.84 m); of two rectangles overlapping > 0.5 m² the larger stays (43 z-fighting pairs → 0); the canopy and scan points in or within 0.5 m of one are dropped at build time (1 232; see Canopy fill). Gate (phase 0): 16 of 20 sampled on the spawn tile are structures against the DOP (March 2024; the misses two roof-edge slivers, now ruled out, and a shadow). Unverified on a real GPU. Remaining temporaries: 3 probable stalls on the Schloßstraße pavement (a living street), container stacks in the Alberthafen. More sites (2026-09-30): Saxony's scan is now fetched and baked (bun run fetch <site> --lsc) for Meißen, Grimma and Leipzig too (Meißen: 475 small structures; Leipzig: 1 303), and NRW's "3D-Messdaten" for Unna (1 596 small structures, 2 385 extra crowns) (the same AdV classes, 2 ground / 20 non-ground; providers/nw.py merges the four 1 km LAZ of a tile, see Hedges below), and (2026-10-01) the LDBV's laser points for München (526 small structures: 147 / 88 / 168 / 123 in tile order 32690_5334, 32692_5334, 32690_5332, 32692_5332; 32 552 extra crowns), its classes mapped to AdV with buildings (6) kept in the non-ground class, so the sheds' roofs are in the surface (providers/by.py, see Hedges). The rules were tuned on Dresden's flight alone. In the photo's shadow (2026-10-21, G28): a candidate whose DOP luminance median is under 45 must fit its plane to < 0.2 m and fill 75 % of its rectangle — a crown's or a house's shadow on a courtyard read as a shed. Unna 877 → 872, München 599 → 542 (about half of the München drops look like shade in the orthophoto, some like courtyard roofs).

  • Structures beyond LoD2: columns ◎detected ❝taken (plan 050, ADR 0036) — the chimneys, towers, masts, water towers, communications towers and lighthouses the LoD2 leaves out. pipeline/bake/structures.py reads DOM1, DGM1 and the burned LoD2 roofs over the tile and a 60 m margin at 1 m and takes gap = DOM − max(DGM, LoD2 roof); an OSM man_made=* of those kinds (point or outline) is kept where the gap within 6 m (2 m of an outline) reaches 5 m and the structure 8 m. Axis at the gap's peak, foot at the lowest ground there, height the DOM's — OSM's height where the DOM undershoots it by more than a quarter (a lattice mast); radius from the outline, OSM diameter or the gap blob low down, a chimney at least h/24 (a 1 m surface model sees a slim shaft too thin: Unna's 48 m Lindenbrauerei chimney, the "Fibonacci-Reihe", measured 1.7 m), clamped per kind, tapered per kind. Nothing is added from the surface model alone (🗃️ below). Nor twice: a tower, lighthouse or water tower (BUILT_TOWERS) whose mapped foot lies under a LoD2 roof is LoD2's own — a church's or a castle's tower is in the building mesh already, and a lathed column inside it doubled it (Meißen's cathedral towers, 81 and 83 m, Grimma's churches, 13 of Munich's 14 towers); masts and chimneys on a roof are not in LoD2 and stay (🗃️ below). lib/city/structures.ts lathes each from its kind's ring profile (16 sides: a chimney's head band, a TV tower's cabin and antenna, a water tower's tank on a slim shaft, a lighthouse's lantern); bakeCityMesh appends it (appendGapStructures) as an object of its own, source = 2, not a Building (the HUD's count stays houses), chimneys, towers and lighthouses in the brick palette, the head in the wall's tint (no terracotta cap); clay, shadows, picking, collision, demolish and the minimap as for any building. 386 on the seven committed sites (2026-09-28; Dresden 180, mostly masts and chimneys). → data/<site>/dlm/structures_<tile>.geojson (ODbL); a site without a surface model writes an empty file. Look unverified on a real GPU.

  • Structures beyond LoD2: missing buildings ◎detected ❝taken =computed (plan 050) — the buildings OSM maps and LoD2 does not carry yet (LoD2 is a year or more behind; the newest buildings of a city are the usual case): an OSM building outline or building:part ≥ 40 m² (not roof, carport, construction, ruins, underground or on a negative layer) that LoD2 covers < 20 % and the surface model fills (≥ 2.5 m above ground on ≥ 60 % of it), extruded from its lowest ground to the 60th percentile of the measured height (the mass, not its antenna) — or to OSM's height (height, else building:levels × 3.2 m) where the model was flown before the building stood: it measured under 60 % of that and at least 1.5 storeys short, or did not fill the outline at all (Hamburg's DOM1, 2022-11-21, saw the Überseequartier as shells: The Lyte 21.6 → 51.2 m). Not where OSM dates the building before 2010, where the model does not fill it and OSM maps it under two storeys, or where DOM1 covers < 95 % of the outline (the site's edge, where nothing says the flight predates it); a building OSM maps no height of needs the model. Such a building carries src: "osm", and its card says the height is OpenStreetMap's. A part OSM lifts off the ground (min_height, else building:min_level storeys) starts at that floor (base): a bridge between two buildings, an overhang. Outlines overlapping one already taken by more than half are left out (a block mapped over its own parts). The canopy's trees inside a missing building are dropped at build time (lib/city/roof-trees.ts gapRoofs): DOM1 − DGM1 there was the building, or the shell, never a crown. Flat roof (or the measured roof's shape, below), its DOP roof colour where in sun, else the slate palette, storey bands from that height, source = 2, building true. Same bake and build as the columns. 866 on the seven committed sites (2026-10-21; Dresden 369, Leipzig 235, of them 50 at OSM's height, Hamburg 81, of them 21: The Lyte 51.2 m, Luv 44.8 m, LEE and The Yard 16.0 m). Windowless; a tower inside its podium may leave coincident walls. Roofs from DOM1 (2026-10-21, G27): a missing building the model measured (not at OSM's height, not lifted off the ground) takes its roof's shape from it the way roofs.py rebuilds a LoD2 roof (shaped_roof): flat levels and measured faces as parts ([{coordinates, z, surface?}]), written where the model holds two levels or more or a face and misses the scan on at most 30 % of its cells (roofs.KEEP_SHARE); the build draws them as a measured roof (shapedGap, measuredRoofMesh), else the flat box stands. Shaped: Leipzig 17, Dresden 13, Meißen 9, Unna 4, Hamburg 3, Grimma 3, München 2. A lattice is no roof (2026-10-21, structures.lattice): a measured level with no face of its own that stands more than 6 m over the building's height, where fewer than 40 % of the surface model's cells lie within 1.5 m of its top, is a mast, a crane or a floodlight, and stands on the highest level below its top that 40 % of the cells lie within 1.5 m of (solid_level: a roof), or at the building's height where none does. (The Rickmer Rickmers' masts stood as 25 m towers on its hull, Dock I's cranes 11 m over the dock, Leipzig's Paralympische Sporthalle a floodlight 50 m up. The median of the cells below the top, tried the same day, stood the ship's masts as 24 m blocks again: its rigging reads at every height from the deck up.) Nine levels on the seven sites.

  • Roofs on posts ◎detected ❝taken =computed (2026-10-21) — a platform roof, a fuel station's canopy, a carport, a bus shelter: a roof with nothing under it but its posts. structures.py takes an OSM building=roof or carport, or an amenity=shelter that is not a walled hut (shelter_type=basic_hut), ≥ 12 m², that LoD2 lacks (< 20 %) and DOM1 shows as a level surface: ≥ 60 % of its cells ≥ 2 m up, their middle half within 1 m (a crown over a carport is a dome), at most 4.5 m (carport), 6 m (shelter) or 16 m (roof) over the ground (higher is a tree over it). Its top is the 60th percentile of the surface there (kind: "roof"). lib/city/open-roofs.ts draws it as a slab under that top, 0.04 × √area deep within 0.2–0.7 m (roofDepth), on 0.3 m posts (roofPosts): a roof under 5 m wide on both long edges, a post every 4 m at most; a wider one on rows under its middle, one row per 10 m of width, a post every 8 m along — a platform roof's centre posts, a fuel station's two rows; only posts 0.4 m inside the outline stand. building false, the flat-roof palette, no windows or doors. 289 on the seven committed sites (Dresden 187, 26 892 m², the Hauptbahnhof's forecourt among them; Leipzig 36, Unna 18, of them seven platform roofs, 1 627 m², 5.8–6.0 m up — LoD2 carries none of the railway's; Hamburg 26, München 15, Meißen 4, Grimma 3). Flat; the canopy's trees under one are dropped at build time like a missing building's.

  • LoD2 Überdachung opened =computed (2026-10-21) — LoD2 draws the ALKIS roofs on posts (51009_1610 Überdachung, 51009_1611 Carport) as solid blocks down to the ground: a platform under one vanished, a fuel station's pumps stood inside it. scripts/city-roofs.ts (withOpenRoofs) cuts their walls to a fascia as deep as the slab above, drops the floor and adds an underside and the same posts, where at least 1.8 m stays clear (a bin shelter or a cellar entrance under that reads as well as a block); they take no windows (WINDOWLESS). Opened on 2026-10-21: München 913 of 989, Hamburg 356 of 368, Unna 524 of 536; Saxony's LoD2 carries no roofs at all.

  • Landmark roof relief ◎detected =computed (plan 050) — a landmark's roof form that LoD2 flattens: on the LoD2 objects of a Wikidata landmark (below) only, where DOM1 stands ≥ 3 m above the object's highest LoD2 roof on ≥ 60 m² (and ≥ 2 % of its footprint), the excess as a height field: per connected patch (≥ 6 m²) its outline and the 1 m grid of the surface model's heights above that highest roof (grid: the north-west corner x, y, res, cols, rows, z row by row, −1 outside the patch), lightly smoothed — half the measured cell, half a Gaussian (σ = 1 cell) normalised over the patch alone, so a tip keeps its height and the 1 m cells stop reading as steps. lib/city/structures.ts reliefMesh builds a surface from it, each corner the mean of the patch cells around it, with walls along the patch's edge down to the roof under each cell and no bottom face (it rests on the roof it was measured above) — the measured roof, not an invented shape. Floor (2026-10-01): the grid carries floor per cell (m relative to z, ≤ 0, at least 0.5 m below the cell's top): the lowest LoD2 roof in the 3 × 3 cells around it — roof cells only, so at the footprint's edge the ground does not pull a wall down along the LoD2 facade — sunk FLOOR_SINK_M = 0.3 m so the roof's slope inside a cell leaves no slit. Voids: a surface-model cell more than VOID_M = 1 m below the LoD2 roof under it is no surface (the laser saw through a sound opening, glass or dark slate) and is filled from the nearest measured cell (fill_voids) before the excess is taken — each connected void on its own, when it is a hole in the roof (roof_holes, since 2026-11-10): at most 25 m² (VOID_HOLE_M2: a sound opening, a skylight), or shut in by the LoD2 roof and spanning at most half of it (VOID_SHARE: a glass roof over a courtyard, its scan within 3 m of the ground: it saw the floor, not a lower roof). A void reaching the roof's edge is a part the LoD2 drew too high, one shut in but standing well over the ground an inner hall, a larger one an open bowl: the fill lifted the Leipzig opera's hall onto its low wing and the trees around the Freilichtbühne in the Großer Garten over its seats. The case: Unna's Evangelische Stadtkirche — LoD2 draws the church as one gabled block (ridge ≈ 137 m, eaves ≈ 122 m), the west tower (walls ≈ 147 m, spire ≈ 187 m) is only in DOM1; the relief used to hover at the ridge height with its walls ending in the air beside the slope, and the voids tore "legs" into it. A file without floor builds every wall down to z, as before. The Elbphilharmonie is LoD2's flat 96.3 m block; the DOM reaches ≈ 109 m over its crests, which now roll; a spire LoD2 cuts short rises to a point (Unna's Stadtkirche: LoD2 stops at 137 m, the surface model at ≈ 178 m). (Until 2026-09-30 the excess was stacked slabs, one per ≥ 2 m band — 🗃️ below.) A relief (kind: relief, of: the host object) copies its host's row — tint, roof, flags, root — so it wears the building's look and demolish takes it along; no footprint of its own. Not on ordinary roofs: there the excess is antennas, dormers, lift housings and trees. Thresholds tuned on synthetic tests and a handful of real cases (plan 050's findings); a site whose landmarks step has not run has no relief (run landmarks before structures), and a relief without a grid (a file baked before the height field) builds nothing.

  • Roofs rebuilt from DOM1 ◎detected (ADR 0036) — a LoD2 roof that misses the surface model is drawn in the form the scan shows, inside the object's own footprint: flat levels where the scan is flat, its measured surface where it slopes or curves (since 2026-10-03; first as stepped flat blocks only — 🗃️ below). The fault: the 2025 edition's tdcFreeFormRoof (10 305 objects) is, on a complex building, often a handful of non-planar facets over the whole footprint — the Westin Bellevue (DESNATPU1000GHDs) was three facets rising from 7 m at the eaves to a 25 m peak, a tent over a slab, wings and two courtyards; and buildings finished after the LoD2's roofs were measured (2016 here) stand as 3 m placeholders (a block on the Ferdinandplatz: 3 m, 19 m in the scan). pipeline/bake/roofs.py: each building's roof burned at 1 m with its vertices' own heights (as the loader draws it; cells under another object's higher roof are not its own); a candidate when ≥ 150 m² and > 40 % of its cells (one inside the edge, NDVI ≤ 0.3) stand > 2 m off DOM1 — unless its form holds (form_holds, 2026-10-03): it rises ≥ 3 m, the scan follows it (Pearson r ≥ 0.85), it misses on ≤ 20 % once shifted by its median offset, and that offset is under the form's own height; the Frauenkirche's modelled dome (r 0.90, 5.6 m under the scan) is kept so. Missing on > 20 %, up to 40 % (since 2026-10-21, level_share, pitch_holds): a candidate only where the scan rises < 0.15 m per metre under ≥ 60 % of the missed cells (a step's edge aside) and the LoD2's pitch is not confirmed — the LoD2 rising > 0.15 m/m on ≥ half the cells and the scan on ≥ half of those keeps it. The 40 % had left the low hipped roofs the 2025 edition models over Leipzig's department stores (a quarter of the roof off, along the ridges, over a level roof: 5 667 m² on DESNATPU1000Bk2M) and many flat roofs a storey off on part of them (Hamburg's); the pitch test keeps a row house's gable or a barrel vault whose flanks are too narrow for faces, which a rebuild would have drawn flat for an annex the LoD2 left out (checked against the DOP on Hamburg's and Dresden's largest and a random draw). The decision is made on the flat model: DOM1 3 × 3-median filtered over the footprint, cut into 1 m bands, pieces < 20 m² into the neighbour closest in height, neighbours < 1.5 m apart merged, each region at its median; kept only when it misses on ≤ half the LoD2's share and ≤ 30 % (the hotel: 78 % → 3 %), not where the scan sees open ground on > 15 % of the footprint (a block torn down since: the footprint is out of date too). What is drawn: the same scan split into its faces first — cells rising > 0.15 m per metre (8.5°) over a patch ≥ 10 m across (a disk-shaped opening: narrower, the scan's metre of blur at each edge rounds a slope into a blanket — the Westin Bellevue's slab did at 5 m — where its few levels read better), or

    0.5 m per metre (27°) over a patch ≥ 5 m across (since 2026-10-20: a mansard's or a hipped roof's flank, which as levels read as a flight of 1 m terraces — the Schauspielhaus; a slab edge's blur is a step, never this steep over 5 m), the 1–2 cells of a ridge or valley closed in, never a cell beside a step of 2 m between neighbours (inclined_cells, step_cells) — and the rest banded and merged as above; faces never merge by height, and a part that holds no disk 7 m across (a step's flank, a parapet run, a plant room) joins the neighbour closest in height when that lies within 3 m (since 2026-10-19; a cell or two wide before — the strips stood as boxes), unless that makes the drawn roof miss on more than half the LoD2's share or 30 % (since 2026-10-21; 30 % alone before): a hall's monitor roofs, 5 m strips 3 m over the rest, had merged into blocks that missed as much as the LoD2 (5 of the 109 rebuilt objects on 33408_5654, one of them worse than its LoD2) — they stay strips; where the faces then still miss more, the flat model the decision was taken on is drawn (13 objects over five cities, one in München whose LoD2 missed on 88 %). A face's plausibility (since 2026-11-06, pipeline/bake/roof_planes.py): the faces also take every plane of the roof rising > 0.15 m/m that holds a disk 4 m across (planar_faces: a gable ring's 8 m flanks at 18°, too narrow and too gentle for the rules above, had been a flight of 1 m levels with a wavy step along the ring — An der Frauenkirche 5). Each face is then drawn in a form roofs are built in, or not at all: planes grown from the smoothest cells (every cell within 0.3 m of its facet's plane, refitted as it grows; facets ≥ 15 m² holding a 4 m disk; the border between two moved to where their planes cross, a ridge or a valley), taken when they explain 75 % of the face and 70 % of their inclined area faces within 15° of one axis or its right angle (orderly: a blanket cut into planes turns every way); else one profile extruded along an axis (a barrel, a shed's curve: the best of 90 axes, the scan within 0.35 m of its 1 m-binned median profile on 75 % of the face, ≥ 200 m², standing ≥ 1 m off its plane); else, on a landmark (landmarks_<tile>.json's objects — the step runs before roofs now), the scan's own surface, smoothed twice by an edge-preserving Gaussian (σ 1 cell, 1 m height σ; the Hauptbahnhof's ribbed vaults); else the face's cells go back to the levels and the model is cut again (at most three rounds). Until 2026-11-06 every face was the smoothed scan (🗃️ below). Flanks first (since 2026-11-10, gable_fit, tried before the planes): the scan smoothed over the roof's calm cells (σ 1.5 cells: a dormer's 2–3 m melts into its flank) falls into patches facing one of the roof's directions (the peaks of its slope-weighted aspect histogram, within 35°, a 3 × 3 majority twice). A patch of ≥ 20 m² that holds a 3 m square and rises ≥ 1 m, and steadily — a fifth to four fifths of its rise spread over ≥ 1.5 m, where the edge of a block raised over a flat roof rises within a metre or two — is a flank, and its cells face cells too (flank_cells, with the crowns between flanks: a ridge the scan rounds). A flank one plane explains on less than 85 % of its cells (on it within 0.5 m, or under a dormer: a blob of ≤ 30 m² up to 3 m over it) is cut along its length where the scan's height off its plane splits into two even runs (a party wall where the ridges of a row of houses step, the middle of a bending wing), and its parts again, as long as the sections (at most 16) fit their own planes on 5 % more of its cells. Then neighbouring flanks facing the same way within 15° are joined (_join_flanks) unless they differ as two houses or two parts of a roof do: a step of 0.75 m where they meet (the 90th percentile along their border), a change of pitch of 0.15 m/m (a mansard's break). A wing its dormers broke into four or five patches, each on a plane a little off the next — kinks and steps along An der Frauenkirche 5's wings — is one flank again. Each flank's plane rises along its direction alone, so eaves and ridge are level: fitted through the median of each metre across it (the first and last left out), then on its cells within 0.5 m (dormers and chimneys trimmed away); a flank turned within 15° of square to one of the building's walls rises square to it (since 2026-11-13, _along_walls: the eave runs along its wall, where the scan alone turns a flank by a few degrees with its eave's blur and what stands on it — the Schauspielhaus's flanks fell to 45°, 63°, 68° and 77° from walls at 60.5°) — only where the walls leave one such way (a footprint of many directions does not say which wall is the flank's) and its plane then lies on at most 3 % less of its cells, and the flanks so turned are tried first, the scan's own ways where they are no roof's; two neighbouring flanks facing within 15° of opposite share one axis (a level ridge). A level patch between them (a crown, a mansard's top) is a level plane; a cell near a flank (4 cells) whose plane holds it — the scan on it or up to 3 m over it, a dormer — is the flank's; a block of more than 30 m² its plane does not explain (off it, or over it wider than a dormer) is no part of the flank but a roof of its own on it or cut into it (since 2026-11-13, _own_blocks: the Schauspielhaus's plant deck cut into the auditorium's roof, the roof beyond it, its corner pavilion); what is left takes planes of its own (plane_facets) — a level where that holds as much of the patch as its plane, within 3 % (_level: what stands on a deck tilts a plane fitted to it; the Schauspielhaus's decks were slabs 9–14° askew; where the face so drawn is no roof's, the patches keep their planes) —, every other cell the nearby plane closest to the scan. Two neighbouring planes that each rise toward the other (a ridge, a hip; a level crown counts as either side) meet on the lower of the two, two that each fall toward the other (a valley) on the higher (creases): the crease is the line where the planes cross, straight and without a step, wherever the scan's rounding had left the border (a 3 m step along the ridge of An der Frauenkirche 5's east wing) — within 6 cells of it, so where the planes cross further off (a flank joined over a bend) no island of the other plane stands far from the border; other borders, a row of houses stepping, stay where the scan put them. The flanks are the face when, judged two cells in from its edge (the eave's blur: a 1 m scan rounds a steep flank's eave into the wall over two cells; one where two would leave less than half the face), 85 % of its cells — a landmark's too: its blurred scan is the alternative only where flanks fail (since 2026-11-11; at a 95 % bar the Schauspielhaus's 23° roof over the auditorium, 84 % on its flanks, was drawn as its scan) — lie within 0.5 m of their plane or under a blob of ≤ 30 m² up to 3 m over it, 60 % on it; no flank sags (a parabola through its metre strips' lower quartiles — a row of dormers lifts a strip's median, not its lower cells, and a street's dormers had made An der Frauenkirche's and Schloßstraße's flanks read as vaults — rising > 0.5 m over its chord across the middle 70 % of its fall, over ≥ 6 strips, and four times the scatter it leaves: a vault's half); the flanks hold ≥ 40 % of the face and turn orderly; and flanks holding ≥ 70 % of their cells rise from an eave — their lowest 30 % of cells lie along the face's edge and face into it within 25° (_eave; a blanket's patches rise from wherever the scan dips: none of 288 random blankets passes). Straight eaves (_straight_eaves): a face drawn as flanks ends where each flank reaches its eave — the lowest of its plane where the scan lies on it (the 1st percentile), a level line — against a level ≥ 1 m below it (a courtyard's roof): that level's cells the plane still stands over within 3 cells, downhill of the face (not past a gable's end, where the plane runs on level along the ridge), join the face, and the face's cells past the eave go to the level where the scan lies on that level (within 1 m; a terrace the scan shows on the roof stays roof), so the wall between them runs along the eave instead of the scan's blur; a column whose centre lies within half a cell's fall of the eave line counts as on it (a flank turned off the grid by a fraction of a degree ends the same way along its length), and a pocket of the level smaller than a part that the joined cells shut in joins too; the face's cells past the eave go only where they reach the level (a strip shut off from it stays the face's). All the gables are judged together, their takes before their gives, a cell two take going to the higher plane (as a valley is drawn): a level they would leave smaller than a part or narrower than 7 m (NARROW_M, a strip between two wings standing as a box) keeps its cells. A wall's blur under a flank (_split_collars, since 2026-11-11): a level narrower than 7 m below a face drawn as flanks (its median over 1 m under the planes there) that the scan lies flat on (within 0.5 m) over less than half its cells is no roof but the scan grading a wall over a few cells — a stage house's 8 m wall banded into a collar round its hipped roof, the Schauspielhaus's at 145 m: it is split at half height, each cell to the face's plane or to the part beyond (a lower face's plane only within 1 m of the scan) whichever its scan lies nearer, so the wall stands where the scan crosses its middle; those cells are not judged when the model's miss is measured (Model.judged: no sharp roof lies on a blur, and the split had sent a small gable at 2 % over its bar back to the flat model). 31 buildings on Dresden's, Leipzig's, Hamburg's and Grimma's tiles lost such a collar or ledge, 0.4 % fewer outline points. An der Frauenkirche 5's ring is now 16 flanks meeting at hips; the block at Schloßstraße and Sporergasse 23 flanks, the scan within 0.5 m of the drawn roof on 70 % of its cells (35 % with the planes first). A face most of whose cells only the flank rule found is drawn as flanks or not at all: the planes or a landmark's surface would draw the dormers' blobs it was found past; not at all, it gives back its flank cells alone, and a sloped plane closed into it is cut again as a face of its own. The surface is written as a 1 m grid in cm from the part's z (its median), two cells past the outline and carried on there on the face's own local plane. Polygonised from the labels smoothed by a 2 m majority (twice: a border diagonal to the grid no longer a staircase of 1 m teeth), coverage_simplify 2 m, clipped to the footprint, then built of walls (regular_cover, since 2026-11-10): the majority rounds every corner and the scan's blur waves every border, so a courtyard read as an oval and a wing's inner eave as a dune's foot. Each edge of 2 m or more is turned onto the nearest of the building's walls (in plan) among those within 15° of its direction — the walls ≥ 4 m of its own outline (a vertex within 0.1 m of its neighbours' line, a roof face's corner left in the union, splits no wall), before another roof's cells are cut from it, walls within 2° of a longer one one direction: a wing's inner edge runs along its own outer wall, not along whichever wall of a bending block is closest in direction (An der Frauenkirche 5's ten directions cut its courtyard's sides into teeth), and an edge within 1 m of that wall runs on it, on the wall's own line (an edge along the outline is the outline, no sliver beside it); neighbours parallel within 1 m (where their edges lie, before one moved onto its wall; within 2°, one front's walls) are one line — two on two walls only on one line, a jog in the footprint stays —, an edge off every wall shorter than 4 m (a rounded corner) is left out, and the lines meet where they cross (two within 2° of each other at a jog, each square to the other), unless a corner would stand 1.5 m off the outline (a round tower keeps its points). The parts are then tiled again, the highest first, each taking what its new outline holds and nothing before it took; a gap none took goes to the lowest part it lies beside along a metre or more and a quarter of its longest contact (a lower roof runs on up to a higher one's straight wall), else back to its old part. Then the parts are one coverage again (clean_cover: their borders noded, a sliver under 1 m² none holds given to the part beside it), their inner borders simplified by 0.3 m (coverage_simplify, the outline kept: the notches where two borders meet a few centimetres apart), and points on a straight edge within the file's centimetre (where a neighbour's border met it) dropped, as one coverage so neighbours drop the same ones of their shared border (_straightened, its edges noded again) — each would be one more wall in the mesh; a step of these fix-ups that changes the parts' union by more than 2 %, halves a part, leaves a part none of the 4 m² it is drawn from, or adds a spike (a corner under 15° between edges over 0.5 m: a sliver two turned borders left), keeps the parts as they were (_holds), and parts laid on the walls that, as drawn, miss DOM1 on more than 30 % of their judged cells or on more than half the LoD2's share are drawn as the scan put them where those miss less (drawn_miss; a structure's roof the same): Dresden's rebuilt roofs have 20 % fewer outline points than before the regularising (the other sites 13 to 32 %). The footprint they are cut to is the building's less what another roof covers, each covered piece whole but not one holding no 3 × 3 cells (a neighbour's eave over the party wall, a notch a metre deep every few metres along it), the cut's raster steps simplified by 0.75 m. Parts standing under 2 m over the ground, or under 4 m² in plan, dropped → data/<site>/dlm/roofs_<tile>.geojson (Polygons with the object id, the roof z and, for a face, its surface; the provider's credit). A face drawn as flanks or planes also carries them as facets (since 2026-11-13, roof_facets.py): each plane's polygon in plan and its plane. The labels lose their strips a cell wide, their raster borders are straightened (coverage_simplify, 1.25 m), then every point of a border between two planes that meet at a crease — the cells either side on their own side of the planes' crossing, 80 % of them — moves onto the line the two cross along, a corner several creases share to their common point (≤ 3 cells); the polygons are cut to the part's outline on a 1 cm grid and noded, each run between two corners one edge (a step's stair within 0.75 m of straight), so two facets share a border point for point; facets that miss the outline by 0.1 % leave the face to its grid. A ridge, a hip, a valley is one straight edge where the grid cut it off between two cell centres. 1 080 objects on Dresden's fifteen tiles, 2 727 parts, 343 of them faces, 319 of those with facets (2026-11-13; 2 731 and 335 before flanks rose square to their walls and blocks on them took planes of their own; 2 741 and 334 before a wall's blur was split, 2 738 and 335 before a landmark's flanks took every building's bar; 2 814 and 367 before the flanks; 1 081, 2 755 and 222 before the plausibility rule; 886 objects, 2 227 parts before the band below 40 %; 2 956 parts before the strips merged; the Altstadt most); the step runs on every site — Hamburg 587 objects (365 before the band), Leipzig 357 (269), Unna 61 (43), Meißen 56 (46), Grimma 17 (13), München 121 (120). bakeCityMesh (withMeasuredRoofs, scripts/measured-roofs.ts) drops the object's LoD2 triangles and stands each part from the object's lowest LoD2 vertex — a flat part as a prism, a face with facets as those (facetPrism: each polygon flat on its plane and shading with the plane's normal, the heights of facets meeting at a corner within 0.5 m one height — a crease, or two planes the scan cannot tell from two that meet (the bake's FLANK_TOL_M), where a step would be a sliver of wall a hand high across the roof; two meeting at a crease there, within 5 cm, are never parted —, a wall where a facet's edge has no neighbour or a lower one), any other face as an error-bounded TIN of its grid (Delatin, 0.25 m) whose inner vertices are triangulated with the outline's points where the height along it bends (Delaunay, the outline's edges enforced by splitting), its walls on those same points, shading with normals smoothed over the face but not across a crease of > 30° (a ridge stays sharp) — same object id, so the table row, picking, demolish, collision and minimap stay; eave and storey rows follow the new shape, and the card says geformt wie gemessen where a face is part of it (MEASURED_SHAPE). The rebuilt objects' triangles: 215 k before and after on the site (a face's TIN costs what its terraces did); the facets take 9 % off the measured parts' own (Dresden 123.7 k → 111.4 k, Leipzig 80.4 k → 74.6 k, Hamburg 130.8 k → 121.1 k, München 8.2 k → 7.4 k, Meißen 3.2 k → 2.7 k, Grimma 2.4 k → 1.7 k, Unna 8.0 k → 7.3 k: a plane is a polygon's few triangles, where its TIN ran along every crease in steps). The sky-view and horizon bake (skyview.py) burns the rebuilt parts — flat tops, faces on their grid — in place of their objects' LoD2 surfaces, so the far shadows match the mesh (roofs runs before skyview). Limits: a face without facets (a vault, a landmark's scan) is the scan's, smoothed: a sharp ridge reads rounded over a metre or two, a dormer or a chimney is smoothed away; such a face's TIN keeps its 0.25 m lumps, which the smoothed shading hides more than it removes. A hip ends where its two planes meet the outline, which may lie a few decimetres beside the outline's corner (the flanks' eaves are fitted each on its own); the step bakes that read a face (skyview, dormers, roof_colour) read its grid, which the facets leave as it was. Judged in headless SwiftShader renders only (before/after), not on a real GPU.

  • Ground-clamp =computed — the loaded terrains' grids (fine level first) sampled to seat trees, lamps, monuments, rails, walls and the player on terrain. lib/city/ground-clamp.ts, heightAt in create-app.ts.

Building detailing (all keyed off CityJSON attrs + the loader's surfacetype)

  • Per-building clay tint ≈assumed (Farbvariation) — deterministic hash(objectid) + function family + measuredHeight nudge → muted per-building wall colour. Source preference: real per-building colour (planned: DOP) would replace the hash; the hash exists precisely so the look survives when function is 86 % "unspecified". Palette: "render" (sand, ochre, soft terracotta plaster) or "brick" (four clinker swatches from orange brick to dark red-brown plus one pale render; in a brick context every building is brick, civic ones too — Hamburg's schools, the Kaispeicher: 471 objects, Dresden 14, München 9; 2026-10-21, G24/G25 —, while structures that are no building, a non-31001 function — tanks, canopies, towers — keep their cool stone), chosen per building by its facade context (below) — no longer per site. The brick swatches are saturated because the shader mixes them 60 % into the pale clay in linear light, where they land on a washed, dusty brick. lib/city/building-tint.ts (at bake time, into the property table's tint), visual-style.ts.
  • Facade context ≈assumed (Farbvariation; 2026-10-01, ADR 0012) — the walls a building without a mapped material wears: what its OSM neighbourhood is mapped as. osm_buildings.py apply_contexts: every OSM outline with a wall material votes — brick for brick; plaster, stone, concrete and wood against it; glass and metal (curtain walls) not at all — weighted by 1 / (distance + 50 m) within 300 m of the LoD2 object (local_contexts, ≥ 6 votes); with fewer, the tile and ≈ 350 m around it vote (area_context, ≥ 20 votes), else plaster. Brick wins at ≥ 50 %. The file carries the tile's vote as a top-level context and an object's own context only where it differs (meta.objects_context counts them); an object with a mapped material keeps its own. The build passes the tile's context (prepare-data.ts), bake-city-mesh.ts uses look.context ?? facades, and a part inherits its root's (inheritedLook). Mapped wall materials in the baked files: Hamburg 106 brick / 11 plaster, Unna 45 plaster / 0 brick, Munich 272 plaster / 38 brick, Dresden 866 plaster / 246 brick. Hamburg's four tiles vote brick, and 128 objects on 32564_5934 whose neighbourhood is mapped otherwise wear plaster; HafenCity's mapped glass keeps its own material (glass does not vote, so its unmapped neighbours follow the brick around them). Replaces the per-site Site.facades switch (🗃️ below). Not yet judged on a real GPU.
  • Rounded walls and roofs =computed (2026-10-19) — LoD2 draws a curved wall or roof as flat facets a few metres wide (the Kathedrale's apses turn 10–20° a facet, the Frauenkirche's drum 20°, the domes of the Frauenkirche, the Yenidze and the Kunstakademie); shaded flat, each facet caught its own share of the sun and the curve read as light and dark stripes. The bake now gives every triangle corner the angle-weighted mean of the faces meeting at its point within 30° of its own (lib/city/crease-normals.ts; the measured roofs round their faces at the same 30°, ADR 0036), per object, walls apart from roofs, only between facets no wider than 8 m in plan that lean to about the same side; a measured face keeps its own normals. Corners of 45° and more (an octagonal tower, a house, the eave), a plain wall's slight kink and a low gable's ridge stay sharp. Only the lit shading changes: the clay lays its windows, relief and glow cells in the facet's own flat plane (normalFlat in visual-style.ts), and a facet turned from the sun takes none, since the buildings cast with their back faces and its shadow-map depth is its own. The welded tiles are smaller: Dresden's city glTF 70.5 → 63.6 MB gzipped. bake-city-mesh.ts vertexNormals.
  • Roof colour =computed ≈assumed (Dachfarbe) — real DOP-sampled colour per building when available (roofColor() + the per-tile LUT, ~83 % coverage), else the synthesized palette (surfacetype==RoofSurface + roofType / Dachneigung → terracotta pitched / slate flat). Bake: pipeline/bake/roof_colour.py. The colour is read from the orthophoto's sunlit texels only, face by face (pipeline/bake/dop_light.py, 2026-10-19): the DOP is flown in sun, and a face turned from it or shaded by a neighbour reads three to five times darker and bluer than its tiles — one median over a whole roof had painted the Frauenkirche's dome slate grey, half the Altstadt's red roofs brown and the Taschenbergpalais' red mansard slopes the grey of its zinc top. A texel is in shadow when dim and blue (B > 1.3 R under luminance 110, or under 40), and a whole face is when the sun meets it at a cosine under 0.17 (red tiles in skylight alone pass the blue test); the flight's sun is the one whose shadows of the LoD2 roofs over the DGM best match the image's near the roofs, searched only along the flight day's solar path (2026-10-21, G2): the fetch writes the days a tile was flown from the provider's DOP metadata (dop/<tile>.flight.json: GeoSN's download-link service, Hamburg's AdV metadata per 1 km, Bavaria's and NRW's DOP info WMS; Berlin none, which keeps the old whole-sky grid), and the sun is scored every 10 min of that day, then every 2 min around the best, at the elevation the clock gives (dop_light.estimate_sun) — Dresden 38.5–38.7°, 173–184° (a whole-sky grid had fitted 33416_5654 at 29° and Grimma's 33342_5678 at 23° under a 38.5° and a 40° sun, 🗃️ below). Flight days: Dresden 2024-03-19, Leipzig 2025-03-09, Meißen 2024-04-07, Grimma 2025-04-04, Hamburg 2024-05-08, München 2024-08-24, Unna 2025-03-19. A fit whose shadows hardly prefer their sun to the opposite one (contrast < 2.8, MIN_SUN_CONTRAST) takes the sun the flight's sure tiles agree on together (pooled_sun: Dresden's 33416_5656, all fields to its east, 33416_5658 and 33408_5654; three of Leipzig's tiles), else none. A flight whose sun-and-sky over sky alone is under 2 (OVERCAST_RATIO; Meißen's 1.87) is read as overcast: no face scaled by its slope, none dropped as turned from the sun. The tile's own unsure fit still finds the roofs' lean (lean_sun). The file's meta keeps flight, sun_from, overcast, lean_sun. Roofs changed by more than 6 sRGB steps: Dresden's 33416_5654 (own fit, 172.9° / 38.5°) 44 %, 33416_5656 (pooled) 58 %, Grimma's 33342_5678 (146.7° / 40.1°) 58 %, Meißen (overcast) 65 %, the well-fitted tiles 3 %. Watch: on Meißen the faces read apart rose 1 987 → 5 875 — striped roofs, unless faces get no colours of their own under an overcast flight. Each sunlit texel is scaled by its face's slope to how a flat roof in that sun reads it (0.7–1.4), so a roof's faces turned to and from the sun come out alike. A building's colour is its faces' sunlit texels together (the old median where fewer than 12 are lit); a face with ≥ 12 sunlit texels and a quarter of it in sun takes its own where it differs from the building's by more than 6 sRGB steps (faces, keyed by the loader's geometry and surface ids), carried in the roof vertex's free _FACADE slots — no vertex buffer more. One flat roof is one roof (2026-10-21, _planes): faces whose normals agree within 0.05 and whose planes lie within 0.3 m, touching at a corner, are read together and take one colour; a rebuilt part stays its own. (LoD2 splits many a flat roof into pieces, and each piece read alone took the colour of what stood on it: Leipzig's department stores read from above as pyramids of three shades.) A face's own colour is read from its core (2026-10-21, _face_mask): its texels two in from its drawn edge, so a face too narrow to have one — a coping, a parapet's top, a gutter strip — reads no colour of its own and takes the building's. (Three to four in a hundred faces read apart were such slivers, coloured by the street or the neighbour's roof beside them.) The face is drawn with an empty margin, so one along the building's edge is eroded there too (the filter repeats an image's edge, and the eave and the coping kept their outer texels). Faces read apart, both changes together: Dresden 38 730 → 33 818, Leipzig 10 741 → 9 157, Hamburg 8 302 → 7 805, München 45 120 → 42 821, Unna 7 749 → 7 067, Grimma 4 183 → 4 013, Meißen 5 847 → 5 365. A roof drawn as its measured parts (ADR 0036) is read part by part instead of by its LoD2 faces (parts, keyed by the part's index in the roofs file; the roof-colour step runs after roofs): LoD2 often lays one flat lid over a whole block of red tiles, zinc and courtyards (the Neumarkt's quarters, the Schauspielhaus), and one colour for all of it read as beige or brown. Where the fetch has DOM1, a face more than 4 m over the ground is read only where the surface model shows its roof (at least half way up from the ground to the face), so a street, forecourt or courtyard inside its outline drops out. A roof face is read where the image shows it: the DOP leans a roof off its outline by its height times the strip's off-nadir angle (Dresden up to 0.15 m per metre, a 30 m roof 4–5 m), so dop_lean.py fits that lean on a 125 m grid from how well the tall roofs' expected shading (bright to the sun, dark from it) matches the image, and each roof's rings are shifted by it before they are read. A wall at most 1.5 m tall standing on its building's roof (a step between roof levels, a parapet, a dormer's cheek) is drawn as roof (withRoofSteps): as sunlit plaster with an eave stroke it read from above as a white dash. A gable is not (a triangle, one top corner over a level foot at the eave, where the dataset keeps it apart from the wall under it): on a shallow roof the rule had coloured 362 gables of 5 m and wider as roof on Meißen's tile, 173 on Dresden's spawn tile. Every flight's photo is first taken to Dresden's colour balance (pipeline/bake/dop_balance.py, 2026-10-20): each flight has its own cast, and every colour read off it carried it — Leipzig's photo is cyan (its sunlit carriageways read r/g 0.79, b/g 1.15 in linear RGB, Dresden's 0.95 and 0.94), so its grey roofs came out blue-green, and München's is yellow (b/g 0.82); Hamburg, Unna, Meißen and Grimma lie within 8 %. The cast is the flight's, not the asphalt's: the photo's grey pixels (its sunlit texels within 12 % of grey) give it to 3 %. The site's sunlit carriageways (the class raster's roads, the brighter half of their texels, pooled over its tiles) are scaled to Dresden's chromaticity per channel in linear RGB, their brightness kept (the exposure stays the flight's), and the whole photo goes through the same factors before its shadows are found (a blue cast reads as shade). Leipzig's factors are (1.24, 1.01, 0.83), München's (0.93, 0.98, 1.12), Dresden's 1 within half a percent; the file's meta.balance keeps them. On to grey (2026-10-23): Dresden's balance is its own flight's, and that flight is about 5 % green (its sunlit roads r/g 0.95, b/g 0.94), so every roof read in it carried a faint green. The roof colours now go on from Dresden's balance to neutral grey roads (dop_balance.neutral_gains, ≈ (1.013, 0.963, 1.026), the brightness kept) — every city's, as every flight is first taken to Dresden's. The water's tint and the sports grounds' surfaces keep Dresden's balance: their thresholds were measured in it. A part raised on another leans with its building (2026-10-23, bases): the landmarks' modelled parts stand on each other — a dome on its drum, a lantern on the dome (Dresden 190 parts, Leipzig 28; no other city's LoD2 stacks them) — and the photo shows them as far off the ground as their building. Leaned from their own lowest point instead, the Frauenkirche's dome (41 m over the square, 0.13 m/m) was read 5 m off its place, on its drum's shade and its lantern, and came out a blue-green grey (sRGB 164, 173, 171; now 200, 201, 197, the sandstone's light grey); the Yenidze's dark glass dome had read the white drum round it. A part whose lowest point lies more than 2 m over its building's foot (the lowest point of it and its parents) and over the DGM1 under it is leaned from that foot, takes its building's lean and measures none (a landmark's dozens of parts outvoted the houses round a lean node). A part on the ground keeps its own foot, one on a slope too (its foot metres over its building's lowest corner, but on the terrain). And any part of a building (a CityJSON BuildingPart) not in sun takes its building's roof, as a face in shadow does (parts_carried): none of it lit, or under a quarter of its roof lit and read the shadow test's blue there — skylight on a light roof, too bright for the test's dimness: the Frauenkirche's corner towers stand in the dome's shadow, read 26 of their 959 texels lit and came out sky blue (sRGB 100, 150, 171), its lantern and the platform round it slate blue. A dark reading stays (the Yenidze's glass dome, most of it dark enough to count as shadow). Discontinued the same day: one colour for a roof's pitched (≥ 20°) and one for its low faces (roofPitched, a fifth texel band), the first answer to the mansards — the face colours cover it, and it still read the shaded slopes dark. Sheds and structures, read (2026-10-21, G5; the bake side): the scan's small structures and DOM1's buildings and roofs on posts are read the same way — the outline at its top (a pent shed by its corners), sunlit texels only, the nearest lean node — into scan / gaps ([index, z, r, g, b], by index in their file and z); one in shadow gets no row. Unna 32408_5710: 503 of 513 sheds, 20 of 20 structures; Grimma 33340: 794 of 804, 53 of 53; München 32690_5334: 130 of 156, 23 of 23. The build takes a row whose index and z match its feature's (within 5 mm; appendedRoofColour, G2 2026-10-21), the card says dop or palette; one in shadow keeps the slate palette. Meißen: 460 of 475 sheds and 61 of 61 gap buildings take the DOP colour. Before, every shed was slate.
  • Roof vividness ≈assumed (Dachsättigung) — raw DOP reads drab/hazy (audited: 38 % near-grey, mean R−B slightly negative). ROOF_VIBRANCE (0.5, Dachsättigung, fixed) applies a hue-preserving chroma boost: it lifts each roof's saturation around the grey axis, weighted so dull/hazy roofs lift most and vivid ones barely (nothing blows out), plus a tiny warm nudge on the muddy greys only. Keeps each roof's TRUE hue — copper-green stays green, terracotta red, slate cool — curing drabness without homogenising toward terracotta. (Rejected: blending toward a terracotta target — it destroyed the ~187 genuine copper-patina-green roofs the DOP captured.) The clay a roof mixes into is a grey (2026-10-21, G6, visual-style.ts clayColour): the pale clay at its luminance, warmed only as far as the roof is warm (Dachwärme, ROOF_WARM_FROM/TO: smoothstep(0.1, 0.4, (R − B) / L)) — the warm clay had turned a cool grey roof warm (Hamburg 49 % of them, München 40, Meißen 36, Unna 31, Grimma 22, Dresden 21, Leipzig 19; now none). At runtime, no re-bake.
  • Storey bands ❝taken ≈assumed (Höhenlinien) — where OSM counts the storeys (building:levels, a part's own or its Building's), the walls up to the eave divided by that count (mappedStoreyHeight), kept when the storey comes out 2.4–5.5 m (outside it the count is another part's or a typo); else storeyHeight(measuredHeight), whole ~3.2 m storeys (storeysAboveGround is only ~4 % populated). Spawn tile: 2 172 objects on mapped storeys (median 3.67 m, p10 3.06, p90 4.79), 298 back on the estimate. None on a facade of its own (flag 512: a church, a palace, a hall, a listed public building OSM counts no storeys of, and the parts that share it): the 3.2 m estimate drew six lines on the Kulturpalast's hall and ten on the Ständehaus, whose storeys are 5–7 m. Storeys from the neighbours (2026-10-21, G21): where OSM counts no storeys (or a count outside 2.4–5.5 m), osm_buildings.py writes storey, the median of eave ÷ building:levels over at least three mapped LoD2 objects within 200 m with the same roof form (flat or pitched) and an eave within max(3 m, 35 %) of its own, else the tile's median of its roof form (at least ten mapped); the build splits the walls up to the eave into whole storeys of about that, held to 2.5–4.5 m (neighbourStorey), before it falls back on the 3.2 m estimate. The card says von der Traufe gezählt, in der Geschosshöhe der kartierten Nachbarhäuser (storey: "neighbours"). Storey counts that match OSM's, each mapped building left out of its own neighbours: Leipzig 34 → 65 %, Hamburg 36 → 49, München 49 → 63, Meißen 53 → 79, Grimma 56 → 84, Unna 84 → 92, Dresden 49 → 75.
  • Gliederung ≈assumed (plinth, cornices, shop zone) — painted in the clay shader from heights the object table already carries, no new data: a 0.6 m stone plinth (the ledges at the first storey line and under the eave are modelled now, with the plinth: see Plinths), and on a building with a mapped shop the ground floor a shade darker as its recessed shop front. Walls only, not on a door, and not on a facade that is its own (flag 16: a Wikidata landmark, a monumental ALKIS use — church, theatre, palace, museum … —, OSM storeys taller than 5.5 m, a hall, or a listed public building, OSM heritage on an ALKIS 31001_3xxx, whose storeys OSM does not count: a court's or an academy's storeys are 5–7 m, and the height's 3.2 m drew twice its rows of windows; a part shares its building's where OSM's count does not fit it, the Kulturpalast's raised roof; ownFacade, partFacade), which gets no drawn door either, and not under a flat roof (flag 256: three quarters of the LoD2 roof area level — post-war slabs and modern blocks have no town house's cornices); the ground floor's three only on a part standing on the ground (flag 128: within 3 m of its tree's lowest base, so a tower's part on a roof gets no plinth). Gliederung, fixed at 80 % (ARTICULATION). The windows are the clay's too (Windows on every house), on a gate of their own: a Wikidata landmark's flag alone does not keep them off, a monumental use or a hall's storeys do.
  • Plinths =computed ≈assumed — modelled, not painted: every open stretch of a LoD2 footprint edge (its outside not in another footprint, so no party wall; 1.5 m and longer; objects 5 m tall and more) on the DGM1 read 0.6 m in front of the wall and smoothed over 3 m (pipeline/bake/plinths.py), cut into level pieces wherever the ground under one spans more than 0.45 m (the steps of a plinth on a sloping street). Each piece's top stands 0.9 m over its highest ground, its foot under the lowest. A stretch already covered by another object's run (within 0.25 m, running the same way: a BuildingPart standing in its Building, two parts on one facade line) is that run's, the tallest object's first, so no two bands stack. Runs end on the footprint's vertices. The building bake lays each piece on its host's LoD2 wall, which stands up to a couple of decimetres off the footprint line and is often turned a little against it, and bends a few centimetres along a long stretch (rays every metre against the host's triangles, wallShiftKnots — one line fitted through five rays sank the 6 cm Gurtgesims into the wall where it bulged; the band is laid piece by piece through the knots the wall needs, thinned: a knot stays only where the line between its neighbours would sink the band more than 1 cm into the wall as read there or stand it more than 3 cm off, so a flat wall is one piece and a tile's bands a fifth of the triangles a piece every metre made; a mitre meets where the two shifted walls meet), and appends a stone band 7 cm proud of it whose front rolls over a 12 cm quarter-round shoulder into its top, shaded smooth (no hard edge), with square ends (lib/city/plinths.ts, appendPlinths); the pieces that meet round a corner of the footprint (turning up to 135°) are mitred into one another on the bisector and share one foot and one top, so a corner closes without overlap. One object per host in a shade a touch darker and cooler than its tint — the form carries it, not the colour (source 5, flag 64); not on a host that is a landmark, glass, metal, flat-roofed or its own colour, lower than 3 m of wall, or standing more than 1.5 m over the ground at its plinth (a part on a roof) — the painted Gliederung's gate. The painted 0.6 m plinth stays under it, hidden where the band stands. Along the same stretches the Gurtgesims is modelled too, in the same register (corniceMesh): level at the first storey line over the host's base (where the wall holds two storeys and the line stands a metre over the plinth's top), a rounded nose 6 cm proud and 14 cm tall over a flat underside, its top washed 3 cm back to the wall, shaded smooth; the pieces in line joined into one run, the runs mitred round corners; a plaster shade barely off the wall's. The painted Gurtgesims is gone: its painted shadow read as a smear. A first cut with hard bevels and ends run past each corner read too hard and overlapped at corners. The Traufgesims follows the same stretches under the eave, where the host's roof runs level along the wall (eaveAlong: the topmost roof face just inside the wall at five points, its plane carried out to the wall; four of them within 0.3 m — not along a gable), from 5 m of wall and 1.5 m over the band below: a quarter-round 16 cm proud rolling out of the wall into a short lip, 32 cm tall, its top flush with the eave. The painted ledge under the eave is gone: it read as a stain under the roof. All three run on from house to house (joinStretches): LoD2's footprints stop a few decimetres short of each other and the bake leaves a party wall's end out, so a row's bands broke at every house. A gap of up to 1.2 m between one stretch's end and the next's start, in line (turning by up to 20°, 0.4 m aside), is closed at its middle; neighbours whose levels lie within 0.6 m (Gurtgesims), 0.25 m (Traufgesims) or 0.2 m (plinth top) take one level, the nearest first, so a row on a slope steps where its houses do rather than drifting. On the spawn tile 2 439 ends run on, most over gaps of 0.4–0.6 m. Dresden: 175 570 pieces on 35 850 objects on fifteen tiles (13 777 on 3 026 on the spawn tile; the files up to 2 MB each). Ground joins: plinths in JOIN_PARTS, budget 1 %. Cornices only on town houses (2026-10-21, G19): the Gurtgesims and the Traufgesims stand only on a town house — storeys from 3.3 m, listed (OSM heritage), or a flat-roofed town house (Windows, below; carriesCornices); a post-war house keeps its plinth but has no cornices. On Dresden's sample tile 558 bands went.
  • Doors ❝taken ◎detected — OSM entrance=* nodes on the ground floor (no level, or one with a 0; not no, entry_only, emergency_ward_entrance) snapped onto the nearest LoD2 footprint edge within 3 m (pipeline/bake/ doors.py), with the outward normal; dropped on a party wall (its outside lies in another footprint), an edge too short for 0.7 m, or a wall too low for the door plus 0.6 m; one door per 1.2 m of wall. The sill stands on the lowest DGM1 sample in front of it. The building bake lays each door onto its host's wall triangles (wallShift in lib/city/doors.ts — LoD2 walls stand up to ±0.15 m off the footprint line, and a wall proud of it swallowed the leaf) and appends a pale surround and a darker leaf set back between it, as two objects in the host's building tree (source 3, flag 64: their tint at full strength). Dresden: 6 628 doors on fifteen tiles (3 208 main, 2 794 yes, 299 staircase), 489 on the spawn tile. Never an invented door: a building without a mapped entrance has none. Ground joins: doors in JOIN_PARTS, budget 1 % (the misses are sills at the top of mapped steps).
  • Passages through buildings (Durchgänge, 2026-10-23) ❝taken ≈assumed — LoD2 draws every building as a closed shell from the ground up, so a gateway into a courtyard, a city gate or the Zwinger's passages to its pond were solid wall. Nothing measures one (the surface model sees the roof over it; the street photos' poses miss by half a metre, Plan 075), but OSM maps them: a way tagged tunnel=building_passage where it runs through the building (pipeline/bake/passages.py). Per way on the ground floor (a level without a 0 or a layer above 0 is skipped; layer=-1 stays) and chains of them merged end to end, the stretches inside the tile's LoD2 footprints (joined across a gap under 2 m, each at least 1 m long, one that meets no outline dropped), run on 0.6 m past either wall — but where the way ends inside the buildings, more than 0.6 m from their joint outline (a drive into a garage, a yard track into its shed), the passage ends there and a wall across it shuts it: run on, it had opened into the hollow LoD2 shell, and one saw through the building; its width from width (1.5–12 m), else 3.5 m for a drive, 4.5 m for a track (a railway, a single track's structure gauge; before, the tracks into Dresden's and Hamburg's sheds ran into 2.5 m slots) and 2.5 m for a walk; the lowest DGM1 sample within 1 m of each axis vertex as its ground. Per LoD2 object the passage's strip reaches — a part of the building beside the way as much as the one it runs through, for the build cuts every tree the passage runs through and a part cut but not lined stood open into its interior — the bake writes the ceiling's plan, the walls along it (and across it at a shut end) and its portals (dlm/passages_<t>.geojson). The building bake (city-passages.ts withPassages, the geometry in lib/city/building-passages.ts) gives it its clear height — maxheight:physical or maxheight where plausible, else the larger of 3.8 m for a drive, 5 m for a track (a contact wire's lowest in a structure) or 3 m for a walk and the ground storey less a 0.6 m lintel (a palace's 5.5 m storey: 4.9 m), never closer than 1 m under the lowest eave over it, none under 2.2 m —, cuts every triangle but a roof's of the building trees it runs through to what lies outside its volume (a prism per segment, mitred at a bend, from SINK.wall under the ground to the clear height over it, so it falls with its ground: the Zwinger's to its pond falls 4 m), and lines it with walls and a ceiling in the host's own clay (source 7, flag 64, no windows, no eave); where two passages cross or run side by side through one gateway, each one's lining is cut out of the other's opening (it had stood across it, and the collider blocked the way). The ground floor's windows keep off a portal as off a door (its width up to 6 m in the door slot), and no OSM door is drawn in an opening. The openings are rectangles: no source gives an arch's form. Last of the building bake, so a cut piece keeps its wall's place in _FACADE. Dresden: 482 passages on fifteen tiles (106 on the Zwinger's), Leipzig 245, München 763, Hamburg 353, Meißen 10, Unna 12, Grimma none.
  • Murals as drawings (Wandbilder, 2026-11-01) ❝taken =computed ≈assumed — a mural OSM maps (tourism=artwork, artwork_type mural, mosaic, graffiti or sgraffito; a node, or a way drawn along its wall, never an area) is drawn as a pen-and-wash drawing after a photograph of itself on its LoD2 wall where three things are known (pipeline/bake/murals.py): its size — OSM width and height, a way's length along the wall, else its Wikidata item's P2049 / P2048 (the item of wikidata, never subject:wikidata, which names whom it shows); a free photo on Wikimedia Commons, at least 500 px wide — OSM image or wikimedia_commons, else the item's P18 (a renamed file followed to its new title); and that the photo shows it straight on — its aspect within 15 % of the size's (FIT), or, for a way, wider than that and showing at least half its height (BAND_MIN): the band of its whole length the photo shows, as tall as the length over the photo's aspect. A slanted shot, a detail, a drawing or a photo of another size fits neither, and the mural is left out: the rule never guesses what a mural shows. The wall is =computed: the outline of the LoD2 building within 3 m — a way follows the ring between its ends' projections the shorter way round, a node is centred on the straight stretch of the outline nearest it and kept on the wall — 3 cm in front of it, its outside on the right (the photo reads from the street), corner stubs under 0.6 m trimmed; it is drawn as long as that wall runs, so a whole photo must fit the wall's length too (a band keeps the photo's proportions). Its height: OSM min_height where mapped; else ≈assumed one storey (3 m) above the highest ground 0.8 m in front of it, lower where the wall has no room, none where the mural stands more than 1 m taller than its wall (a mapping error). The photo is fetched once by bun run fetch (at a standard thumbnail width: upload.wikimedia.org refuses the originals) and resized to 64 px a metre (at most 3 840). A photograph broke the pastel clay model (its own light, its grain, its contrast), so the bake redraws it (pipeline/bake/mural_drawing.py, =computed, one rule for every mural): its own light flattened, its grain calmed by an edge-keeping filter, pen lines where its lightness turns (XDoG, a fleck of ink with too little ink around it dropped), a wash of its own colours (five k-means fields in CIELAB, seeded, lifted into the clay's light register and muted, at 12 px a metre) and an ink, its darkest field darkened. The lines and the wash are committed as WebP beside data/<site>/murals/murals.json, the photo is not; the drawing is an adaptation of the photo under its licence, whose author and licence the HUD's sources list ("gezeichnet nach Foto …"). The viewer (mural-layer.ts) draws each as one strip, a glaze: it multiplies the wall's own lit colour by the drawing taken relative to the clay's base colour (the paper: the wall shows where nothing is drawn), so the wall's sun, shadow, sky, floodlight and night carry over it, and the fog fades it to white, which multiplies nothing. On the full LoD2 only (the far level's walls are simplified by half a metre), the lines one channel and decoded one at a time (at 32 px a metre on a phone), their texels dropped once uploaded, the textures' bytes on the tile's cache budget (2.7 MB for Dresden's two, 2.0 MB on a phone — a photo in RGBA weighed 7.9 and 5.2), hidden with its building (CityLayer.hang); Papier and the Schwarzplan leave it out. Dresden: the Fürstenzug (the band of its 102 m frieze its photo shows, 6.3 m of the 10.5 m mapped with the window zone, a storey up) and Der Weg der Roten Fahne on the Kulturpalast (30 × 10.5 m; it starts lower than a storey up — about 2 m too high until OSM maps its min_height) — 2 of the 37 murals Dresden maps. No other city has one that passes: Leipzig maps 4 murals, Hamburg 24, München 15, none with both a size and a fitting photo (Georg Elser's 23 m graffito has its size, its photos are slanted); Meißen, Grimma and Unna map none. A site without the file draws its walls as before.
  • Dormers and chimneys ◎detected — what DOM1 shows over a LoD2 roof that LoD2 leaves out (pipeline/bake/dormers.py): the excess DOM − roof against its own 9 m median, every blob 0.7 m and more over it, at any size and spacing (2026-10-31; until then 3–30 m², no side over 8 m), clear of every tree point. A building LoD2 drew as a whole over DOM1 (the Taschenbergpalais 1.2 m, flat parts and all) is read against DOM1 raised by that much first: where its cells up to 62° read 0.4–3 m under the roof with a median deviation of 0.3 m at most (only that way: a surface over the roof may be its own dormers and eaves); until 2026-11-02 such a roof failed the fit and had no dormers at all. A part LoD2 drew off in plan (the Taschenbergpalais' north wing about a metre south of DOM1, so that its north mansard's dormers fell on the wall's edge and the eaves) is read a second time moved onto DOM1: a part with steep (over 62°) faces where a move of up to 1 m (0.5 m steps, the shortest of equal fits) brings the median deviation over the whole part to 0.7 of what it was and at least 0.08 m lower, and over its steep faces to 0.7 as well; the dormers that second reading finds on a moved part and the first did not (none of one kind within their width) are moved back into LoD2's frame and kept. Seams are read on LoD2's roof as drawn. On a pitched roof (25–62°, a mansard to 80°) a dormer; a mansard is LoD2's roofType 3400 or, as Saxony never writes 3400, read from the LoD2 roof faces: a face over 62° whose top edge a face of 62° or less goes on from (within 0.5 m in height and plan), where such faces make ≥ 15 % and ≥ 50 m² of the building's roof; only those faces take the 80° cap (a spire's faces meet at its tip, a steep annex under a tall flat roof touches nothing at its top; read from the faces, as the burned roof reads a step between two parts as a steep band); the break between the two is no seam to the dormers on the lower roof. A dormer: its top 0.9–3.5 m over the roof and under the object's ridge, compact, at most 8 m down the slope and as wide across it as it runs; dormers the 1 m cells join are parted at their cores ≥ 1.2 m over the roof; none along a seam of the LoD2 roof (a ridge, hip, valley or step drawn half a metre off reads as a band: on Dresden's 33410_5656 a third of the dormers drawn until then were such bands in hipped roofs' valleys). A chimney: a blob of one or two cells reaching within 0.5 m of the ridge (one as small lower down is a small dormer), up to four cells over the ridge, or one or two on a flat roof clear of its outline and seams; 0.8–5 m over the roof (1.2 m on a seam), at most 3 m over the ridge; on a flat roof not one of three within 4 m (masts on a platform roof, a plant room's ducts), on a pitched roof every stack, however close. Rows (in_rows): a building's dormers falling the same way (within 20°) at one depth (1.5 m) form a row, set straight at its median depth and in one shape (the row's median width, depth and height over the roof; shed or flat as most of them are). One grid a stretch (2026-11-03; a row parts where two stand more than 2.5 times its median gap apart, two wings or a tower between): a spacing of 1.8–7 m about one dormer a slot (0.75–1.3 of the median gap of those 1.8 m or more apart), where the phases of all of them agree best, the slots then fitted by least squares to the nearest dormer on each; taken where half of them stand within 0.22 of a slot (at random a quarter). Every dormer goes on its nearest slot; one whose slot a nearer one takes goes on the free slot on its side, else it is that one's blob (within 0.35 of the spacing, or a fragment narrower than 0.6 of the stretch's median, which also chooses last) or something else and stays (a stair head). On each LoD2 part the row stands at one height: the median roof height and slope under it (the 1 m cells read a mansard's 70° face ±1.4 m up and down: one cell either way). A single empty slot between two set takes a dormer like its neighbour (DOM1 misses a small dormer more often than a roof leaves one slot bare), but not on a chimney or a dormer of a grid; a smaller stray there gives way. Until then (2026-11-02) the row had parted into runs of their own rhythm: the palais' 24 south dormers in five, 2.6–4.6 m, and the dormers off them kept DOM1's ±0.5 m. Where the street photos measured the window axes of the wall below (dlm/windows_<t>.json, within 12 m, facing the same way) the slots are those axes, every one, second or third, if the dormers' phases on them reach 0.8 (on Dresden's 33410_5656 twelve rows stand over a measured wall and none does: DOM1 reads a dormer's centre ±0.5 m, and most walls' axes are unphased). Two rows that read one dormer twice (its front and back) are one, a dormer of a grid kept first; one more than 1.5 m or 20° off its row, or 2.5 times its stretch's median width (a Zwerchhaus), stays as it was. On a face of the palais the height spread under a row (the median of its range) fell from 0.65 m to 0; its south row stands on one 3.47 m grid. Dormers with one grid a stretch: Dresden 13 651 (13 518 before, the filled slots), Leipzig 6 922, München 7 309, Unna 2 080, Meißen 1 802, Hamburg 1 739, Grimma 1 429; the Taschenbergpalais 95. On the exact roof (2026-11-03): last, every dormer and chimney takes its height from its host's LoD2 roof face under its centre (the highest of the faces over it, else the nearest within 0.3 m of its outline: past a ridge, a hip or a mansard's break the neighbour's plane runs on above the face beside it), a dormer its slope too and its top moved with it, a chimney's measured top kept; a dormer whose face is flatter than 25° keeps its row's height (on_planes). On a roof rebuilt from DOM1 (roofs) the host's face is its rebuilt part (since 2026-11-10: a flat part's z, a face's surface grid read between its cells, its slope on the quarter of the 3 × 3 cells around that lies on one plane — beside a ridge, hip or valley not a mix of two flanks —, its ridge the highest of them; a dormer where the rebuilt roof is flatter than 25°, and a dormer or chimney off every rebuilt part of its host, is dropped before the rows are set (RoofPlanes.carries) — the building draws no pitch, or no roof there, to carry it; a chimney there keeps its measured height over the roof it was read on, which shifts may have lifted off DOM1, not its measured top; a part's crest is the highest of its cells and its outline, so a strip narrower than a cell has one); until then such a roof's dormers and chimneys stood at the LoD2 faces it replaces, metres off the drawn roof — some chimneys' measured tops under their own feet (Dresden's rebuilt roofs hold 371 dormers now, 50 before (359 before a landmark's flanks took every building's bar: the Kreuzschule's 12, 2 before); the block at Schloßstraße and Sporergasse 26, 4 before). The burned 1 m roof reads a cell's height, up to half a metre off a 45° face under a centre: over all seven cities 7 % of the dormers stood more than 0.3 m over the roof the building draws (0.6 % over 0.7 m), their cheeks clear of it — dormers floating over the roof — and as many sank as deep into it; now each stands on the face. The Taschenbergpalais' street photos show one identical dormer over every window axis; DOM1 had read their centres up to 3 m apart down the slope and 1–4 m wide. Checked against the 20 cm orthophoto: rows of small gabled dormers now in; DOM1 still misses the smallest, and a roof LoD2 drew flat over a pitched one has none. A chimney is a box 0.6 m a cell (at most 1.6 m) from under the roof to its highest measured point, along the roof's fall or the footprint's longer side (chimneyMesh). Each dormer becomes a front facing down the slope, two cheeks and a roof running back into the main roof: flat at the measured top, or a shed dormer (Schleppgaube) falling with the main roof (lib/city/dormers.ts). Under its roof, 2 cm in, an inner lid facing in that no camera sees: the sun's shadow map draws back faces, so the open box let the sun onto the roof inside it, and the map's soft edge bled that light out as a bright line between the cheek and its shadow (2026-11-03; two triangles a dormer, the same under a chimney's top; +0.3 % on a building tile for the dormers). Its fall (tf) is read along its axis through the centre: DOM1 every 0.5 m over the blob and up to 3 m behind it while it stands ≥ 0.3 m over the roof, the front 1.5 m left out (front wall and eave smear it), a line through ≥ 3 samples over ≥ 1 m. A fall of 8° to the main slope − 8° is a shed dormer; its top is then where its roof meets the main roof (under the ridge), its front ≥ 0.6 m over the roof. A plane over the whole blob read 45 % of Meißen's dormers as shed (front and cheek cells tilt it). 2026-10-21 (G11): 8–25 % shed dormers per tile (Meißen 15.6, Grimma 16–25, Leipzig 12, Dresden 11–14, Hamburg 8–10, München 21, Unna 11–13 %); about half the blobs are under 3 m deep and stay flat. A host's chimneys are appended as one object, its dormers as one per band of feet 0.15 m deep (dormerParts: the sill is read from the object's base), in its tint and roof colour (source 4), no storey band, eave line, plinth or shop zone on them. Where the host has windows, a dormer's front takes one row of its own (dormerWindows, style bit dormer): 0.8 × 1.3 m on a 1.35 m axis, 0.2 m in from the front's ends, its sill 0.2 m over the roof, its head 0.15 m under the dormer's eave (the front's own WALL_EAVE), shorter where the front is lower; the cheeks stay blind (facade length 0). The street photos' shop reading (facade-reading.ts) is the host's ground floor and never reaches a dormer's row: on Dresden's 33410_5656 146 of 237 read buildings carry one, and it had taken their dormers' windows. Dresden: 13 231 dormers (10 399 before) and 19 676 chimneys on fifteen tiles; Leipzig 6 852 and 10 114, München 7 095 (10 233 before: the edges of flat roofs and the roof terraces it had drawn as dormers are gone; checked on the 20 cm orthophoto, where small Altstadt dormers stand under 0.5 m out of DOM1 and stay undrawn) and 5 302, Hamburg 1 755 and 7 746, Unna 2 070 and 3 621, Meißen 1 803 and 1 520, Grimma 1 421 and 1 660. More than half the chimneys stand on flat roofs, where some are rooftop plant or the foot of a higher wall rather than a stack.
  • Eave line =computed (Traufkante) — cornice stroke at min RoofSurface-Z per building (geometry-derived; the attribute is ~4 %).
  • Night light on facades ❝taken ≈assumed (Abendlicht) — after dark (the sun rig's nightFactor) a building of commerce or public use (function 31001_2xxx/3xxx, buildingGlows; structures that are no building stay dark — 22 % of Hamburg's lit objects were tanks and canopies, 16 % of München's; 2026-10-21, G30) is lit from in front, never glowing itself: soft cones of warm light from the wall's foot every 6 m, about half of them on, gone by a storey and a half (Lichtkegel, GROUND_LIGHT); three in five of the landmarks and monumental buildings (flag 16), whole buildings at once and whatever their use, floodlit (Anstrahlung, FLOODLIGHT, isFloodlit in lib/city/city-mesh.ts): cones every 7 m up to the eave over a faint wash, seven in ten on, each a little off its slot and of its own strength — every one of them, evenly, was "schick aber übertrieben" (the maintainer, 2026-10-20). Lit objects: Dresden 8 867 → 8 654, Leipzig 3 036 → 2 970, Meißen 563 → 542, Grimma 401 → 393, Hamburg 3 861 → 3 020 (422 tanks, 373 canopies), München 6 865 → 5 794 (1 004 canopies), Unna 6 156 → 5 587 (549 canopies). A shop's band (Ladenlicht) lights two cells in three and none on a landmark: on the Zwinger and the terrace parts of the Kunstakademie it lit the low walls all round. Streiflicht is off at night (it kept a tenth until 2026-11-06: every wall seen along a street and every roof from the air stood pale against the night). Every wall facing a lamp near the eye takes its light (streetLight, lamp-reflections.ts), in the cone a street lamp's shade throws down, so a house front is lit at its foot and dark above the lamps. visual-style.ts nightLight.
  • Dusk glow (whole walls) 🗃️ discontinued 2026-10-19 — a warm emissive wash over every wall of a commerce/public/special building (0.25 at full night) plus the full-strength rim: the centre glowed beige from afar and every grazing roof shone pale, "Gebäude viel zu hell" (the maintainer). Replaced by the light on the facades above; compared at six night views against dimming the wash (still a lit block) and floodlighting every listed building (too many, too flat).
  • Attributes through the building tree ❝taken — the Saxon LoD2 puts the geometry of 7 950 Buildings (fifteen tiles; 7 866 have none of their own — 2 391 / 2 327 on the four first) on 24 143 BuildingParts (7 239), and the parts carry no function: only the parent does. The bake resolves each object's attributes through its root (the part's own value first, inheritedAttributes in lib/city/building-tint.ts), so tint, roof palette and glow see the building's use; heights stay the part's own. Before the fix every part read as housing: 1 682 parts of 338 commerce/public/special buildings were dark at dusk (646 / 177 / 412 / 447 on 33410_5656 / 33410_5658 / 33412_5656 / 33412_5658). Plan 027 phase 0.
  • Shop fronts at dusk ❝taken ◎detected (Abendlicht) — OSM shop=* and the places to eat and drink (amenity=cafe/restaurant/bar/pub/fast_food/ice_cream/ biergarten) on the ground floor (no level, or a list with 0), joined to the LoD2 footprints by pipeline/bake/osm_buildings.py: the footprint a point lies in, else the nearest within 3 m (entrances sit on the facade line); an OSM building outline carrying one marks the objects it covers ≥ 50 %. The bit rides in the property table's flags column (band 2's free float); the clay shader lays a warm wash under the first storey line (soft top edge, walls only, a ≈3.5 m hash along the facade so a long front is not one strip) at 0.14 of the dusk glow's strength (SHOP_LIGHT) × nightFactor (0.4 until 2026-10-19), a shopfront's pane lit from inside (PANE_LIGHT). No window structure (the grid is a recorded veto, 🗃️ below). 1 555 of 1 855 ground- floor points placed (per-tile sums); 264–323 objects with a shop per tile, 1 188 in all, matched directly — a part also carries its root Building's flags (a Saxon LoD2 Building with parts has no geometry of its own; inheritedFlags, like the attributes), so 2 060 drawn objects show the wash (759 / 359 / 513 / 429). STOP check: 8.2 % of the placed points (2 of a seeded 20) land in a footprint more than 20 m from any street centreline (a courtyard building) — under the plan's 20 %, so the join stays containment-first. Strength 0.4, not yet judged on a real GPU (plan 019). Plan 027 phase 1.
  • Listed facades ❝taken ◎detected (Farbvariation, Traufkante) — OSM building outlines with heritage=* (167 touch the four first tiles) mark the LoD2 objects they cover ≥ 50 % (1 302 over the fifteen tiles, 2026-09-26; on the four first 842 objects: 478 / 28 / 139 / 197 on 33410_5656 / 33410_5658 / 33412_5656 / 33412_5658 — the Zwinger, the Schloss and the Altstadt blocks are many parts each; with a root's flag handed to its parts, 882 drawn: 513 / 28 / 141 / 200). A barely-there warm lift of the wall tint (+3.5 % red at full Farbvariation) and a finer second cornice line 0.45 m under the eave. The official list (LfD Sachsen, Kulturdenkmale_Flaeche) is reachable only as a WMS here: a possible second source. Not yet judged on a real GPU — off if it reads as highlighting. Plan 027 phase 2.
  • Mapped wall and roof colour, wall material ❝taken ◎detected (Farbvariation, Dachfarbe; plan 050) — OSM building:colour, roof:colour (a CSS name or hex) and building:material / building:facade:material / facade:material (normalised to glass, metal, brick, stone, concrete, wood, plaster) on the building and building:part outlines covering an LoD2 object ≥ 50 % (osm_buildings.py; buildings first, then parts, so a part's own tags win; a part without its own inherits its root's, inheritedLook). Colours are brought into the clay's register, never used raw (osmColourTint in lib/city/building-tint.ts): hue kept, saturation ≤ 0.45, lightness held in [0.45, 0.9] for walls and [0.3, 0.8] for roofs, plus the palette's per-building jitter — a red wall is a dusty terracotta, not a signal red. Walls: OSM colour > the material's palette family (brick → brick, stone / glass / metal → cool, concrete → neutral, wood / plaster → warm) > the use family (function). Roofs: DOP colour > OSM roof:colour > palette. For a landmark Wikidata's material (P186) fills in where OSM names none. Counts per tile in osmbuild_<tile>.json (meta.objects_look).
  • Glass and metal facades ❝taken ≈assumed (Streiflicht; plan 050) — the material's glass and metal (OSM's, or glass the street photos found, Glass facades from street photos below) set object flags (4, 8; a part takes its root's mapped material, never its photo glass): the clay turns a little cooler (×(0.93, 0.99, 1.07), metal at 60 %) and smoother (roughness 0.42 glass, 0.5 metal), and glass mirrors the sky (Spiegelung, sky-reflection.ts: a mipmapped cube map of the sky dome, 10 % face on to all of it at a grazing view; metal a third as much; fixed at full, REFLECTIONS — below it the old pale Fresnel³ sheen at the rim's strength would show). No panes, no mullions, no texture and no drawn windows (Windows on every house leaves glass and metal plain) — the window-grid veto holds (🗃️ below). osmColour / clayGlow in visual-style.ts. Conservative strengths, not yet judged on a real GPU.
  • Glass facades from street photos ◎detected (ADR 0060) — where no source names a building's material, its walls are read on the street photos of every camera (facade_glass.py): bun run fetch asks Mapillary for every photo near the tile (<t>.photos.json), picks per LoD2 outer wall the photo of each of up to eight sequences that sees it best — twice the four the decision keeps, since a tree or a bus that hides a view shows only in the reading — (within 60 m, ≤ 60° off its normal, near enough that its pixel at the wall's nearest point spans no more than the reading's 10 cm — a panorama at ~33 m —, a free line of sight, a framed photo with ≥ 30 % of the wall's upper part in its frame), turns its upright thumbnail and segmentation back into the frame its pose is in (the EXIF orientation: ~6 % of Dresden's framed photos are stored upside down; a mirrored one is not read), rectifies it onto the wall (facade_measure.rectify, the camera's own model) and asks CLIP (ViT-B/32, ONNX) whether the upper wall, where the photo sees a quarter of it, is "a glass facade" or "a wall with windows"; the probability is cached per photo and wall. A photo Mapillary drew no building in (none of the uploads before 2021 on the Congress Center's tile, many since) is read on the wall the geometry puts in it, less what its segmentation found in front, and only where a third description, sky, trees or a street, is the less likely (a crop of sky above a cantilevered roof read 0.92 glass against the two walls alone). The cache is <t>.glass.<key>.jsonl, named by a hash of the prompts, the model, the crop's limits and the rectification's, and ends with a line saying every photo its plan picked was read, for the walls and planner limits it was planned for (plan_key), and which photo it picked for which wall — the decision reads those alone, not one an earlier plan under other limits picked (where the cache is missing or not read through, the windows bake keeps the glass its file holds rather than drop it, and says so; a wall planned later on a photo already read, or numbered anew by a new LoD2 edition — a wall is its object, number and length —, is read at the next fetch); the photo is deleted. windows.py keeps one reading a sequence (the photo that saw most of the wall), of the four sequences that saw most, and calls a wall glass where at least two such readings exist and two thirds of them reach 0.95 (wall_glass; at 0.9 metal cladding, ribbon-window blocks and a plastered house with balconies came through); the wall gets no window model: a curtain wall's mullions are no rhythm for the stone wing, the house next door, the dormers over it, or its own house where a glazed stair tower too short to make it glass or a material OSM names keeps it from drawing as glass. A Brown camera's parameters are OpenSfM's (fx, fy, cx, cy, k1, k2, …): its focal and radial terms are read from there, and no lens is read past where its distortion turns (a strong barrel folds the world beyond back into the frame); a framed photo's frame is its own shape where Mapillary gives it (the photo list is asked anew when the fields it asks for change). A photo the network, the token or the model's loading failed is read again at the next fetch; one the CDN or the API refuses on its own, or a wall whose crop the model fails on (a worker out of memory over a very long wall; the photo's other walls are kept), is read again and given up after three fetches since it was last read — a first failure is not counted where ten or more and most of a run failed: that is a proxy's, the CDN's or the machine's failure. One Mapillary has taken down since (HTTP 404, 410, or its "does not exist" answer) is cached as skipped. A photo is planned at its thumbnail's own size, which is smaller where the original was. The building bake gives an object material glass, from photos, when its glass walls make at least half the length of its walls read and a quarter of all its outer walls, or 15 m and a twentieth of them (windows_<t>.json's outer, every outer edge however short; lib/city/photo-glass.ts) — a stone house whose one glazed stair tower was all the photos read stays stone, and so does a mall with a 16 m glazed entrance in 547 m of walls; a file without outer asks the 15 m alone. The photos' glass is the object's own: a Building's passes to none of its parts (inheritedLook), each part is read on its own walls. It then draws as Glass and metal facades above. On the probe (Dresden, 2026-11-10) every known glass front read 0.91–1.00 on its clear views (Prager Spitze, Glaskugelhaus, Gläserne Manufaktur, Ammonhof, Congress Center) and no plastered or stone wall above 0.9 but one clad in white panels; the facade profiles' openings share had read a stucco house at 0.99.
  • Landmarks from Wikidata ◎detected ❝taken (HUD Erkunden → Orte; plan 050) — bun run fetch asks Wikidata per tile (wikibase:box, instance of a subclass of architectural structure Q811979, ≥ 2 sitelinks, top 80; height P2048, material P186, building = subclass of Q41176; a box that times out is asked again in quarters) → data/_raw/<provider>/wikidata/ landmarks_<tile>.json (CC0). pipeline/bake/landmarks.py matches each item to the LoD2 objects that draw it — OSM outlines tagged wikidata=<Q> covering objects by half, else, for a building, the LoD2 building under its point (root and parts), else the nearest (below) — drops what matches nothing drawn (districts, streets, harbours) and keeps, most notable first, those with at least NOTABLE_SHARE = 8 % of the sitelinks of the tile's most notable one (never fewer than MIN_SITELINKS = 2; MAX_PER_TILE = 40 is a safety cap, not the rule): a relative floor per tile, so Unna gets its own (its churches have two sitelinks each; an absolute floor gave it one, 🗃️ below) and a dense old town keeps its second rank (a fixed 12 per tile cut it, 🗃️ below). → data/<site>/dlm/landmarks_<tile>.json. The build flags the objects (OBJECT_FLAG_LANDMARK = 16; no look of its own), gives them the Wikidata material where OSM names none (withLandmarks, lib/city/landmark-lut.ts), and writes the site's twelve most notable into the tileset's extras.landmarks (siteLandmarks) — one per building drawn: an item whose LoD2 objects a more notable one already holds is that one (the Rüstkammer and the Staatliche Kunstsammlungen in the Residenzschloss, the Galerie Neue Meister in the Albertinum), its name kept in the host's also for the search. The HUD merges them with the authored vantages into one list, Orte (lib/city/places.ts sitePlaces): a landmark a vantage names — every word of one in the other, German endings, fillers, the city's name and Wikidata's brackets aside ("Zwinger" in "Zwinger & Semperoper", "Dresden Hauptbahnhof" in "Hauptbahnhof") — or one within 60 m of where a vantage looking down ≥ 20° looks is folded into that vantage, the rest follow the vantages; a click on a landmark glides to an aerial vantage from the south-south-west (landmarkVantage: altitude clamp(h + 50, 60, 300) m, pitch −22°, fov 55°). Examples: the Elbphilharmonie (41 sitelinks, material glass), the Chilehaus (35) and St. Michaelis (27, brick) in Hamburg, the Frauenkirche in Dresden, the Lindenbrauerei and the Zentrum für Internationale Lichtkunst in it in Unna (Wikidata's point lies beside the building: matched to the nearest within 10 m, within 25 m only where OSM names it alike — a word not generic shared with its OSM name; 2026-10-21, it had taken Krenkelstraße 7 for "Mietvilla Krenkelstraße 3"). What is gone stays out (2026-10-21, G31): the fetch also asks each item's P576 (dissolved/demolished) and P5816 (state of conservation) (fetch_status, asked apart, so an older cache is completed in place, complete_status); an item demolished or destroyed (P5816, or P576 without a standing state) is matched only through an OSM outline tagged with it (gone) — its point had flagged the building on its site now (Dresden: the Alte Reformierte Kirche on the Neues Rathaus's 30 objects, the Gothic Frauenkirche on 59, the Opernhaus am Taschenberg on 65, the Palais Brühl, the Sophienkirche; München's Braunes Haus among the site's twelve, now the Museum Brandhorst). Dresden's listed landmarks 284 → 238, flagged objects 1 411 → 1 342 (37 demolished items dropped, 13 by the 10 m rule); Leipzig 76 → 69, Hamburg 68 → 63 (the Bucerius Kunst Forum, 11 m off, is lost), München 124 → 121. Districts are kept out by matching only OSM outlines of buildings, building parts and man_made structures (HafenCity, Wachwitz got in through their place outlines before). On a dense tile the floor follows the tile's top (2026-10-01): an old town's second rank stays on the list beside the Frauenkirche, the Zwinger and the Semperoper, where a fixed 12 cut it (Dresden's tiles now list up to 40). The roof relief stays with each tile's 12 most notable (RELIEF_LANDMARKS): over an ordinary villa among old trees the crowns would read as its roof. Dresden's Congress Center is not in Wikidata's answer for its tile (the query keeps a box's 80 most linked items), so it is still not listed. Coverage follows Wikidata's, which varies by city; a site fetched before the query existed needs bun run fetch <site> again (it only adds what is missing) before bun run bake <site> --step landmarks.
  • Roughness jitter ≈assumed (Materialstreuung) — hash(objectid) → roughness clamped to [0.55, 1.0] (stays matte).

The twin: identity, facts and provenance (asked on demand)

The scene says nothing on its own; asked (a click, a long press, I under the pointer), it answers in a card (ADR 0042, plan 052).

  • Object identity and semantics in the tileset ❝taken — CityJSON attributes → ten columns of the building glTF's EXT_structural_metadata table (lib/city/object-facts.ts, scripts/bake-city-mesh.ts): buildingId (the Building's gml:id; a part carries its Building's — the Saxon parts' UUID_… ids are dropped), function (ALKIS Gebäudefunktion, the root's for a part), roofType, created (the object's export date, date only) as ENUM columns of the codes a tile uses; height (measuredHeight, else the model's top − base), roofPitch (Dachneigung), area (the GroundSurface rings' area) as FLOAT32, noData −1. A scan structure (plan 034) carries its key, height and area only. The code lists' German labels: lib/city/adv-codes.ts (AdV CityGML profile). Spawn tile +46 kB gzipped (+3.5 %). No look: read one row when asked. Found on the way: the bake read the footprints after the CityJSON loader had rewritten the document (a Solid's semantic values come back flattened), so every Solid lost its GroundSurface — on the spawn tile 2 012 of 2 041 BuildingParts had no minimap footprint (and would have had no area); the footprints are now read before the loader runs.

  • How a building is drawn, and from what ❝taken — the bake's own decisions behind the walls, roof, storeys and windows, which the style columns carry only as their outcome: where the wall colour came from (OSM building:colour, else the palette of OSM's building:material, the neighbourhood's brick, or the use's), the mapped material (or Wikidata's for a landmark), where the roof colour came from (DOP, OSM roof:colour, palette), where the storey came from (OSM building:levels, the street photos, the estimate from the height), where the window rhythm came from (its own walls, another part, a neighbour, its type) or why there are none (a facade of its own and why — its use or a hall's storeys —, glass or metal, a low wall, a use without windows, no window fitting the storey), and the landmark's Wikidata item → one ENUM column drawn of the same table, a short key=value;… text (lib/city/drawn-facts.ts, drawnOf and assignWindows in scripts/city-windows.ts). The card's Daten section reads it with the object's style row and lists every drawn value with its source, its badge and a link to the place (OSM's objects there, Mapillary's photos there, the Wikidata item: lib/city/drawn-lines.ts, plan 052) — the values as the bake wrote them, never recomputed. Added 2026-10-19 so a strange rendering can be traced to its source from the card.

  • Names, addresses and storeys ❝taken ◎detected — OSM building outlines (name, addr:street + addr:housenumber, building:levels) covering ≥ 50 % of a LoD2 footprint (the one covering most), else the addr:housenumber points on it (inside, or ≤ 3 m off the facade), grouped by street (pipeline/bake/osm_buildings.py → osmbuild_<tile>.json, per object, never handed to the root) → the name, addr, levels columns. Spawn tile: 2 363 of 4 404 objects addressed, 444 named, 2 470 with storeys (BBBike extract 2026-09-26). The card merges a building's parts.

  • Ensembles ◎detected (2026-10-19) — LoD2 Buildings that are one building to a visitor: the cadastre splits Dresden's Residenzschloss into three (the main complex, the Hausmannsturm's spire one of its parts, and the tower's shaft among them), and OSM maps it as one named outline. osm_buildings.py tree_outlines writes on each root the key of the named OSM building outline covering at least half of its whole tree (outline, w<way> / r<relation>); lib/city/ensembles.ts joins trees sharing it, and, for a wing the outline misses, trees of the same name whose footprints touch (≤ 0.5 m, overlap or crossing) — never two trees whose outlines differ (two halls named alike). Joined in bake-city-mesh.ts → the ensemble column (the tile's lowest root, its facts the tile's whole ensemble's: the union of the footprints, lowest base to highest top) and group (the outline's key hashed, FNV-1a: the same in every tile, so a building a tile edge cuts in two is asked, outlined and demolished in both tiles on screen; the card takes the highest of the tiles' tops and sums their grounds). Asked, outlined, listed in Hier auch and demolished as one; the card counts its Buildings (Gebäude: 3 im Stadtmodell) and keeps the picked one's buildingId. Unnamed outlines are left out: OSM often draws a whole block as one (up to 38 LoD2 Buildings under one in Dresden, each house with its own address). Dresden, extract of 2026-10-17: 262 named outlines over 732 trees, 4 of them across a tile edge (the World Trade Center among them); the name rule alone had made 292 groups, 14 of them not one outline. A first cut joined by name and touch only (no outline, not across tiles); the outline replaced it on the maintainer's ask.

  • Provenance manifest ❝taken — data/provenance.json (hand-kept) → provenance.json next to the tileset (lib/city/provenance.ts, at build time): per source its label, credit and licence, per tile its edition (LoD2 split into the model year and its inputs' years). The card's source lines; fetched with the first question.

  • The asked building ≈assumed (no slider) — the picked object's building tree, or its ensemble's trees (the set demolish takes), gets flag 32 in the packed object texture at runtime; the clay lifts it 40 % towards paper white, adds a faint paper light (brighter at night) and draws a pencil hatch: near, strokes on the building every 0.9 m (up the facade at 45°, across the roof; fwidth-constant); where those would crowd closer than a few pixels, 45° strokes every 7 px in screen space instead, so the mark reads from the air too. A first cut (0.45 m strokes, a flat graphite wash far off, a 35 % lift) was checked headless: the wash and the lift cancelled, and from 100 m the building looked unchanged. Not in Papier (its override material). Not yet judged on a real GPU.

  • The inquiry card ❝taken (asked by a click, a long press on touch, or I under the pointer — at the centre without one — no mode since 2026-10-01; the HUD's card surface, no longer a warm paper sheet; a tap slightly off a small house still finds it — two rings of rays vote; a bottom sheet on touch screens) — lib/city/inquiry.ts → inquiry-card.tsx: the OSM name or the ALKIS use as the title, the address, then only the facts that are known (the Building's height — its tree's base to top unless the survey measured it —, eave, roof form and pitch, the ground area of its parts' union, overlaps counted once, mapped storeys, parts, Kulturdenkmal, a shop on the ground floor), the buildingId to copy, and one source line per source quoted, with edition and licence. Unspecified use (86 % of the objects) is said as such, never guessed. A roof rebuilt from DOM1 (ADR 0036) is said to be measured there: no LoD2 form, pitch or height, DOM1 as its source.

  • Tree facts ❝taken (plan 052 phase 4) — the street-tree register's art_deutsch, art_botanisch, name (the location), standort_nr, jalter (age), aend_dat (the record's date) and which of height, crown and trunk it measured rather than the bake filled in → treefacts_<tile>.json (pipeline/bake/trees.py), columns aligned with the trees file, names, places and dates as tables; an OSM tree its tagged taxon, German name and sizes. ≈ 37 kB gzipped a tile, fetched with the first question about a tree there, never to draw. The card: species, location and tree number, measured sizes only, age with the record's date.

  • Asking trees, monuments, bridges ❝taken — lib/city/ask-items.ts builds each tile's askable things from the features the dressing already has: a tree's trunk and crown cylinders (as the inventory sizes them; packed per 64 m cell), a monument's marker or measured form, a basin's prism; bridges are met on their drawn meshes (bridge-ask.ts), the point naming its deck. The card (lib/city/inquiry-features.ts): a monument's official name (Basis-DLM), basin form (OSM), measured height; a bridge's name and measured deck (Basis-DLM, DOM1), structure and main span (Wikidata or OSM), clearance (OSM), keyed by its Wikidata item.

  • Asking the traffic layers ❝taken — while a data layer shows, its flows and counters answer like the rest. A counted section is met on its drawn glass bodies, on whichever terrain level shows them (traffic-ask.ts; the bodies' BVH built on the first question that reaches them), the vertex the ray meets names its section (userData.trafficSection, a CPU-side table — WebGPU allows eight vertex buffers and the bodies use them). A counter is met on its columns as tall as they stand now (bike-ask.ts, cylinders from bikeColumns). The cards (lib/city/inquiry-traffic.ts): a section's vehicles a day both ways, per direction named by the compass point it heads to with its heavy-goods share (where the source counted per direction; an even split says so instead), the year and how it was counted, the scene's hour as an estimate (≈ the day's count on the typical curve, rounded to tens, the curve's source named) and the layer's credit; a counter's bicycles in the last hour per direction, when they were counted, and whether that is older than three hours. The outline takes a section's own triangles, drawn with the layer's own position node (grown with the hour, widened from the air); the glass writes no depth, so its mask tests only what stands in front. Trams are not asked yet.

  • The outline =computed (no slider) — whatever is asked, building, tree, monument, bridge, traffic flow or bicycle counter, gets one line along its silhouette as the camera sees it now, not along its edges: the element's triangles (a building's own out of its tile's mesh, a bridge's out of the tile's bridge meshes within 12 m of its deck; a tree's crown ellipsoid — a cone for a conifer — on its trunk, a monument's cylinder, a basin's prism) are drawn into a mask where the scene pass's depth shows them (a stand-in only where the scene's own surface lies inside it), the mask is blurred at half resolution, and the band of the blur between 0.07 and 0.66 is the line: 4.5 CSS px in the HUD's accent pink (SELECTION_ACCENT, the theme's --sidebar-primary; the hatch on a building is --primary), on a hair of its pale outside (0.012–0.07, 55 %), anti-aliased by its own gradient. The selection leaves the city's palette on purpose: it is the interface's, and the card's kicker and the chosen candidate carry the same colour. A new question flashes: the pale fringe widens to 0.002 and glows in the accent, easing out over 650 ms (OUTLINE_PULSE); a hover over a candidate does not. (Graphite and paper until 2026-10-07: maintainer review asked for a bolder mark.) The blur rounds the corners and drops what is finer than the line; its reach follows the device pixel ratio, so the width is in screen pixels at any distance (lib/city/outline.ts, selection-outline.ts, selection-shape.ts; over the finished frame, before the output transform, so every picture style shows it). It replaced the pencil loop (🗃️ below). Not yet judged on a real GPU.

  • The nameplate: the thing asked, named in the city ❝taken (2026-10-31) — what the card's title and address say, set as lettering in the scene beside it (lib/city/nameplate.ts): a Wikidata landmark (flag 16) by its name large, with up to two lines of facts under it — "erbaut 1726–1734 · George Bähr", "wiederaufgebaut 2005 · 91 m" (each fact where Wikidata has one, the card's height where it has none, factLines in lib/city/landmark-facts.ts); a named building, bridge or monument by its name with its address or kind; an ordinary house by its address, small; a tree, a lamp, a bench, a stop, a landing stage or an unnamed monument by what its card calls it, with one line (a tree's species from its tile's facts file, the one the card reads, else its kind); nothing for a house without an address, a counted street, a design. The facts are a site file baked once (landmark-facts.json, Wikidata CC0: inception P571 or opening P1619, architect P84, height P2048, and the P793 events that begin, end or span the building's construction or its reconstruction, each with its own dates — years only where Wikidata states the year, not a decade or a century; an end more than 100 years after the start left out; a rebuild with no end stated reads "wiederaufgebaut ab"), matched to the Buildings through the landmark tiles' objects, fetched with the first landmark asked (an organisation's inception, often its founding, stands back where construction events date the building). Where several items share the asked Buildings, the one the card's title names, else none. The card lists the same facts as rows (Bauzeit, Architekt, Wiederaufbau, Höhe laut Wikidata) with Wikidata as their source, so the plate never says what the card does not. Where it stands follows the camera (lib/city/nameplate-place.ts), in multiples of the building's size S (its height, at least half its long side): looking down on it steeper than 50° from over its top (from the air or in Modell) it lies on the ground beside the footprint, reading along the footprint's axis nearest the screen's — but only where that ground is seen steeper than 35°, else it stands up (lettering on the ground seen flat could not be read, maintainer review 2026-11-01); in the street it stands upright before the facade the camera faces best (1.5 m, less for a small thing; kept until that facade is seen flatter than 25°), standing on it like rooftop letters — a landmark's under its tower on the highest eave (in height) of its broad LoD2 objects where that lies clearly under a slender one's (never a lower annex in its stead; a tower: twice as high at its eave as the square root of its own footprint; the carrier's eave 1.75 times lower, each over its object's own base, and under the tower's in height; a root's or an ensemble lead's area is its whole tree's, so it is never judged slender), where that lies a quarter of the height or more under the top, so a church's name stands on its nave, not on its spire's tip; any other building's stand on the top; else on the top — or beside it in the sky (no line to it: the selection outline marks the thing, in the lettering's ink); within 1 S an inscription before the facade, its foot a quarter of the height up (2.5–15 m); beyond 8 S its name alone. The letters are built, not printed: 0.14 of their capital height deep, their returns lit by the sun, casting a shadow on whatever stands behind them from the sun, down to the ground and only where the sun reaches. A long name breaks between words into even lines (an A / B name at its slash). Thresholds hold 20 % of hysteresis; a new place must stay the best 900 ms before the plate moves (maintainer review 2026-11-02: it moved too often); the place is chosen every 250 ms among its candidates by what of it would be off screen, hidden by the city (rays through the buildings' BVHs, never a tile whose BVH is being built), under the HUD's panels (data-scene-cover) or over its own building, and a change cross-fades. Its capitals are 9 % of the footprint's long side, kept on screen between a size per rank (a landmark 36–96 CSS px, 26–56 on a phone; its facts at 0.4–0.44 of that). Where the city covers it anyway, it shows as a pale outline (the scene's depth). How it is set follows the picture style (lib/city/nameplate-type.ts): Pastell and Papier the HUD's grotesque (Inter, heavy, set tight) with tracked capitals for the facts, in loud colours — Pastell the selection's magenta with plum returns and its facts in the same magenta lifted a fifth toward white (plum facts read poorly over a pale facade, maintainer review 2026-11-02; translucent ones showed their own plum returns through), Papier a printer's violet (vermilion for a day, 2026-11-01: the selection's hue reads as one hand) — and no halo (a book serif with a pale halo until 2026-11-01: maintainer review found it old-fashioned and too much like a tooltip); Comic a caption box (Bangers, yellow, its returns black; opaque, so it keeps off the building and stands over it), Noir and Sin City a condensed grotesque (Oswald) with a monospaced line, Sin City's returns red, Strich spaced capitals, Schwarzplan on the ground like a plan's lettering from above the building where it reads; by night the ink turns light and dims a little, and the shadow goes with the sun. The selection outline takes each style's lettering ink (Schwarzplan red on white, Sin City red, Comic black on yellow, Noir white, Strich graphite, Pastell and Papier the violet). Drawn after the frame (nameplate-layer.ts, rendering.md). Not yet judged on a real GPU.

  • Stadtwissen ❝taken — what the twin knows, as files an agent reads (ADR 0055): per tile every root Building that OSM names or addresses, that is listed (heritage) or that is part of a Wikidata landmark — its gml:id, OSM name, address, storeys and shops, the ALKIS use and roof form decoded (adv-codes.ts), the height (the largest measuredHeight of the Building and its parts), its position (the extent's centre and base) and Q-id —, and the tile's named monuments (stadtwissen_<tile>.json, scripts/bake-stadtwissen.ts); and per site the landmarks with an excerpt of Wikidata's statements in words (instance of, architect, style, named after, location, creator, material, commemorates, inception, dissolved, significant event, visitors, heritage status, motto, opening, height, current use), the German article and the sources with their licences (stadtwissen.json). The excerpt is fetched once (bun scripts/fetch-stadtwissen.ts <site>, wbgetentities, CC0) and committed (data/<site>/wissen/wikidata.json). Dresden: 15 tiles of 150–360 kB, the index 168 kB, 238 of 238 landmarks with facts. No look: the viewer never loads it; bun scripts/stadtwissen.ts <site> such|nah|wahrzeichen|quellen searches it.

  • Erzählte Spaziergänge ❝taken (the facts; the narration is none of the four methods — generated text, marked as such wherever it shows) — walks written ahead by Claude from the Stadtwissen and the sources each claim names, committed as data/<site>/walks/<id>.json (lib/city/walks.ts) and bundled into walks.json by the build. A stop is a view (eye in the site CRS on the DGM's ground, the point it looks at, lens, hour, picture style; never Modell), a narration of at most 560 characters, a closer look, notes for days of the year and seasons, and its facts with an https source or a Stadtwissen id each. The walk card (walk-player.tsx) glides there at 2.2× the pace, sets today at the stop's hour and the style, tells the stop and today's notes in a book face (Newsreader, self-hosted) and, if asked, reads it aloud from the start of the glide, folds out the sources, and says on every stop that the narration is Claude's. scripts/walks.test.ts holds every walk to its data (form, extent, ground within 0.2 m, references). Dresden: four walks, 31 stops (Augusts Bühne, Narben und Nähte, Gold, Milch und Eigensinn, Wunder am Elbhang); one each in Leipzig (6 stops), Hamburg (6), München (5), Meißen (3), Grimma (3) and Unna (2); Berlin none yet; how they are written: walks.md.

Vegetation

  • Cultivated land ❝taken ≈assumed (plan 028) — OSM landuse=allotments|orchard|vineyard (and leisure=garden plots inside a colony) → pipeline/bake/cultivated.py: cultivated_<tile>.geojson (colonies, parcels, orchards with their trees, vineyards with their rows) and a 2048² colony raster (cultivated_<tile>.png, two bytes per texel: the signed distance to the edge of the garden land and the colony's long axis; the viewer gets it cropped to the colonies, phones at half resolution). No new land-cover class (ADR 0010): dressing only. Allotments get little gardens in the terrain's fragment pass (cultivated-layer.ts colonyGarden, a node function): the colony's edge is the baked distance field (the colony less its paths — OSM footways, paths, service roads, tracks — roads, rail and water), sampled LINEAR and faded over a metre with a slow wobble, so it is soft and organic from walking height to 200 m up; inside, everything is analytic in the data frame — plots ≈12 × 17 m (a jittered Voronoi in the colony's axis frame, borders meandering by a domain warp and drawn as thin soft paths), each a lawn in one of a few soft greens, a third with warm vegetable beds, some with sparse pastel flower dots, the rest with darker shrub mottles, and a faint hedge green inside the rim. Every line and band is box-filtered over the pixel footprint; with distance the detail gives way to the plots' tones and then to one calm colony tone (no moiré). Redesigned 2026-09-25 after the maintainer looked at it on a phone — see 🗃️ Allotment bed bands. A mapped parcel takes its own axis and a 0.5 m seam along its border. STOP measured: of the 66 colonies (77.9 ha) in the four first tiles, 0 carry mapped parcels — and 0 of the ≈195 (208 ha) over the fifteen tiles (2026-09-26; the four first tiles' counts: the 351 leisure=garden areas lie elsewhere; inside the colonies OSM maps sheds, 196 footways and 123 fences) — under the plan's one-third, so the plots are invented texture, kept low in contrast (COLONY_GARDEN.strength 0.85 of Bodendetail), and no parcel outline is claimed as data. Colonies (2026-10-21, P23): the plots' axis follows the colony's mapped paths (footways, paths, service ways and tracks inside it: their length-weighted prevailing direction, at least 40 m of them, cultivated.py path_axis), the long axis without them; plots ≈ 15 × 24 m (COLONY_GARDEN; were ≈ 12 × 17). Turned more than 20°: Dresden 7 of 24, 9 of 47 and 9 of 30 colonies on three tiles, Meißen 5 of 19, Grimma 3 of 19 and 2 of 6, Leipzig 1 of 4. Orchards (2026-10-21, P21): where the tile has DOM1, a tree at each crown peak inside the orchard (crown_peaks, orchard_dom: nDOM smoothed 1 m, 2–12 m tall, off LoD2), as tall as measured, its crown 0.85 × its height (2–9 m) (src: "dom", m: 1 in the viewer's file); without DOM1 the mapped natural=tree inside, else a grid 8 m apart along the long axis, centred; either way less the spots a measured tree already fills (a canopy, scan or inventory tree of any tile within 4 m, or within its crown radius when wider — so the step runs after the canopy, its scan crowns and the register). Their genus from the orchard's trees=* (else produce=*); a Christmas-tree plantation's trees evergreen cones. Dresden's 33408_5654: 30 grid trees → 4 measured; the four invented trees of Dresden's embankment "orchard" became the one DOM1 shows; Unna 11 → 1. Drawn by the tree layer as the cadastre's "small" archetype (a round crown on a ≈1.3 m stem), a measured one at its measured size; asked, a measured tree's card gives its height and crown and names DOM1 for its place, height and crown (inquiry-features.ts orchardCard, lineage.ts orchardLineage; before, it said the orchard's default). (Before: 5 orchards and 35 trees over Dresden's fifteen tiles, all on the grid or mapped.) Vineyards (2026-10-21, P22): row direction and spacing from the orthophoto (dop_rows: an FFT of 40 m windows, the direction most windows agree on within 15°, the spacing where one wave stands alone, 1.5–3.6 m); else down the fall line below 25 % slope and along the contour above (slope_layout; the DGM's mean gradient over the whole polygon, read from every committed DGM it touches), 2.3 m apart (were 1.8 m along the contour everywhere); dir, gap, src in the file. Meißen: 12 of 15 from the photo, the gentle vineyards turned 61–85° to their fall line, the terraces unchanged; rows 890 → 597. Dresden's rows 305 → 252 (33416_5654 turned 83°), Hamburg's 30 → 25. Drawn as chains of low boxes 1.3 m tall and 0.5 m wide, chunked into 250 m cells, and the seasonal bare canes wait for plan 025's season plumbing. ≈18 s per tile. Not yet judged on a GPU (the chessboard check is the plan's other STOP).

  • Tree/hedge rows ❝taken ◎detected ≈assumed — Basis-DLM (or OSM) hedge & tree-row lines → instanced sets (Instances), chunked into 250 m cells for frustum culling. pipeline/bake/landcover.py (vegrows_<tile>.geojson) → vegetation-layer.ts (the tree rows) and low-vegetation-layer.ts (the hedges, since 2026-11-01 drawn like the OSM hedges at the size their card states, 1.6 m × 1.2 m — they were dark 0.9 m boxes every 1.1 m), per fine terrain tile. Where DOM1 exists the canopy step places a row's trees itself (2026-10-21, P7, canopy.py row_trees): DOM1's crown peaks within 3 m of the line (nDOM smoothed 1 m, over 3 m, off LoD2 and the traffic), projected onto it, as tall as measured, thinned tallest-first 3 m apart, none within 4.5 m of a canopy or scan crown; written into canopy_<t>.geojson as row: 1 with a rows member, which the pack carries as ROWS_MEASURED (lib/city/point-pack.ts), and the viewer then draws only the hedges from vegrows. Without DOM1 the trees stand every 9 m along the line as before. Row trees: Dresden's 33408_5656 1 176 → 540 (doubles 26 % → 2 trees, standing on a crown 80 → 97 %), 33412_5656 927 → 274, Leipzig's 33316_5688 1 378 → 408, Unna 904 → 85 and 362 → 64, Hamburg's 32564_5934 157 → 37, Grimma 208 → 15 and 98 → 16, Meißen 67 → 7; heights now measured 7–20 m (were ~7). Why: the 9 m grid stood 25–61 % of its trees within 4.5 m of a measured crown (Unna 61 %, Dresden 35 %) and 9–41 % of them over open ground, at 5–9 m where DOM1 measures 9–20 m. A DLM hedge within 2 m of an OSM hedge is left to the OSM one (hedges.without_mapped, 2026-10-21; Dresden 0.43 of 4.48 km). Without a DLM only natural=tree_row is written here (Hamburg's 2.26 km of OSM hedges were drawn twice).

  • Canopy fill ◎detected =computed — nDOM = DOM1 − DGM1, one tree per ~7 m cell at the tallest pixel, as tall as measured and 0.58 × its height wide (CROWN_TO_HEIGHT, lib/city/tree-placement.ts: the median of Dresden's 50 123 surveyed trees with height and crown, the same in every 5 m class; until 2026-10-07 the generic crown was scaled uniformly by h / 5.8, 0.88 × h wide and 1.1 × h tall, so a park point stood under ~5 crowns). A crown starts at clamp(0.15 + 0.012 · H, 0.22, 0.6) · H (2026-10-21, P8, crownBase; the laser scans' crown base ratios: street trees 0.31–0.35, tall ones 0.39–0.50), at least 3 m over the ground where it reaches over a carriageway (c: 1 in the canopy file, canopy.py over_carriageway: one within 0.29 · H + 0.5 m of the trunk; the pack then 5 values a point, 20 B, else 3) — up to the share's top; before, every crown started at 0.28 · H. Over a road: Dresden 769 of 6 823 and 747 of 8 252 on two tiles, Leipzig 1 829 of 14 986, Hamburg 3 986 of 24 680, Unna 640 of 6 537 and 773 of 8 737, Grimma 971 of 21 855 and 616 of 15 302, Meißen 454 of 14 506. The trunk drops to the crown's base (vegetation-layer.ts crownFrame, fitTrunks); the coarse level's crowns carry the same base (The coarse level's trees below), and an asked canopy tree's solid starts there too (ask-items.ts). Gated off paths and water via the DLM class raster, off the carriageways and track beds (ways.traffic, below), and off everything within 10 m of a DLM bridge (a bridge's steel on a park bank is as tall as a crown: the Blaues Wunder's pylons came out as two trees; without a DLM the decks the rail step baked from OSM, bridge_<tile>.geojson — so rail runs before canopy). And off a building's roof (2026-10-20, roof_mask): inside an OSM building outline (shrunk by 1 m, its courtyards out) a texel no more than 2 m above the outline's roof — the upper quartile of its nDOM, so a stadium's pitch or a hall's skylight wells do not pull it down — is that roof, not a crown. (The DLM files a stadium's stands, a fair's halls and a park's glasshouses as sports or park ground; LoD2 lacks some of them, and their roofs became rows of trees: Dresden's stadium, the Leipzig and Dresden fair halls, Planten un Blomen's glasshouses in Hamburg, a tennis club's halls in Unna. 0.7–1.6 % of the canopy points, per city.) Without a laser scan (Hamburg, Berlin; 2026-10-20, green_mask) the canopy also takes the texels outside the vegetation mask that the DOP shows green (NDVI ≥ 0.3, the GLI where the DOP has no infrared) and DOM1 rough (a 3 × 3 spread above 0.5 m on most of the 5 × 5 around), off every LoD2 outline grown by 1 m — the same gates as the rest of the canopy after it. A Land with a scan leaves those trees to lowveg (canopyx, below), whose crowns the scan measures. (Hamburg had no tree in its courtyards and gardens: no scan step finds them there, and OSM maps a park, not a back garden. Its crown cover against DOM1's rough tall vegetation off the buildings went from 0.68 to 0.99 — 0.61 to 0.96 in the built-up class — at the same precision (0.67 → 0.66), 19 600 → 46 900 canopy trees; the summer DOP's crowns are green, and its harbour's cranes, piles and scrap are grey.) On transport land the veto is the traffic, not the parcel (2026-10-20, pipeline/bake/ways.py, verge_mask): no tree stands on a carriageway (the DLM's road centrelines at their surveyed width, the buffers the class raster burns; without a DLM OSM's roads at the class raster's widths) or a track bed (the baked rails by their track count and every OSM track above ground — sidings, yards, trams in their own bed — one track each, at the drawn ballast's widths: 4.5 m, 4 m more per further track), each grown by 0.5 m. Beyond them the road and railway land the DLM files to its fences is planted from the crowns DOM1 shows there — rough, 3–45 m, off every LoD2 outline, a blob of at least 15 m² after a 3 × 3 opening — with or without a scan, and not gated on the DOP's green: Saxony's is flown leaf-off, where 1–4 % of a crown's texels read green (Leipzig), against 19 % in Unna's summer flight. (The class gate vetoed the parcel: the station forecourts, embankments, verges, junction islands and the strips of wood beside a road the DLM files as road or railway, with their trees — on Leipzig's tiles 5.2 % of the crown cover DOM1 shows, 6.9 % of what the model left out; Dresden 1.6 % and 3.9 %. Leipzig gains 5 373 canopy trees (+11 %), Unna 840; on Unna's tile with a scan every added point on rail and road land is tall vegetation to the scan (multi-echo ≥ 0.5 within 1.5 m).) Nor is a bare plane a crown there (2026-10-21, bare_planes): a texel whose 5 × 5 neighbourhood is at least half plane — the residual of its 3 × 3 window about the least-squares plane at most 0.25 m (PLANE_M; a heap's flank reads 0.05–0.15 m, a Saxon crown over 1 m) — and whose calibrated index reads below 0.1 (BARE; asphalt, gravel, sand and roofs 0.0–0.05, a mown lawn 0.33) is no crown, on transport land only. (rough_mask read a gravel heap's smooth flank as rough for its slope alone, and leafless trees are no greener than gravel in Saxony's DOP: Leipzig's freight yard on 33318_5690 grew 22 trees on its heaps. Not in the forests and parks: there Dresden's DOM1 of bare limes and München's smooth DOM under a visible-band index read like planes too. Removed: Leipzig 26, Dresden 6 (on four tiles), Meißen 1, München 52 (cranes, containers, kiosks, a covered rail approach; one tree loses one cell), none in Grimma, Unna and Hamburg; 33 of the 37 Saxon removals were no tree.) pipeline/bake/canopy.py. At build time (scripts/prepare/side-bakes.ts, treesOffStructures in lib/city/small-buildings.ts) a canopy or laser-scan point in or within 0.5 m of one of the tile's small structures is dropped: DOM1 reads a shed's roof as a 3–4 m tree (859 canopy points stood in 636 sheds, 807 within 1 m of the shed's own height). Not in the canopy bake: it runs before the structures are found. Likewise a point more than 1 m inside a LoD2 footprint and at most 2 m taller than that building is its roof (treesOffRoofs, lib/city/roof-trees.ts): where the Basis-DLM still files a since-built hall's ground as park, DOM1 over its roof made a grove on it (2 153 points over Dresden's 15 tiles, most on the halls west of the Yenidze; 2026-10-19). A taller point is a tree over a shed or garage and stays. At build time, so every site has it without a re-bake.

  • NDVI crown colour =computed — per-tree DOP greenness (ndvi_<tile>.png, sampled on the CPU at placement) shifts the crown dry pale-sage → lush deep green. pipeline/bake/ndvi.py → vegetation-layer.ts crownColor; falls back to the hash-only sage when no NDVI raster (graceful — see portability). Sampled with a 5×5 footprint max + a remap recentred on the low NDVI median: the raster is ~2 m/px and median-zero, so a single-pixel sample left ~95 % of crowns reading "dry" (invisible); the footprint max + recentre make lush↔dry read clearly. The low median has a cause: the DOP was flown on 2024-03-19, leaf-off (data/dresden/provenance.json), so deciduous crowns are bare in the imagery and the index mostly separates evergreens and grass from everything else. A summer DOP would make the recentre less necessary and the meadow tint truer.

  • Vegetation index from the visible bands =computed (GLI; 2026-10-01, ADR 0012) — a DOP without an infrared band (Bavaria's open DOP is RGB only) gets the same ndvi_<tile>.png from its colours: the Green Leaf Index GLI = (2G − R − B) / (2G + R + B), mapped onto the NDVI's scale by GLI_SCALE 3.3 · GLI + GLI_OFFSET 0.14, clamped to 0..1 (ndvi.py _gli). The line was fitted where both exist — Unna's two summer DOP tiles, NRW's RGBI: r ≈ 0.7 against the NDVI, 85 % agreement on NDVI > 0.3; checked on Hamburg's four tiles, r 0.75–0.86. Every consumer (crown colour, meadow tint, the tram's lawn bed, the hedge cue) reads it as it reads the NDVI. It tells green from grey well and vigour less well — a stand-in, marked 🟡 in the Sources by city page. München's four tiles have it. A DOP with fewer than three bands still skips.

  • One vegetation-index scale for every flight =computed (2026-10-20) — the orthophotos are not calibrated reflectances: each flight's bands carry their own gain and offset, so the same asphalt read NDVI −0.30 in Dresden, −0.37 in Meißen, −0.15 in Leipzig, −0.07 in Hamburg and +0.02 in Unna, and the same mown lawn 0.33, 0.41, 0.12, 0.14 and 0.22 (Munich's GLI: 0.20 and 0.43). Every threshold downstream (the lawn edges' and the urban green's 0.3, the hedge cue's 0.12, the canopy's courtyard crowns, the sheds' and the rebuilt roofs' green veto, the tram's lawn bed, the crown colour and the meadow tint) was tuned on Dresden's: on Leipzig's flight 1–4 % of the OSM lawns passed the urban green's 0.3 (Dresden's 18–82 % a tile), so its courtyards drew grey; on Munich's GLI the asphalt read 0.19 and passed the hedge cue. ndvi.py now maps the raw index by one line per site onto Dresden's scale: the medians on the site's carriageways (ways.carriageways) and its mown lawns (OSM grass, meadow, village green, park, garden), both where DOM1 stands within 0.5 m of the ground, pooled over the site's tiles, go to Dresden's own (−0.298, 0.332), so Dresden's raster stays as it was and a threshold means the same green in every city. Two materials every city shows and that hardly change between flights, asphalt and a mown lawn, and per site rather than per tile: Dresden's lawns alone range 0.14–0.46 a tile under one flight (the Großer Garten against the Heide), which a per-tile scale would have turned into seams. A site with too few of either keeps the raw index. The step now runs before the canopy, which reads its raster.

  • Evergreen canopy crowns ≈ detected (2026-10-20) — a canopy crown has no species, so every one followed the generic deciduous year and the Dresdner Heide's pines stood bare from November to April. Two sources tell a conifer without a register. The Basis-DLM's forest kind (veg02_f VEG: 1100 broadleaf, 1200 needle, 1300 mixed; without a DLM OSM's leaf_type on a wood or forest): a crown in a needle forest is an evergreen (l: "e"). And in a mixed forest the orthophoto, where it was flown leaf-off: on Dresden's March flight the crowns of the Heide's needle stands read NDVI 0.39 (median, 3 × 3 max at the peak) and its broadleaf stands' 0.16 — a cut at 0.354 parts them best (Youden's J 0.40) — so a mixed forest's crown at 0.35 or more on the calibrated index is one too (l: "e", ls: "dop"), but only where the site's broadleaf crowns read below 0.30 (at least 200 of them): under leaves a summer flight shows every crown green. 44 % of the Heide's mixed-forest crowns pass. Baked: Dresden 91 076 of its 303 113 canopy crowns (the Heide's two northern tiles 63 % and 60 %), Grimma 807, Meißen 129, Unna 68; none on Leipzig's, Hamburg's and München's tiles. The DOM's form (a cone's peak over its shoulders) does not part them (J 0.17) and is not used. An evergreen crown keeps its leaves and colour all year (no season jitter) on both terrain levels (lib/city/coarse-crowns.ts byte 10), and the card says Nadelbaum, immergrün with the source that told it. It stands as a conifer (2026-10-21, app/_components/conifer-crown.ts): a three-tier lathe cone in the lobed crown's box, on the same trunk, its own chunk set beside the broad crowns, sharing their node build (the same attributes, no index, the same material), in the register conifers' deeper blue-green (evergreenColor; before, the NDVI's lush↔dry green), outlined as its cone when asked; the coarse level draws it as a six-sided cone of one skirt, and the register's conifers carry their cone there too (lib/city/coarse-crowns.ts, the code byte's 0x80). A canopy evergreen is a needle forest's crown or a mixed forest's green in March, so a pine, spruce or fir — an evergreen broadleaf (holly, box, cherry laurel) is a shrub under the canopy's 3 m — and a larch in a needle forest is drawn as one (the data tells no larch apart). Pines (2026-10-21, P12–P14, pipeline/bake/crown_form.py, canopy.py mark_pines): a crown whose surface a fifth of its height out stays within 0.3 H of its top is a pine's (pine: 1; on the registers' conifers DOM1 sees, 73 % of the pines and 29 % of the cones fall under it): it wears the pine's flat-topped crown from 0.55 H (buildPineGeo, lib/city/tree-inventory.ts PINE_CROWN_BASE, pineBase), the rest the cone — Dresden 54 600 of 91 076, Grimma 493 of 806, Meißen 77 of 129, Unna 32 of 68. Asked, a canopy pine's solid starts its crown at 0.55 H too, and only a tree drawn as a cone is outlined as one (ask-items.ts bit 16, CanopyInquiry.cone).

  • Crown shaping ≈assumed — radial crown normals (free), organic trunk, base darkening; every tree its own growth (treeGrowth, vertex stage, from where it stands): a lean of up to 0.7 m at the trunk's top (the hash squared, so most stand nearly upright) that carries the crown along, and a crook of up to ±0.12 m mid-way up the trunk, both in metres turned into each set's frame so a register tree's trunk and crown still meet; the trunk forks at 0.78 of its height into three limbs ending inside the crown (trunkLimbs, part of the shared trunk geometry; a cone or pine keeps an unforked trunk, and so do the register's own trunks, a set of their own beside the canopy's in a cell that has both), which a bare crown shows in winter; the crown moves as the trunk's top does, whole. Not away from walls: the trunk knows nothing of them; the mid crown is a detail-2 icosphere (180 tris) with lobes; three-tier crown LOD per 250 m chunk (lib/city/vegetation-lod.ts), planned over every loaded tile at once: the rich ~1 440-tri multi-tuft crown near (220 m in / 300 m out) but only while the site's rich trees fit a budget of 2 500, nearest chunks first; the mid crown + trunk; and past 650 m (back at 550 m) a detail-1 crown (80 tris) without trunk, dense chunks (≥ 400 trees) thinned to every other tree drawn 1.35× wider. Before the budget, a camera in the Dresdner Heide's forest tiles (one tree per 7 m) put ~9 000 rich crowns on screen — ~80 M triangles per pass over the site, which stalled the GPU into a lost context; the same view is now ~23 M. No backlight glow (see Backlit crown shimmer and translucency under rejected).

  • The coarse level's trees ≈assumed (2026-10-05) — each terrain level carries its own trees. The fine level draws every tree (the tiers above). The coarse level, which the tile renderer shows wherever the fine one is not loaded — in the distance and from the air, in a Modell picture past 2.5 m/px, at any scale on a phone whose memory governor coarsens the stream, and while fine tiles load — draws a fixed third of them, chosen by a hash of where each stands (drawnCoarse, lib/city/coarse-crowns.ts): as tall, their crowns √3 wider (COARSE_TREE_WIDEN) so a wood covers what the whole one did, in the same colours (NDVI, the register's foliage) and seasons, on the scene's crown material, the far crown's geometry one subdivision coarser and without trunks (coarse-crowns-layer.ts); its conifers — a canopy evergreen, a register tree whose crown is the cone — as a six-sided cone of one skirt in a set of their own (coarseConifer, 2026-10-21), and its pines with the pine's crown in a third set (coarseForm). A canopy or row crown's record carries its crown base in millimetres (the third uint16; 2026-10-21), so a distant crown starts where its near one does (crownBase above); a coarse crown had ignored it. The build step runs the fine level's own placement over each tile's tree files (the packed canopy as the browser unpacks it, the scan crowns, the rows, the register and the orchards, its veto, on the fine level's ground; lib/city/tree-placement.ts, lib/city/tree-inventory.ts, scripts/coarse-crowns.ts) and packs that third, 20 bytes a crown (crowns_<t>.crw.gz, ≈ 136 000 crowns and 1.5 MB gzipped for Dresden's fifteen tiles). The coarse level fetches it with its other side files and builds it with its dressing, whatever the mode or scale, so a tree stays where it stood when the level changes and no scale and no device is left without trees. A level that takes over from the other on screen shows only once its trees are built (handsOver, tile-stream.ts): the old level's trees stay until the new level's arrive, in the same frame — first built, the new level showed bare and its trees followed a moment later, a blink at every change of level. The cost of the switch: where the levels meet, a wood goes from all its trees to a third of them, wider. Since 2026-10-25 (plan 076, step 4) a coarse level's dressing — light: its crowns and bridges — goes ahead of every fine level's in the dressings' one-at-a-time turn (COARSE_DRESSING_RANK), so every tile in view has its third before the nearest has all; and where the fine level shows without its own trees — a phone shows it the moment it loads, and its dressing waited behind the nearer tiles' — the coarse level's crowns stand in on it until they hang (coarseCrownsStandIn, DressingPlugin.standInCrowns). They are the coarse level's own, kept loaded under the fine one anyway: no byte more on the GPU or the wire. A site file of every tile's crowns at boot (the plan's first idea) would have held all ≈ 136 000 crowns on a phone for the whole visit, ≈ 12 MB.

  • Canopy motion ≈assumed — per-frame in buildCrownMaterial, main pass only (the shadow pass draws the rigid castShadowPositionNode → no shadow-pass cost, no extra buffers): wind sway (vertex bend, stiff base → loose top, per-tree phase from the instance origin) and sway-coupled brightness (the crown brightens leaning into the same gust, × (1 + 0.09 · gust), centred so the mean colour is unchanged). The brightness is fixed at what its slider (Windhelligkeit) defaulted to; the leaf flutter is gone (🗃️ below).

  • Tree inventory from the Dresden street-tree cadastre ❝taken ≈assumed (every tile, by default) — inputs: the city's Stadtbaumkataster (WFS cls:L1261, dl-de/by-2-0 "Landeshauptstadt Dresden"; street trees, parks, schools — not the Großer Garten, not private ground): position, height, crown diameter, taxon. pipeline/bake/trees.py bakes data/<site>/dlm/trees_<tile>.geojson (from the WFS cache bun run fetch writes, cadastre.py) (h, d, archetype id, leaf type, foliage colour; missing h/d imputed from the genus median / the archetype's d:h), with the taxonomy in pipeline/bake/tree_archetypes.py: genus + cultivar + German name → eight archetypes (round 66 %, oval 15 %, small ornamental 12 % incl. 276 globe cultivars, columnar 5 %, conifer 1.8 %, weeping 0.1 % of the block's 18 444 trees; since 2026-10-21, P12–P14, the conifers split three ways: the cone — Picea, Abies, Thuja, Chamaecyparis, Larix, Pseudotsuga —, the pine (archetype 6: Pinus but cembra/peuce, a flat-topped crown from 0.55 H on a long clear stem) and the bush (7: Taxus, Cephalotaxus, the dwarf pines and spreading junipers, a dwarf cultivar: the lobed crown from 0.04 H); Dresden's 2 174 conifers → 805 cones, 883 pines, 480 bushes, Leipzig's 257 → 111 / 78 / 68, Hamburg's 117 → 76 / 39 / 2), leaf type (318 evergreen; Larix/Metasequoia/Taxodium are leaf-off) and 193 purple / 40 golden cultivars. What it does: tree-inventory-layer.ts plants each tree at its surveyed spot; the per-instance scale is non-uniform (crown width = d, crown depth = h minus an archetype clear stem), so round/oval/small share the lobed broadleaf crown and only four silhouettes get geometry of their own (a flame for fastigiate cultivars, a tiered lathe cone for conifers, a pine's flat-topped crown — the lobed crown reshaped, conifer-crown.ts reshapeCrown, pineSilhouette —, a curtained dome for weeping trees; a bush is the lobed crown from the ground) — same crown material, LOD swap and 250 m chunks as the canopy. Evergreens, purple and golden cultivars are tinted from the cadastre; deciduous crowns keep the NDVI remap. Row and canopy trees inside a cadastre crown (radius max(d/2, 3.5 m)) are dropped unless they overtop it by max(5 m, 30 %) (lib/city/tree-inventory.ts) — except around a cadastre tree in DLM forest/copse (classes 2/3; the bake flags it f = 1, 599 of 18 444 trees): there the measured canopy is kept whole, because the veto made parks and woods visibly thinner (the canopy draws every 7 m cell, the register only the trees it tends). Draw calls: the trunks and the broadleaf crowns (≈93 % of the trees) are handed to buildVegetation as precomputed TreeInstances and ride in the canopy's own chunk meshes (their trunks one unforked set beside the canopy's forked one); only the flame/cone/dome silhouettes are meshes of the inventory layer. Findings (scripts/eval/kataster-eval.py): only 31 % of cadastre trees have a canopy point (21 % on the spawn tile) — the canopy mask misses 99 % of the trees standing on road pixels and 67 % of those on built-up land; 7 743 of the missing trees are ≥ 8 m tall. 41 % of the DLM tree-row samples duplicate a cadastre tree. Where both exist, heights agree to a 2.8 m median absolute difference (canopy 1.0 m lower, r = 0.72). Leaf-off NDVI cannot assign a leaf type: AUC 0.87, but at the best balanced threshold (NDVI ≥ 0.37) precision for "evergreen" is 6.5 % (265 of 4 073 flagged), and it would flag 31 % of all canopy points; the best raw accuracy (98.3 %) is no better than calling everything leaf-off. Look (prototype, shots/kat-*): bare streets and squares (Radeberger Straße, Albertplatz, Stolpener Straße) become avenues, the fly-over reads as an inhabited city; dense parks (Rosengarten) get visibly thinner, because the canopy's uniform scale draws every 7 m cell as a 0.77·h-wide crown where the cadastre draws measured crowns. Cost (scripts/eval/kataster-cost.ts, block, main-pass vegetation draw calls after the LOD swap; canopy only → prototype as a separate layer → merged, as shipped): Albertstraße 222 → 581 → 401, Albertplatz 168 → 432 → 297, Rosengarten 140 → 385 → 259, fly-over 61 → 180 → 126 (the merge removes ~50 % of the added calls; what is left is the reshaped silhouettes, up to three per chunk). Built vegetation meshes 675 → 1 944 → 1 367; +179 KB gzipped transfer; the prototype measured −4 to −9 % fps on an M1 Max (scripts/eval/kataster-perf.ts) and no change in time-to-ready. Fallback: no trees_<tile>.geojson → rows + canopy, unchanged. Portability: any city's tree register (or segmented LiDAR trees) fills the same contract.

  • Tree registers of other cities ❝taken (2026-10-01; Hamburg and Leipzig baked, Berlin configured) — inputs: pipeline/bake/cadastre.py holds a register per city as data: its WFS (service, one or more feature types, output format, CRS) and a field mapping (Fields: botanical and German name, height, crown, trunk diameter or circumference — a circumference is divided by π —, planting year, felled date, position from two properties or from the geometry). trees.py reads every register through it into one record; no per-city branch. Registers: Hamburg's Straßenbaumkataster (de.hh.up:strassenbaumkataster, dl-de/by-2-0, BUKEA; street trees only; taxon from the cultivar, else species, else genus), Leipzig's Baumkataster (OpenData:Baeume, dl-de/by-2-0, Amt für Stadtgrün und Gewässer; street and park trees; felled trees and the "Bestandsfläche"/"waldartiger Bestand" points — a stand, not a tree — skipped), Berlin's Baumbestand (baumbestand:strassenbaeume + baumbestand:anlagenbaeume, dl-de/zero-2-0; its root CA is only in certifi's bundle, which the fetch adds to the system trust — verification stays on). The taxon table gained Calocedrus (conifer), Eriolobus and Crataemespilus (small) and the registers' placeholders ("unbekannt", "nicht bestimmt") as no taxon: trees of no genus the classifier knows are 2.3 % in Hamburg (all of them records without any name), 0.3 % in Leipzig, 0.1 % in Berlin's spawn tile (2.4 % in Dresden). Hamburg records no height. Since 2026-10-21 (P10) a register tree without a height takes it from its genus and trunk diameter (pipeline/bake/allometry.py, the table tree_allometry.json: per genus the median height of each trunk class, made non-decreasing, fitted on Dresden's and Leipzig's registers, ~74 000 trees; a genus with few trees its archetype's curve; cross-validated, 1.8 m median absolute error against 3–4 m for a genus median), else from its crown, else its genus's median (_missing_height). Hamburg's DOM1 is matched from a leaf-off flight and read its street trees 10–40 % low (a 25–40 cm lime 9.7 m against 13 m in Saxony); it now only drops absent trees (75) and raises stale heights (32; Register heights checked against the surface model below). Hamburg's median register height 8.3 → 12.9 m (25–40 cm limes 9.7 → 14.1 m). (Until then the bake measured the height in the surface model — the highest texel within a quarter of the crown diameter, 1–3 m, of the trunk, kept when it fit the recorded crown, 0.6·d ≤ h ≤ max(3·d, 6 m) + 2 m —, else imputed it from the crown; on Leipzig's register that read 0.6 m low in the median and 1.5 m off, but Hamburg's November flight is not Saxony's, and 8 575 of its 10 418 trees had taken that low height.) Baked (before P10): Hamburg 10 418 register

    • 3 224 OSM trees (1 838 / 3 330 / 2 139 / 3 111 + 694 / 875 / 757 / 898 on 32564_5932, 32564_5934, 32566_5932, 32566_5934), Leipzig 25 498 + 2 826 (5 452 / 5 424 / 7 237 / 7 385 + 794 / 735 / 612 / 685 on 33316_5688, 33316_5690, 33318_5688, 33318_5690); Leipzig's scan trees (canopyx) were thinned against the new register (8 814 of 27 893 dropped as register trees). Dresden's output is unchanged by the generalisation (re-bake of 33412_5656: the same 6 121 trees, property for property).
  • Register heights checked against the surface model ❝taken ◎detected (2026-10-21, P11; trees.py surface_check) — the surface model has a say over every register tree, and over an OSM tree with a height. Stale: where DOM1 by the trunk (the highest texel within a quarter of the crown diameter, 1–3 m) stands over the recorded height by max(5 m, 30 %), fits the recorded crown (plausible_measure) and is at most 1.9 × the allometry's height for the trunk on record (99 % of Dresden's and Leipzig's measured trees are; a 5 cm sapling read 9 m under a neighbour's crown), DOM1's height is taken (the crown filled in from the height again; the card's measured bit set): raised in the registers Dresden 1 920, Leipzig 705, Hamburg 32; OSM trees Dresden 555, Leipzig 638, Hamburg 718. Absent: a tree the model must see (8 m or more expected) where DOM1 by its trunk stays under 2 m, the disc all in view, the orthophoto bare there and the register not dating its planting within 6 years of the flight is left out: Dresden 51, Leipzig 8, Hamburg 75 (DOM1 alone does not tell: Hamburg's leaf-off model shows nothing at 4 % of its 25 cm trunks, most of them green in its summer photo). A conifer of no known genus takes its form from DOM1 (crown_form.py: pine or cone). OSM trees stand clear of a register tree by max(3 m, its crown radius ≤ 15 m) (were 3 m): the doubles 3–6 m from their register twin fell from 1 364 / 609 / 1 288 to 440 / 278 / 227 in Dresden / Leipzig / Hamburg (OSM trees dropped: 274, 372, 60). The registers' weeping trees: Hamburg 51 → 52.

  • OSM trees beside the cadastre, genus and trunk ❝taken (plan 025 phase A, every tile) — inputs: the site's OSM extract (natural=tree nodes, ODbL) and the cadastre's stammdurchmesser_akt. trees.py writes each tree's genus gn (an index into tree_archetypes.GENERA, 38 entries, the file's genera member; the red maples and red oaks are entries of their own because their autumn is) — 17 228 of the 18 444 cadastre trees (93 %) have one, the rest follow the generic curve — and the trunk diameter t (cm, 17 233 trees; a diameter implausible for the tree's height — over 15 cm per metre, or over 4 m, 0.1 % of the register — is dropped, not clamped). The viewer fits a measured trunk's girth to the radius the trunk geometry draws at 1.3 m, its flared foot included (× 1.3 for the nine-sided, barkless trunk; tree-inventory.ts trunkGirth / trunkRadiusAt; a taper-only fit drew a 25 m tree's trunk half as thick again), the height rule stays where none is measured. OSM trees: the BBBike extract (2026-09-19) has 10 854 natural=tree in the fifteen tiles (6 018 in the four first); 8 292 (4 385) carry a taxon that names a genus the classifier knows (species/taxon/genus, then species:de/genus:de; German names such as "Platane" or "Gemeine Fichte" mapped by their last word, a lower-case genus capitalised, an unambiguous bare epithet such as hippocastanum read — a family or an English name is no taxon) or at least leaf_type (which wins over a taxon it contradicts; leaf_cycle sets the leaf type), the rest are dropped (the DOM canopy covers unknown trees; no species is invented). The cadastre wins: an OSM tree within 3 m of a cadastre tree — any of the cached WFS answer, its 10 m margin across the seams included — is taken to be it (4 839 dropped; 2 565 on the four first), leaving 3 453 over the fifteen tiles (0–854 per tile; 1 820 on the four first, 854 / 252 / 368 / 346; s: "osm", 298 with a genus) — courts, the Zwinger, Free-State and private ground. Their height/crown/circumference tags are read (53 / 1 / 2 in the four first tiles), the gaps filled from the cadastre's own statistics for the tile. They join the cadastre's veto path (keepTree), so an OSM tree replaces the canopy or laser-scan crown it stands in. The re-bake left the cadastre trees unchanged (same set, positions, properties); features are now sorted by position so a re-bake diffs by content. Since 2026-10-21 (P11) the clearance is the register tree's crown radius, at least 3 m (Register heights checked against the surface model above). A site without a register keeps its OSM trees (2026-10-21): the trees step runs there too and writes trees_<t>.geojson and treefacts_<t>.json on every tile, the OSM trees alone, their heights measured in DOM1 (measure_heights), vetoing the canopy crowns at their place as a register's do — München 3 271 (3 055 of them measured), Unna 191, Grimma 100, Meißen 92. Before, the step skipped a site without a register.

  • Trees by season ≈assumed (plan 025 phases B–C, every tile; look not yet judged on a GPU) — inputs: the genus above and the scene date. lib/city/tree-season.ts maps (day of year, genus, per-tree jitter of ±6 days) → { leaf, autumn }: leaf-out mid-April → early May, colouring in October, leaf fall late October → late November, per genus (limes early and butter yellow, horse-chestnuts browned from August by the leaf miner, oaks late, russet and holding 30 % dead leaves through winter, the ginkgo bare within days, larch and dawn redwood rust then bare); evergreens never change; canopy and row trees (species unknown) follow a generic deciduous curve. Another city's year is Dresden's warped by its climate (2026-10-21, P17, SITE_SEASON, climateDay): per site a spring and an autumn shift, the median over the species the Deutscher Wetterdienst observes of the difference between the city's and Dresden's median day of each phase (leaf-out and flowering; colouring and leaf fall), from DWD's annual phenology reporters 2010–2023 within 30 km and 150 m of height of the city (dl-de/by-2-0, "Quelle: Deutscher Wetterdienst"; read once offline, the shifts are a table in the code — the browser asks DWD nothing): Hamburg +3 / +5 days, München +2 / +4, Unna −6 / +1, Leipzig −1 / +1, Meißen +1 / +1, Grimma 0 / −1, Berlin 0 / −3; the shift holds from March to June and from August through November and runs over between. The birch turns and drops two weeks later than before (DWD around Dresden: colouring on day 284, half its leaves gone on 304; the table had it 16 and 20 days early). crown-season.ts writes, on a change of calendar day only (the local calendar the HUD composes the date in; never per frame; throttled to one run per 150 ms), each crown's colour (summer green → the genus hue, divided by the crown material's base so it lands as itself) into the sets' per-instance tint (instanceTints), and aBare = 1 − leaf into a per-chunk instanced attribute (read with instanceFloat). A chunk with any bare crown switches to the seasonal crown material: a hashed alpha test in the crown's own local space (Wyman & McGuire's method, as three's alphaHash: cells about 1.25 px wide at every distance, fixed to the tree so they sway with it) discards the leafless part down to a 25 % grey-brown twig stipple; the test is the material's maskNode, which three's shadow pass honours, so the shadow thins the same way (PCF softens the stipple into a lighter shadow; until plan 020 a matching custom depth material did this). The first cut used fixed cells in crown space (~8 across a crown) with a solid, twig-tinted crown past 160–320 m: headless plates showed big trees shattered into flat brown shards up close, so it was replaced before shipping. A chunk in full leaf keeps the plain material (no discard, early depth test intact). A season change redraws the shadow map. Both crown materials are scene-wide (sceneCrowns) and are compiled once per scene before the first tree lands (crownWarmup, stand-ins through PostStack.compile), so the first date change across the leaf fall builds nothing inside a frame. Cost (scripts/eval/season-cost.ts, the whole site, 58 988 crowns, CPU): 4–7 ms median per date change (July → October 6.4 ms, October → January 3.6 ms), under the plan's 16 ms bar; the upload is 16 B per crown. Not seasonal: hedges and trunks. Unverified: the autumn hues and whether the bare-crown dither reads as noise in motion (the plan's STOP condition) need the --headed plates on a real GPU.

  • Tree families ≈assumed (2026-10-07, every register tree) — inputs: the genus gn (and the archetype's leaf for a conifer without one). lib/city/tree-family.ts groups the 37 genera into ten families an illustrator tells apart (lime, maple, oak/beech, plane, chestnut, pinnate — robinia, ash, honey locust —, birch/poplar/willow, blossom — cherry, apple, pear, hawthorn, rowan —, hornbeam/elm, conifer) and gives each a lean of its summer green (HSL offsets ≤ 0.07 on the NDVI colour: a chestnut deeper, a robinia yellower, a plane lighter; not on the copper, golden or evergreen cultivars) and a bark colour, written as the trunk's per-instance tint (a birch's white stem, a plane's pale bark, a pine's reddish one; canopy trunks keep the material's own). Blossom: Phenology.bloom gives eight genera a flowering span, colour and strength (cherries pink 2–26 April, before their leaves; pears, apples, hawthorn white; rowan, horse-chestnut candles, robinia and catalpa a partial cream), bloomAt a rise and fall over it; crown-season.ts mixes the crown's colour toward the blossom and keeps a flowering crown full (aBare = 1 − max(leaf, bloom)) — on the same day change, no new attribute. An illustrator's table, not botany.

  • Tree age ❝taken ≈assumed (2026-10-07, Dresden; the other registers on their next bake) — inputs: the planting year y (trees.py planting_year: a register's pflanzjahr, else the year of its record less Dresden's jalter; 43 471 of Dresden's 60 266 trees). lib/city/tree-age.ts turns the age into what the measured sizes do not say: a tree under 15 years has, without a measured trunk, a trunk down to 0.6 × the height rule's; under 5 years (1 960 trees) it stands between stakes (tree-stakes.ts: three posts and two rails, one set per tile, a scene-wide material); past 80 years an unmeasured trunk thickens to 1.3 × by 150. The crown's colour leans with the age too: a sapling's lighter and fresher (fading out by 15 years), an ancient tree's deeper and duller (from 80 to 150 years); on the fine level only, the coarse level's crowns carry no age. The age is counted to the current year, so the picture ages with the calendar. Height and crown size stay as measured.

  • Trees by picture style ≈assumed (2026-10-07) — what the tree changes above mean per style, with one new mark. Pastell, Film noir, Sin City: everything (families, blossom, age, stakes). Comic: its cloud crowns copy every per-instance attribute (style-dressing.ts), so they turn and flower with the season, printed in marker inks (green, an autumn crown yellow, a bare or blossoming one the paper; printedCrown), a bare one whole. Papier and Strich: the paper override has no mask, so their card crowns stay closed outlines (a plan's cloud, not a noise). Schwarzplan: until now no trees; now the register's trees as a site plan draws them (tree-plan.ts: the measured crown as a circle, its line 8 % of the radius, the stem a dot, 0.4 m above the foot), one hidden set per tile shown only in its frames (paper-scene.ts shows what is tagged planOnly) and drawn in the figure's black by the swap. Only the register's trees: measured, so they may stand on a plan; the canopy's are inferred. On the fine terrain level only, so a Schwarzplan zoomed out past it shows none.

  • Tree feet ≈assumed (Baumfuß, 2026-11-02) — every trunk grows out of the ground rather than standing on it:

    • the bark's own foot (buildTrunkMaterial, every trunk, every tier): three or four roots (a hash of where the tree stands) bulge the foot out by up to 55 % of its radius, fading out 0.6 m up the bark (trunkRoots, two extra rings low in the trunk, TRUNK_RINGS, nine sides); up to 0.45 m over the ground the bark turns to damp earth along them, the troughs about half as high, darkest at the ground. A separate root-flare skirt round the trunk read as a shoe and was dropped; the flat disc before it floated on a slope and let the paving show through. The earth is the ground's own where the tile has foot rasters (trunkEarthGraph, read per vertex at the foot, a trunk material per tile on one build per raster layout): the pit's mulch in paving, leaf mould in a wood, the grass's roots on a lawn or meadow, elsewhere the class colour darkened as damp soil (× 0.72); one fixed earth elsewhere.
    • bark and moss (barkSkin, every trunk): 17 furrows round the stem, wavering up it, tilt the normal across them (± 0.2) and shade their troughs to × 0.91, faded to their mean between 10 and 35 m (they would shimmer); moss (sRGB 118, 138, 80) on the weather side (west, a little north), highest low on the trunk and gone by 1.9 m, in patches, and a little all round at the foot, at most 50 %.
    • the trunk in its crown's shade (crownShade, every trunk): the trunk takes no shadow (in the shadow map the bark shadows itself, a black sleeve), so it stood lit under a dark crown and read as pasted on. Per vertex, the ray from the bark to the sun is tested against the sphere the generic crown is drawn round (in the matrix's frame, the sun turned in by its columns); where it passes through, the bark darkens to half, softly to the rim, by day; and under the crown the hidden sky darkens it toward the crown's base (× 0.8). A bare season keeps 30 % of it (crownLeaf, the generic phenology's leaf; a cone's or pine's trunk keeps its crown's shade the year round).
    • the crown darker inside and underneath (crownOcclusion, every crown, every form): a factor on the crown's colour per vertex, cool and dark toward its underside (the radial normal's height, down to × (0.56, 0.62, 0.66)) and in the troughs between its lobes (the shaped vertex nearer the crown's centre than 1.15 R, down to × 0.8), so the crown's mass reads as one volume over the trunk instead of a lit ball. Not while Comic prints the crowns.
    • in paving — the class raster's built-up, road, railway or unclassified land, off the baked lawn (the edge raster's meadow side) — a tree pit: mulched earth with a grain, litter flecks and the odd tuft of green, 0.7 m beyond the trunk's flared foot (0.85–1.4 m), its rim wobbling ±6 % by a world-space noise, in a 12 cm ring of setts (a joint every 14 cm and on both edges). Rings are drawn before every pit's earth, so overlapping pits become one bed with one ring; a pit stops 0.32 m short of the road edge (the edge raster's road distance), off the kerb. Alpha-blended and lit as the ground is (its sun, shadows, sky view and horizon).
    • under the crown (off forest and copse, whose floor is its own; the crown's reach 0.8 × the trunk's drawn length, 1.5–7 m): a multiplied glaze darkens and cools the ground in the crown's shade, ragged at the edge, with moss on a lawn (a bare broadleaf crown keeps 30 % of the shade, crownLeaf; a cone's or pine's all of it the year round; fogged as the ground under it is, faded to white); where the crown's shadow falls (thrown off along the sun by the tree's height, at most three crowns away, drawn out by a low sun) and the shadow map shades the ground, sun flecks lift it back up to about the sun, by day, within 45 m (the glaze catches the shadow term through its receivedShadowNode and draws its factor as its outputNode). A fleck is the sun's own image through a gap in the leaves, a pinhole camera: round, the gap's height × the sun's 0.53° wide (6–23 cm), soft-rimmed, drawn out along the sun by 1 / its elevation (up to 3×), in clusters where the crown is thin, in two grids of cells turned against each other (worked out only where one can show: the quad reaches far past the shadow at a low sun). The wind does not carry them along: a branch sways a cluster 5 cm to and fro (about 7 s), a leaf trembles a fleck by 1 cm at 0.6–1.8 Hz and blinks almost half of them on and off; past a few pixels they are their mean. (Until 2026-11-03 they were noise cells drifting on the wind clock, which read as flowing water.) The glaze's quad lies along the throw, back past the shade under the crown and forward past the thrown ellipse; a quad centred between the two cut the shade off in a straight line; on a lawn or meadow a ring of unmown grass round the trunk (the mower keeps off it), reaching 0.85 m past the bark with a ragged edge: long blades streaked out from the trunk, darker and yellower, and seed heads catching the light where they are larger than a pixel, and in it a tuft of 88 real blades (tuftGraph, a set of its own on the trunks' buffers: thin triangles 0.08–0.24 m high, both faces lit from above, 0.8–1.3 × by a hash, leaning out from 0.28–0.8 m round a 0.25 m foot moved out to the trunk's own roots, turned with the trunk), opaque, lit as the ground, swaying 3.5 cm, shrinking into the ground as the foot fades, so the cut between bark and lawn is hidden — not in paving or a wood; and in autumn fallen leaves (none under a cone or a pine) — covering flecks in yellow, orange, russet and brown in 11 cm cells, thick under the crown, thinning 25 % beyond it and swept against the kerb (the edge raster's road distance), a mean tint where a cell is under a few pixels. How much leaf lies follows the scene's day (leafLitterAt: from late September, all of it in November, a third left into January, none from mid-March).
    • worn lawn under every other big park tree (on a lawn or meadow, a crown of 2.5 m or more, 55 % of them by a hash of where they stand): bare earth where people sit, on one side of the trunk, and the path they wear to it, 2.6–6 m long, tapering and wandering, grass showing through in tufts — covering, lit as the ground (the earth set, never with a pit). A path leads off in a hashed direction, not to a neighbour: the trees know nothing of each other.
    • the city's details round a street tree in paving: its roots (the trunk material's three to five lobes, same hash) break the mulch as a bark-brown root, heave the ring's setts over each crest and open their joints, and run on under the paving as thin cracks up to 1.5 m past the ring; a pit at least 5 m off any road (a pedestrian zone, by the edge raster) wears a cast-iron grate (rings and spokes, the mulch showing through); a pit 0.9–3 m off the kerb gets, for 70 % of them, an anthracite guard hoop between it and the road, 0.45 m short of the kerb and along it (the road distance's gradient is the way to the road; none on a tile without the edge raster, nor where the trunk's foot and its roots' bulge, 1.55 × its radius, reach past it) — a small mesh, opaque, standard-lit, not a shadow caster, its posts on 0.2 m under the ground. Every part lies on the ground's slope at its trunk (8 × 8 cells; the slope read off the terrain 1.5 m either side of the foot when the chunk is built, two signed bytes a tree, groundSlope); where the ground bends away from that plane each vertex is drawn toward the eye along its own ray by 8 % of its distance from the trunk (at most 1 m), so it keeps its place on the screen and stays over the ground. The guard hoop stands on the same slope. (Until 2026-11-03 a flat quad sank into rising ground and the leaves and the shade stopped in a straight line where the slope's crease crossed it.) What a tree gets is read once per vertex at the trunk's foot. All of it fades out between 70 and 150 m; a chunk whose nearest tree is past 160 m draws none of it (the chunk's crowns, trunks and the register's own trunks count; a cell of the register's trunks alone, whose crowns are drawn elsewhere, is switched by its trunks the same way). One Instances set per part (ring, earth, leaves, shade, guard, tuft) per set of trunks on the trunks' own attribute objects (matrices, drop), and a quad, a hoop and a tuft per tile, with materials per tile whose slots hold the tile's class and edge rasters and its light — two bytes a tree (the slope) and about 15 KB a tile. The paper styles leave the blended parts out. Before, the trunk stabbed straight into a flat ground.
  • Hedges (OSM, laser-scan height) ◎detected ❝taken and trees outside the canopy mask (laser scan) — on by default. The laser scan is baked for every tile of the site (bun run fetch <site> --lsc, then bun run bake <site>); a tile without one would bake OSM-only (hedges at their tag / 1.5 m, no extra trees). Where a scan is read (products.lsc): Saxony (Dresden, and since 2026-09-30 Meißen, Grimma and Leipzig — Meißen 6 449 extra trees), NRW (Unna, the same date) and, since 2026-10-01, Bavaria (München, see below); Hamburg and Berlin stay OSM-only. Shipped: the OSM hedges (src osm, osm+lsc, osm+dom), the extra trees and, since 2026-10-21, the scan's own hedges that pass scan_hedge_ok (lsc): at least 6 m long, at least 4× as long as wide, 80 % of the axis at least 1.5 m from any crown over 3 m, median NDVI at least 0.12 where the raster exists. München 123 of 453 (1.42 of 3.12 km); Unna 961 of 2 990 (11.68 of 21.31 km; their precision is not measured yet). Without a scan an OSM hedge takes DOM1's height (the median along the line of the highest nDOM within 1 m, ignoring anything over 3 m, clamped to 0.5–3 m) and is dropped under 0.4 m. Hamburg: median 1.4–1.8 m instead of the 1.5 m default, 7 of 53 dropped. Not shipped (🗃️ below): the rest of the laser-scan-only hedges (short, wide, along crown rims, not green) and all shrubs — the bake still finds them (bun run bake <site> --step lowveg --research writes every candidate under data/_raw/ for research), but the committed lowveg_<tile>.geojson holds only what renders. Drawn (low-vegetation-layer.ts, 2026-11-01): a trimmed hedge — tapered flanks, rounded shoulders, a flattish top, square-cut ends — one instance per ≤ 2.5 m piece, the pieces meeting end to end, mitred at a bend (a closed ring at its closing corner too) and sheared to the ground's slope, so a run reads as one hedge, not a row of sausages. Leafy lumps (~0.8 m, ±9 cm) and a slow rise and fall of the top (~6 m, ±14 cm) are pushed out in the vertex stage by a world-space noise, so pieces agree at every joint; the same noise lights the lumps' crests and shades their hollows. The sage drifts over ~40 m in world space, never per piece or per line, so a tile seam does not show. In Comic the hedge wears a dimmed leafy marker ink and prints in the deep green. Why: the old chain of soft blocks pinched at every join and the DLM's hedges were boxes (Manuel, 2026-11-01).

    • Why: the canopy bake (canopy.py) keeps nothing under MIN_H = 3 m and only forest/copse/sport-and-leisure areas, so courtyard and garden trees are dropped at any height; the Basis-DLM carries a hedge only when it is landscape-shaping (≥ 200 m) — vegrows on 33412_5656 has 0 hedges.
    • Inputs: GeoSN laser scan (LAZ, 2024-11-30, leaf-off) → 0.5 m rasters: ground (classes 2/8/30), surface max (2/20), non-ground count, its multi-echo share, and the mean intensity of the low returns (0.25–4 m above ground). DOP NDVI (2024-03-19). OSM barrier=hedge, natural=scrub|shrubbery|shrub. Exclusions: LoD2 surfaces (+1 m), OSM walls (+0.75 m), bridges, water/rail classes, the rim (1 m) of any > 3 m crown.
    • Cue — measured, and not the one expected: per-pixel AUC of OSM-hedge pixels against cars (road class) / OSM fences / building rims: NDVI 0.94 / 0.71 / 0.89, low-return intensity 0.70 / 0.83 / 0.80, multi-echo ratio 0.76 / 0.28 / 0.42. The echo ratio separates tall trees from roofs perfectly (≥ 0.5 on 100 % of > 5 m forest pixels, 3 % of roofs) but not low vegetation: only 26 % of hedge pixels reach it, against 56 % of fence pixels — a clipped hedge rarely splits a pulse, and 1–2 m above ground the two echoes are too close to separate. The leaf-off NDVI works because the city's hedges are largely evergreen. Rule: NDVI ≥ 0.12 ∨ (intensity ≥ 1250 ∧ echo ≥ 0.3); inside OSM scrub NDVI ≥ 0.06 ∨ echo ≥ 0.3. Then close 3×3, open 2×2, blobs ≥ 2 m². Other scans: intensity normalised per scan (2026-10-01). Two sensors' intensities are not calibrated against each other, so lsc.rasterise scales every scan's low-return intensities by 1352 / the median intensity of its own measured ground (class 2) in the tile — 1352 being that median on the spawn tile 33412_5656 where the 1250 was measured (34.1 M ground points, p10/p90 958/1637), so that tile keeps a factor of exactly 1 and its rasters are byte for byte what they were. Bavaria (LDBV, CC BY 4.0, flight 2022): ground medians 1436–1548 per 1 km file, factors ≈ 0.88–0.94. NRW ("3D-Messdaten", Geobasis NRW, dl-de/zero-2-0) records 16-bit intensities; it had been divided by a hand-set 16 (a ratio of whole- scan medians, not of like surfaces) — still applied while merging, now moot: a fixed factor does not change the median ratio. Assumption: the same mix of ground (asphalt, lawn) reads equally bright to every sensor — a first-order gain correction, not a calibration. Not verified on Unna's or München's hedges by eye. On Unna's 32408_5710 the NRW ground median is 3152 after the old 1/16 (raw 48–52 k per 1 km file), factor 0.43: the old scale had left its intensities 2.3× too bright, yet the shipped result is the same — the 11 OSM hedges (4 with the scan's height) identical, the low-vegetation mask 10.39 → 10.29 ha, the scan trees (echo only) unchanged at 1 258 — because Unna's mask leans on NDVI. So Unna was not re-baked for it. Bavaria's classes are not AdV's: 6 building, 20 "object point, e.g. vegetation" (no 3/4/5), 9 water, 23/24 synthetic ground, 22 bridge (≤ 2020) or cellar entrance (≥ 2021). providers/by.py maps them into the AdV scheme while merging (6 and 20 → 20, 9 → 8, 23/24 → 30, 22 → 20 up to 2020 and dropped after, 0/1/7 dropped): buildings stay in the non-ground class as in GeoSN's, because every rule here was measured with roofs in it and the small structures need the sheds' roofs; LoD2, land cover, NDVI, echo ratio and intensity decide what is vegetation. München (2026-10-01, the four tiles): 108 of 195 OSM hedges take the scan's height (osm+lsc); 6 104 / 8 068 / 7 708 / 10 655 extra trees (no cadastre to thin them; 97 % of the Altstadt tile's on the built-up class, median 13.3 m). Intensity factors per tile 0.92 / 0.90 / 0.89 / 0.86. NRW's classes are not AdV's either (measured 2026-10-01 on 32408_5710, 66.7 M points): class 1 is 22 % of the scan, every point of it a multi-echo return 1.7–22 m above ground (the crowns' non-last echoes), while 20 is 18 % multi-echo (roofs, the last echoes); 17 bridge, 18 noise, 24 below ground, 26 synthetic ground. Read as GeoSN's (20 only) Unna's surface missed most crown tops; providers/nw.py now maps 1/17/20 → 20, 26 → 30 and drops 18/24 (LSC_CLASSES), and Unna's re-bake found 14 457 scan trees (was 2 385) and 897 small structures (was 1 596: crown gaps had read as low roofs). München's open DOP is RGB, so its NDVI branch reads the RGB stand-in (ndvi.py) rather than a near-infrared NDVI, and its flight is leaf-on (June 2022): the cue's thresholds are Dresden's, untested against either.
    • Shape: elongated components (skeleton ≥ 4 m, length/width ≥ 3, width ≤ 3 m from the distance transform on the skeleton, few spurs) → skeleton → polyline → Douglas-Peucker 0.4 m, h = median ridge nDOM, w = 2 × median inscribed radius; compact ones → one shrub (centroid, equivalent radius, p90 height), beds > 12 m² → shrubs at height peaks ≥ 1.5 m apart. OSM geometry wins: a mapped hedge keeps its line and takes the LSC height where ≥ 30 % of it is supported (src: "osm+lsc"), else its height tag or 1.5 m ("osm"); LSC hedges within 2 m of it are dropped, the rest fill the unmapped ones ("lsc").
    • Evaluation (33412_5656): only 47 % of the 4.3 km of OSM hedges is observable at all — 53 % runs under a > 3 m crown, invisible to a first- surface model. Of the observable length, the mask comes within 1.5 m of 58 % (echo ≥ 0.5 alone: 25 %, NDVI alone: 48 %, no cue at all: 76 % but 9.7 ha of mask instead of 2.6 ha). Fences: 3 % of 17.9 km of OSM fence (> 3 m from a hedge) is hit — the 2×2 opening removes them. Mask breakdown: 30 % lies 1–2 m from a > 3 m crown (understory or crown-edge false positives — the main remaining risk), 1.9 % on the road class (cars), 1.8 % 1–2 m from a building, 2 % beside a fence. Checked by eye on DSM-hillshade overlays: a clipped evergreen hedge beside a row of parked cars is taken and the cars are not; grave shrubs on the Trinitatisfriedhof come out as shrubs; a branchy scrub mass was first mis-skeletonised into a "hedge network" (fixed by the inscribed-width + spur test). Candidates: 594 hedges (65 OSM, 51 OSM+LSC, 478 LSC-only; 7.3 km) and 3 226 shrubs. Shipped: the 116 OSM hedges (4.3 km; 54 / 94 / 45 on the OSM-only neighbours).
    • Trees outside the mask (canopyx): crown peaks of the multi-echo (≥ 0.5)

      3 m canopy, ≥ 3 m apart, more than 5 m from any current canopy point, off paths, water, carriageways and track beds (ways.traffic; until 2026-10-20 off the whole road and railway class, see Canopy fill) — 8 006 peaks (96 % built-up class; 57 % more than 15 m from the DLM road area, 1 969 of them enclosed by buildings on ≥ 6 of 8 rays = courtyards; 16 % street-side). Deduplicated against the cadastre in the bake: the cadastre wins position and species, and a peak within max(4 m, the cadastre crown radius) of a cadastre tree is dropped (radius match, not 1:1 — one big crown often yields two peaks) → 5 739 shipped, 2 267 dropped in the first bake; 5 788 shipped, 2 528 dropped (8 316 peaks) since the re-bake of 2026-09-25 with the current bridges and the land cover with the OSM squares carved in, where more trees may stand. (Taking the scan height where the cadastre has none is not done: the cadastre bake imputes missing heights and does not mark them.) Each peak's crown radius r is its watershed share of the crown mass (crown_radii, from every peak and every canopy and cadastre tree, so a neighbour's share stays its own), as a disc of the same area, 1–12 m. (Until 2026-10-20 the distance to the mass's edge at the peak: lone scan crowns came out a third of the area DOM1 shows them, against half for lone cadastre trees with a surveyed crown — Dresden, Leipzig; in Unna the median radius went from 1.8 to 3.5 m, and the circles over the orthophoto now cover the crowns they stand in. Saxony's scan trees keep the old radii until GeoSN's laser-scan share answers again: on 2026-10-20 it answered 404 for every tile.) Rendered as ordinary canopy trees.

    • Rendering (low-vegetation-layer.ts, separate from the tree layer): hedges are chains of superellipsoid "clay" blocks (288 tris, ≤ 2.5 m pieces, 0.6 m overlap); rooted-base darkening + a static world-space foliage mottle; 250 m chunks; cast and receive shadows, no animation (ADR 0009).
    • Cost (prototype, real GPU, 3200×2000, full block, all candidates): hedges + shrubs +11–13 draw calls/frame, frame time 19.8 → 20.2 ms on the fly-over; the 8 006 extra trees +4.9–7 % triangles drawn, 19.8 → 23.0 ms. Shipped (OSM hedges + 5 739 trees), kataster-cost.ts: +9–58 draw calls in the main pass over the merged cadastre (fly-over 126 → 144, Albertstraße 401 → 459).
    • Bake: pipeline/bake/lowveg.py (scipy, scikit-image in the uv environment; lsc.py rasterises the LAZ with laspy by PDAL's binning rules — the committed rasters were PDAL's). The committed files were baked by its predecessor (extract-lowveg.sh, OSM via Overpass); the port reads the hedges from the local extract (ADR 0037).
  • Street lamps ❝taken — OSM lamp points (the local Geofabrik extract, only those the tile owns: west and south edges in, east and north out, so a seam lamp stands once) → instanced lamp posts (ODbL). pipeline/bake/lamps.py; the viewer applies the same ownership to older files (ownsPoint). Gated off water (8) and railway (5) land-cover so no poles stand in the Elbe or the track bed (the rail corridor is now its own layer). Built per fine terrain tile; the three real lights (a phone's one, 2026-10-11) go to the nearest heads of the visible tiles. Lamps on bridges =computed (2026-10-20) — a lamp a drawn road or path deck holds stands on the deck (deck: the deck's top there, from its line along its axis, the viewer's deckAt) when its nearest highway within 15 m is a bridge or the ground below is water or railway. Before, the gate dropped those over water or a railway and stood those over a meadow or a street on the ground under the deck, their 5 m posts ending in it: Dresden 36 (the Budapester Straße's and the Löbtauer Brücke's over the tracks), Hamburg 32 (the Kennedybrücke's over the Alster, the Reesendamm's, the Altmannbrücke's), Grimma 11 (the Pöppelmannbrücke), München 8 (the Boschbrücke), Meißen 1 (Schloßbrücke). One mapped within 2 m of a street's line on the deck — in its lanes — stands at the deck's nearer side, 0.4 m in, straight across its axis: on the ground such a lamp moves to the class raster's kerb, which on a bridge is the ground's under it (Meißen's Schloßbrücke lamp, mapped on its lane). A lamp under a rail deck, or beside a street the bridge spans, lights that street and stays on the ground. Its post height is read in DOM1 over the deck's top, not over the ground below (over the river the deck around it was never clear, so no bridge lamp gave a reading). Street furniture on bridges =computed (2026-10-20) — by the same rule a bench, bin, bollard, bicycle stand or shelter stands on the deck where it is mapped; before, everything on a deck was dropped; one in the lanes (a bench, bin, stand, shelter, post box, column, picnic table or drinking fountain) moves to the deck's side as the lamps do (Meißen's Schloßbrücke bin). Hamburg gains 28 pieces (12 benches, 7 bins, 6 stands: the Busanbrücke, the Brooktorkaibrücke, the Kehrwiedersteg), Meißen 4 (the Johannesbrücke's benches), Dresden 2, Grimma and München 1. A signal or hydrant sign mapped in the lanes stands on the deck too (2026-10-21, furniture.py ON_DECK_KERB; before, it was left out: the kerb it moves to on the ground is the class raster's, the ground's under the bridge). A directed signal goes to the right of its traffic, facing it, at its carriageway's kerb (the street's width, else its class's half-width, + 0.6 m) or the deck's edge if that is nearer — a deck over a station's tracks is 30–46 m wide —; one without a direction to the nearer side, no farther than the kerb. A hydrant sign goes to the deck's nearer edge. A wall clock hangs on a facade within reach, else it is dropped. Kept: 11 signals and 3 hydrant signs (Hamburg 6 + 1, its signals on the Altmann-, Kornhaus- and Rentzelstraßenbrücke; Dresden 3 + 1; München 1 + 1, its signal on the Maximiliansbrücke; Leipzig 1). Lamps on a wall =computed (2026-10-20) — a lamp OSM tags support=wall_mounted or lamp_mount=wall (and not on a wire) hangs on the facade when an OSM outline lies within 1.5 m of it: its head 3.5 m up and 0.5 m out of the facade on the side no building covers, on a short bracket from the wall (wall: its foot), with no post. The 3.5 m is the Mapillary photo at Meißen's Burgkeller, lanterns on brackets at about the first floor. Before, they stood as 5 m posts in the facade line, and lamp-heights read the eave above them as a post. Meißen hangs 16 of its 24 wall lamps on a facade, Dresden 5; no other city tags one where a facade stands. München's lamps on wires were already hung (71 across its streets). Seven more of Meißen's hang on the Burgberg's terrace walls by the Schlossstufen (2026-10-21, on: "wall", lamps.py Walls): with no outline within 1.5 m, a mapped wall, retaining wall or city wall within 1.5 m holds the lantern, its bracket's foot on the wall's line. It hangs on the low face where the DGM 2–4 m either side differs by 0.5 m or more, at min(3.5 m, the step − 0.5 m) in 0.5 m steps (6 at 3.5 m, 1 at 3.0 m); a freestanding wall at its tagged height, else its kind's default, on the side of the nearest street (else the side the lamp is mapped on). A wall lower than 3 m — a garden wall — holds none. One stays a post. The card's lineage names the wall and the side the DGM told. Lamps hung across the street =computed (2026-10-08) — OSM maps the old town's catenary lamps (Äußere Neustadt: Alaunstraße, …) in the carriageway, where they stood as posts in the traffic. A lamp hangs when support/lamp_mount = suspended/wire says so, or, untagged, when it stands in the road class within 2 m of a street's centre line with facades (OSM outlines) closing both sides within 12 m; its wire runs across the street to the facade on each side (within 20 m, a 2.5 m corridor either side so a gateway does not miss the wall) or to a mast at the kerb, its head 7 m up, the wire's ends 0.6 m higher (the sag). Any other lamp in the road class moves to the nearest kerb. Dresden: 19 hung, nearly all tagged (OSM tags support=suspended on 17 lamps, 16 of them in the road class; ~480 untagged lamps stand in it, most a pole the road's width covers). The wires are the trams' ribbons (wire-ribbons.ts), in the posts' dark metal so the head reads as hung. Mapillary's lamps in the carriageway hang by the untagged rule too (Dresden: 142, most of them in the Äußere Neustadt, re-baked 2026-10-08).

  • Lamp post heights, street by street =computed (2026-10-20) — every post stood 5 m tall, from a pedestrian lane to a six-lane road. The surface model shows the posts where the scan caught them: at a lamp clear of trees and facades the highest DOM1 − DGM1 texel within a metre is its head. Leipzig's primary roads' lamps read 10 m (quartiles 7.6–12.6 m), its residential streets' and footpaths' 5 m; Dresden's Ostra-Allee 8.3 m, the Sarrasanistraße 3.4 m; Mapillary's photos of a few of them agree (8–12 m masts with arms on Leipzig's and Hamburg's main roads). One lamp is a poor reading — the scan misses a thin post as often as it hits it (a quarter of Leipzig's read under a metre), a van or a branch reads as one, and the readings along one street spread by 1.6–1.8 m (interquartile) — but a street's lamps are alike, and the streets differ by more (3.4 to 10 m). So lamp_heights.py gives every post on a street with three readings their median, a post on another its kind of street's on the tile (main, mid, residential, walk; ten readings), else the 5 m it had (h, hs); a reading must stand alone (a crown over a lamp is tall over metres: Hamburg's footpath lamps read 7.4 m without that rule, 6.8 m with it) over clear ground. The viewer draws one post set per height and a taller post's light pool wider (lamp-layer.ts); the card says the height and how it was found. Arms (Ausleger) ≈assumed (2026-10-21) — a post from 7 m up (ARM_MIN_H), or of any height that OSM tags lamp_mount=bent_mast (bent, lamps.py), holds its head on an arm (arm: one compass bearing, lamp_heights.py arm_of) reaching to the nearest carriageway line within 15 m, a street's before a service way's (carriageways); a post standing on that line (within 0.5 m) gets none. DOM1 shows why 7 m: at clear lamps the highest texel lies a median 0.3–0.6 m towards the carriageway at 5–7 m (the mapped point's own scatter) and 1.0–1.3 m at 7–8 m (Hamburg's 6.5–11 m masts 2.0–2.7 m); Mapillary's photos show arms and bent masts from 7 m and heads on top of posts of 4–5 m. Where a service way is nearest and a street in reach too, the head reads a median 1.9 m towards the street and none towards the service way. A second arm between a dual carriageway's lines was dropped: DOM1 shows a head on both sides at 1 of 53 such masts, no more often than beside one carriageway. The viewer draws an arm of 0.2 · h (1.2–2.5 m, lib/city/lamps.ts armReach) leaving the pole 0.35 m under the head and rising to it, a flat cap over the head, the head, its light and its pool at the arm's end (lamp-layer.ts); the card's lineage names the carriageway it reaches over. The bearing is set without DOM1 too (a bent mast keeps 5 m and its arm). Dresden 331 arms (131 of them bent masts under 7 m) and 253 on Mapillary's lamps, Hamburg 926 of 1 933 posts (38 bent), Leipzig 31, München 25 (bent masts).

  • Street lamps OSM lacks, from Mapillary ◎detected — Mapillary's detected street lights ("map features": object--street-light, each one triangulated across the street photos that saw it; CC BY-SA 4.0) fill in where OSM is thin. On Dresden's spawn tile OSM maps 339 lamps; Mapillary places ~1 500 street lights, of which 125 lie within 5 m of an OSM one (its positions are 2–6 m off; median 4 m against the OSM lamps it matches). Kept: seen since 2020; seen well (2026-10-21, seen_well): some image within 25 m of where it was placed (seen only from farther, it lay along a long ray, as a bridge's lamps did on the meadow below), and some image taken with the sun above −12° (dop_light.solar_position; at night the detector takes headlights and lit signs for street lights); not a tunnel's lighting laid on the ground above (within 20 m of a road in a tunnel and nearer it than any way on the ground, Tunnels: Meißen's ten over the Schottenbergtunnel); the two detectors' copies of one object merged within 4 m; nothing within 8 m of an OSM lamp; nothing inside a LoD2 footprint (a facade lamp, or a placement through the wall); one in the carriageway moved to the kerb, none on water, railway or a bridge deck (furniture.py's gate); kept lamps 7 m apart (one lamp placed twice from two sequences). QA 2026-10-21: of 45 kept lamps from before the gate (Grimma, Meißen, Unna), 31 were real, 7 wrong, 7 unclear; the gate kept all 31 real lamps and dropped 5 of the 7 wrong ones; two or more images says nothing (98–100 % have that many). Precision after the gate: Grimma ~75 %, Meißen ~57 % before the tunnel rule, Unna ~83 %. Lamps after the re-bake of 2026-10-21: Dresden 5 439 on fifteen tiles (6 792 before the gate), Grimma 949 (944 on 33340_5678, where OSM maps 66), Meißen 122 (OSM 696), Unna 133. Stood by the lamp layer like OSM's; the card names Mapillary. Its own file (dlm/mly_<t>.geojson), never merged into the OSM files: ODbL and CC BY-SA do not mix in one database. pipeline/bake/mapillary.py (the fetch caches the tile's street lights with each image's place and time under <raw>/mapillary/<t>.v2.json, it needs MAPILLARY_TOKEN; without that cache the file is empty and the lamps are OSM's alone), lib/city/mapillary.ts; per site Site.mapillary with objects (Dresden, Grimma, Meißen, Unna), whose credit joins the footer (mapillaryCredit: "Lampen, Schaufenster", or "Lampen" alone where no panorama is measured — Unna, panoramas: false). Its litter bins are no longer taken (🗃️ below). Not yet: Hamburg and München (Hamburg 2 731 raw, 745 kept against OSM's 553; München 5 180 raw, 2 273 kept against 461; 3 of 12 sampled in each were lamps hung on wires drawn as posts — they need a hang decision from the photos or DOM1). Not (yet): Mapillary's benches and bicycle stands (fewer than OSM's: 111 and 48 against 587 and 292 on the spawn tile), traffic signs (no layer draws them), signals (OSM's are placed by the direction they face, which a detection lacks).

  • Facade readings from street photos (experimental) ◎detected — Mapillary's panoramas of a street front (since 2016; on Dresden nearly all from 2025), rectified onto the LoD2 wall with their computed poses and masked to the "building" class, give per wall the share of the upper wall that is not plain render (openings, frames, stucco, glass), the share clearly darker than the render, and how open the ground floor is; two sequences of one wall agree at Spearman ρ 0.66–0.71 (window size, window count, storey height and "ornament" did not, and stucco cannot be told from glass). A building takes the length-weighted median of its walls, graded: busy 1–3 (busy only where two sequences saw it), dark 1–3, and a shop where a shop sign hangs within 4 m of the wall or the ground floor is more than 80 % open. On Dresden 2 798 walls of 1 957 buildings over thirteen tiles (864 walls and 614 buildings on the spawn tile, a third of its street fronts; courtyards hardly at all; two tiles have no usable panorama). Since 2026-10-09 every committed site with panoramas is measured the same way (Site.mapillary; its detected objects — the lamps alone since 2026-10-21 — are taken in Dresden, Grimma, Meißen and Unna, objects): Leipzig 1 635 buildings on four tiles, Hamburg 109, Meißen 49, Grimma 229 (its eastern tile has no panorama), München 3 609; Unna has no panorama at all, only flat photos. Its own file (dlm/facades_<t>.json, CC BY-SA 4.0), read at runtime into the flags column above the OSM bits (lib/city/facade-reading.ts); the clay draws a fine plaster relief lit from above on mid and busy facades, a tone a few per cent darker or lighter, a ledge at every storey of a busy front under a pitched roof, and the shop zone of a signed shop (Fassadenbild, visual-style.ts facadeReading). No windows from this reading: its window size, count and storey height did not agree between two drives (the windows' rhythm comes from the traits measured since, Windows on every house). The photo analysis is the facades step (pipeline/bake/facades.py, facade_measure.py): bun run fetch plans the walls, asks Mapillary for each panorama's pose and segmentation, measures its 2048 px thumbnail on every wall it serves and deletes it, caching only the measurements (<raw>/mapillary/facades/<t>.jsonl); bun run bake --step facades grades them per building. 13 717 panoramas measured for Dresden's fifteen tiles in about an hour (the spawn tile's 3 341 in six minutes); re-baked from the step, the spawn tile agrees with the spike's hand-converted file on busy for 584 of 590 shared buildings and on dark for 581; shop is set on 59 more, since the step joins the store signs per building (the spike's table had matched them to walls by an index from a different wall list).

  • Shopfronts from street photos (experimental) ◎detected ≈assumed — the same panoramas measured once more per wall (<raw>/mapillary/facades/<t>.v2.jsonl; the first fetch's cache stays): the ground floor's profile along the LoD2 wall in 0.5 m bins from its start vertex — how open each 0.3 m row of the band 0.3–3.4 m is against the ground floor's own plaster (the brighter half of its smooth pixels: a mean over piers and glass, as the upper-wall reading takes, makes a whole shopfront read open), where Mapillary's segmentation finds a store sign (object--sign--store, its band in metres), and the top of wide, low openings. Per wall (pipeline/bake/shopfronts.py) each sequence's median profile is shifted onto the best-seen one by the best correlation within ±1.5 m (taken only where it beats no shift by 0.15 and reaches 0.5), every image votes per bin — open where both knee height (0.6–1.2 m) and eye height (1.2–2.4 m) are at least 60 % open, so a house window above its sill and a dark plinth are no opening —, two images must agree, and the runs of open bins at least 1.2 m wide are the bays (three or more of nearly one width take their median width; none is added). A glazed row — the bins any image saw at least 75 % open over at least 6 m of the wall, the modernist shop rows the boulevard's panoramas see one stretch at a time — takes one image's verdict where no other looked and bridges unknown gaps up to 1 m (a tree trunk, a column): its glass runs from pier to pier (row). A wall is a shopfront with a bay over at least half its decided bins, as a glazed row, or with a store sign seen twice (or once beside a Mapillary sign feature). Over a shopfront the surface model may show a canopy (canopy): per metre along the wall nDOM = DOM1 − DGM1 every 0.5 m out to 8 m; where it is 2.6–6.5 m high a metre out, no higher than the building behind, and stays level (−0.6/+1 m) until it drops off, that is the canopy's depth; runs of at least 4 m (one miss bridged) keep their median depth and height. Under a canopy the glass runs its whole length but for runs of 1.5 m or more two images agree are wall — the Hauptstraße's GDR pavilions are glass from end to end, and their canopy's shade and columns break the photo profile up —, and a wall of the same LoD2 object facing the same way that no photo measured gets the row too where its canopy goes on (src: "canopy"). dlm/shopfronts_<t>.json (per Building, per wall: base points in the tile's CRS, bays, sign, ground-floor top, canopies; credited to Mapillary and, with a canopy, GeoSN). All fifteen Dresden tiles measured (2026-10-08: 13 717 panoramas — the spawn tile's 3 341 first, the other fourteen tiles' 10 376 in 30 min of fetch at ≈ 6 images/s on four cores): 1 003 shopfront walls on 780 buildings, 223 of them glazed rows, 19 under a canopy (3 of them continuing one). The other cities (2026-10-09, Site.mapillary, the same step and constants): Leipzig 7 945 panoramas on four tiles, 1 130 shopfront walls on 843 buildings (226 glazed rows, 11 under a canopy — its DOM1 fetched for this); München 22 212 panoramas, the most of any city (11 679 on the Marienplatz tile, 10 362 of all of them from 2026), 3 079 walls on 2 229 buildings (889 glazed rows; no DOM1 on disk, so no canopies) — on the Marienplatz tile 1 741 of the 3 231 walls seen, a larger share than anywhere else, worth a look for false openings (arcades, dark plinths); Hamburg 842 panoramas (61 on the Jungfernstieg tile), 152 walls on 112 buildings (33 glazed rows; no DOM1 on disk, so no canopies); Grimma 855 panoramas, all on the old-town tile, 79 walls on 75 buildings (12 glazed rows); Meißen 196 panoramas, 23 walls on 21 buildings; Unna none — Mapillary has only flat photos there (1 676 in the square kilometre at the old-town tile's centre, not one panorama), which the rectification cannot place on a wall, so Unna is left without mapillary. The panorama search now asks for a cell the API keeps answering 500 on in quarters, as it does a full one (Leipzig's centre). Noisy: two sequences' profiles correlate at a median 0.31 unshifted, 0.56 after the shift; a single image misreads graffiti and painted panels as openings; a long wall is often seen by two images only. The building bake draws them (lib/city/shopfronts.ts, appendShopfronts, column source = 6), soft and abstract (Weiche Nische, 2026-10-08): per run of bays a surround 15 cm proud with rounded outer edges, a niche per bay whose reveal rolls into the front over a 7 cm quarter round (smooth normals), the glass at the niche's back 6 cm out of the wall from 0.5 m (a glazed row: 0.3 m) to the measured ground-floor top (3 m where none; under a canopy's slab, into which the head reaches over a strip of wall under 0.25 m) — no mullions, no frame boards —, the store sign painted flush on the wall at the measured sign band (cut out of the wall and laid back in a darker tone; until 2026-11-10 a rounded fascia 10 cm proud, which read as a slab — Manuel chose "nur aufgemalt") (none under a canopy: its fascia is the sign band); a canopy is a slab 0.3 m thick under its measured top from the wall out to a 0.2 m deep fascia at its edge, at most 1 m tall, with 2.6 m headroom, its outer edges rounded, in the host's own clay (its columns are not measured and not drawn). The glass is a muted tone a little darker and cooler than its wall (not black), with the head's soft shadow over its top (paneShade); it is smooth (roughness 0.2) and catches the sun calmly; the upper pane and a grazing view hold a calm sky (paneSky), so it reads as glass out of the sun too. Shop windows on a shop's ground floor were the first exception to the glass veto (🗃️ Procedural window grid, user decision 2026-10-08; the veto lifted 2026-11-06); a wall no photo shows as a shopfront gets none (🗃️ A shop window at an OSM shop node). Passages (2026-10-21, G20): a bay crossed by an OSM tunnel=building_passage or covered=arcade way (4 m inside the wall to 1 m out) is no shop window; the wall keeps its sign and canopy. Dresden 8 of 1 147 bays, Leipzig 10 of 1 476, München 31 of 4 774, Hamburg 1 of 165, Meißen and Grimma none. Without the facade cache the step only takes the passages out of the committed file (it used to write an empty one). Openings in the wall (2026-11-10, Manuel: the shop windows to follow the windows' logic, "direkt beides zusammen"): each bay is cut out of its LoD2 wall, a 0.2 m box of reveals in the wall's clay behind it and the glass at its back, its sill at the stall riser's top (lib/city/shop-windows.ts); the plinth band breaks for each window (a first cut set the sill 0.2 m over the plinth's levelled top, and the band read as lying in front of the glass: Manuel, same day); the surround 15 cm proud with a soft niche per bay is gone, and with it the risers' feet on the ground (ADR 0030's shopfronts joins). Dresden's Neustadt tile (33412_5656): 327 of 344 bays cut (the rest stand where no LoD2 wall covers them), its vertices 5.71 → 5.63 million (−1.4 %).

  • Windows on every house (Fenster) ◎detected ≈assumed =computed — drawn by the clay (ADR 0010), their rhythm measured where street photos saw the house and inferred where not. The measurement: the same panoramas measured a third time per LoD2 wall, over the upper storeys (from 3.6 m to 0.3 m under the eave; <raw>/mapillary/facades/<t>.v4.jsonl, the v1/v2 caches stay; pipeline/bake/facade_traits.py), up to six drives a wall, each pose turned from Mapillary's east-north-up onto the UTM grid by the meridian convergence (−0.98° in Dresden, +1.9° in München; before 2026-11-01 it was left out, 0.25 m along a wall 15 m away): column and row profiles in 0.1 m bins of how dark the wall is against its local plaster and how much of it is an opening, the upper wall's vertical edges in 5 cm columns (the jamb profile), the plaster's vertical and horizontal edges, its fine texture, the openings' surround against the plaster further out, and the ground floor's plaster rows. One drive's pose is off along a wall by about 0.2 m, two drives' by up to a metre, so the features (pipeline/bake/windows.py, step windows) are read from what does not move with the camera — and where the axes stand from the drives registered onto each other (below). The column profile's period (autocorrelation: the window axis spacing, 1.2–7 m) and its peak (grid regular ≥ 0.45, loose ≥ 0.25), the row profile's period (storey, 2.6–5 m), the openings' share, the profiles' dark duty times their periods (width, height, proportion), the surround (Fasche), horizontal edges recurring at the storey (sill band), vertical edges at the axis (lisenes), the texture (ornament), the ground floor's tone step and joints (plinth, height, rustication) and a fine joint grid both ways (panel joints). Only images that see at least half the upper wall at 7 cm a pixel or finer count (over every image the axis agreed at ρ 0.58, over these 0.72 on the spawn tile); a wall's value is the median of its sequences' medians. Each feature's reliability is Spearman ρ between the two sequences with the most images on each wall; only ρ ≥ 0.6 over at least 30 walls may drive the model, fixed in the code from Dresden's tiles pooled (2026-10-09, 1 931 walls with a usable image on thirteen tiles):

    featureρ pooled (walls)ρ spawn tile (walls)drives the model
    axis spacing0.667 (273)0.717 (122)yes
    grid (peak)0.684 (334)0.676 (154)yes
    storey height0.605 (186)0.745 (90)yes
    opening share0.710 (334)0.733 (154)(recorded)
    darkness0.745 (334)0.784 (154)(recorded)
    width duty0.662 (333)0.650 (153)via width
    height duty0.465 (329)0.390 (150)no
    window width0.640 (273)0.678 (122)yes
    window height0.615 (182)0.616 (87)via proportion
    proportion0.656 (156)0.713 (72)yes
    Fasche (surround)0.509 (239)0.545 (115)no
    Fasche, signed0.559 (239)0.551 (115)no
    sill band0.534 (85)0.649 (48)no (passed on the spawn tile alone)
    lisenes0.565 (143)0.494 (69)no
    ornament0.711 (319)0.743 (149)yes
    plinth tone0.594 (133)0.512 (67)no
    plinth height0.549 (133)0.437 (67)no
    rustication0.528 (143)0.517 (70)no
    panel joints0.466 (334)0.505 (154)no

    Each opening on its own (2026-11-06): the fetch also keeps every window-sized opening an image saw whole (op, facade_traits.openings: its place, width, height and how full its head's corners are). Their median width and height and the median of each one's width over its height (ow, oh, oprop) agree between two drives far better than the profiles' duty on the height and the proportion: ρ 0.69, 0.71 and 0.76 against 0.70, 0.59 and 0.65 (Dresden's tiles pooled, 141 walls), and the model takes them first. The head (straight, a segmental or a round arch) agrees at ρ 0.67 but is not drawn: of five Dresden walls it read as round, two had round arches and three straight windows behind half-drawn blinds.

    A wall gets a model where at least two images found an axis period and 60 % of them lie within 15 % of their median, the period is a window rhythm (1.8–4.8 m: not a stucco band, not its double) and the grid is regular or loose — no window on a wall without that measurement: the axis spacing; the window's width measured (0.7–2.2 m, a 0.6 m pier kept: the openings' own, else the profiles') and its height from the measured proportion (0.8–2.4 m), else Dresden's medians (width 0.44 of the axis, width over height 0.77); ornament where the texture is in the upper third. dlm/windows_<t>.json (per Building, per wall: the wall's base points, its eave, images and sequences, every trait, the model; the tile's ρ table; credited to Mapillary, CC BY-SA 4.0); a tile without traits writes an empty file. Dresden, 2026-10-09: 13 717 panoramas fetched again in ≈ 75 min (three workers), 1 931 walls with traits, 588 modelled (389 regular, 199 loose, 409 with a storey, 108 with ornament) on 527 buildings; two tiles (33408_5656, 33416_5658) have no panorama on their walls and 33410_5654's see too little. The rhythm per building (lib/city/windows.ts windowSpecs, in the building bake; one more band of the object table, WindowSpec: axis spacing, width, height and a style word), best first: the model on the building's own walls (the wall most images saw); another part of its building; the nearest measured building within 30 m (150 m until 2026-10-21, G14: on Dresden's sample tile the rhythms lent fell from 1 211 to 278) with the same roof form and an eave within 4 m or 35 % (score: distance + 10 m a metre of eave — a block's houses share their era); its type — a town house (pitched roof, storeys from 3.3 m), a walk-up (lower storeys), a house (eave ≤ 6.5 m), a block (flat roof, eave ≥ 7 m) or a low flat building. A type's numbers are Dresden's measured medians (scripts/windows-report.ts, 2026-10-09: town house 442 walls, axis 2.72 m, width 1.18 m; walk-up 104, 2.68 m, 1.14 m; block 37, 2.62 m, 1.05 m; house 5, 2.56 m, 0.94 m): axis 2.7 m (house 2.6, low 3), width 1.15 / 1.1 / 0.95 / 1.2 / 1.4 m. A window's height follows its storey: its head 0.75 m under the next floor in a town house, 0.55 m in a walk-up, 0.5 m in a house, 0.45 m in a block (at most 2.4 / 1.8 / 1.45 / 1.6 m), the sill 0.85–0.9 m over the floor; a measured height scales that by its share of the type's median (the photos read openings short — the reveal's shade, curtains —, a town house's 1.79 m median against the 2 m its storeys give; 1.63 m from the profiles until 2026-11-06). A measured rhythm is drawn towards its type's (TOWARD_TYPE: the ratio to the type's value to a power, bounded — the axis ^0.75 within 0.75–1.35, the width and the height's share ^0.8 within 0.65–1.5): the axis is noisy (ρ 0.67 between two drives) and a row of houses whose axes differ by half reads restless, but the window read off each opening keeps most of its own size and shape — until 2026-11-06 it was held within a fifth of its type's (width ^0.5 within 0.8–1.25), and every house's windows came out alike. The Detailansicht names the form: hochkant (width under 0.8 of the height), quadratisch, liegend (from 1.25) or Band (lying and three quarters of its axis wide; windowForm). The storey snaps so that a whole number of floors reaches the eave (OSM's levels first, else the neighbours' storey, else the height's estimate) — never the photos' storey: the row profile's period reads short (the reveal's shade, a sill band, a Kniestock's row), and on Dresden's measured walls of buildings OSM counts it gave the right number of rows on 47 % and one too many on 42 %, the neighbours' storey 74 % and 9 % (2026-11-06; until then it decided wherever OSM counts none). Where OSM's count itself seems to hold the attic (a storey far lower than the neighbours'), it stands: of the Dresden houses checked in street photos most counts were right (new houses of 2.75 m storeys among Gründerzeit ones), and where the attic was counted it was a mansard that LoD2 draws as upright wall up to its break. The storey is written into the object's row before the plinths and shopfronts read it. A town house gets a Fasche round each window and a raised ground floor (its windows 0.45 m higher, their heads where the upper floors' are), and every ground floor's windows open at least 0.2 m over its plinth's top — its highest piece, so a plinth stepping up a slope lifts them —, and over the ground in front of every wall, read in the tile's fine terrain 0.5 m out of the wall every 2 m along it, also of a wall standing on a step of its building, not where the ground lies more than 1 m under the wall (a wall over a lower roof) (wallRises; 2026-10-19/20, Manuel: the windows sank into the Brühlsche Terrasse — the plinth pieces alone reach only open street runs of buildings from 5 m, so the Kunstakademie's terrace walls had no ground under them). The plinth pieces' tops are levels like the ground's, not a floor for every wall. The lift is a class of 0.05 m up to 1.55 m; a wall whose ground or plinth rises past its reach (at most 2.2 m over the base, GROUND_REACH_M, and no higher than leaves the ground floor's windows at least 0.8 of their width tall, overGround) loses its own ground floor, signed like a shopfront's, the building's other walls keep theirs. Such a wall keeps that height (_FACADE's first door slot) and draws no row under it, so where the ground rises more than a storey the upper floors' windows go too (the Kunstakademie on the terrace). Dresden, 2026-10-20: wall feet whose windows the fine ground covered 11 925 → 0 (of ~610 000), objects without a ground floor 6 258 → 3 590 (the rest were shops). Only a wall the photos saw shop bays on gives its ground floor up (2026-11-10, Manuel, Alaunstraße: a house with an OSM shop or café lost the ground floor's windows on every wall, and a wall where one drive read a store sign alone — 318 of Dresden's 1 003 shopfront walls — kept blank plaster under its fascia — the bake's style word and the clay's shopOf gate, which also took a sign the photos read, each blanked it; now both keep the house's windows, the shop zone still painted darker). The storeys count from the street (liftToStreet): a LoD2 base is the lowest ground round the whole footprint — a courtyard's, a light well's, a cellar ramp's — and lay a median 0.3 m under the street on the spawn tile, so each LoD2 object's base moves up to its building's street level where its plinths say where that is (the highest of its pieces' lowest ground, at most 1 m over the lowest), from at most 1.5 m under it, its eave staying where it is; the storey lines, the Gurtgesims and the windows then start from the pavement, not behind the plinth; a measured or borrowed loose grid shifts each axis by up to 8 % of the spacing (a hash per column), measured ornament a Verdachung over each window. None on a facade of its own (a church or hall — ALKIS function or OSM storeys over 5.5 m —, not a Wikidata landmark's flag alone: most are town houses) except a palace's: a public use (ALKIS 31001_3010–3072 but the churches' 3038–3048 — a Schloss, museum, theatre, parliament, court, school, hospital; no barracks, bunker or prison, no station, no sports hall; palatialFacade) gets the type palace (3.6 m axis, 1.45 m wide, up to 3 m tall, a Fasche, 21 cm reveals) on storeys of OSM's count where it gives 2.4–7.5 m, else the eave split into ~5.5 m storeys (palaceStorey), and only its own walls' measured rhythm, never a neighbour's (2026-10-19: the Zwinger, Albertinum, Japanisches Palais, Residenzschloss and Landtag stood blank; 540 objects in Dresden gain windows, none lose any). None on glass or metal, on a garage, shed, warehouse or plant (ALKIS 2461–2464, 2513, 2522–2523, 2720–2729), under 3.2 m of wall, or on a part the bake adds (the scan's sheds, gaps, dormers, shopfronts). Where on its wall (facadeAttribute; the glTF's _FACADE, a snorm16 vec4 per vertex, 512 m a unit): metres along the wall from its middle, the wall's length — negative where its ground floor has no windows (over a shopfront, or with more doors than the two slots), 0 on a party wall —, and up to two of OSM's doors along it, each its place (in 0.1 m) and its width rounded up to a class (0.2 m steps to 2 m, whole metres to 6 m, the widest OSM width the doors bake takes) in one code: a ground-floor window keeps at least 0.6 m of wall between its opening and a door's (the door's surround, a pier); a roof vertex carries −1 in the first door's slot, the clay's roof flag. A wall is one object's coplanar, vertex-connected triangles; a party wall is one with another wall back to back against it (opposite normals within 10°, planes within 0.5 m, over a quarter of it covered). Where the axes stand (pipeline/bake/window_axes.py, 2026-11-01): each pair of a wall's images is registered by its jamb profile (the shift within 2 m with the best normalised correlation, at least 0.5 and 0.12 above any shift 0.4 m away — a period's alias), one offset per image solved from the pairs (robust least squares, Huber at 0.12 m) with the drives' mean offsets averaging to zero, the openings profiles moved by their offsets and averaged, a window a run of it at least half its peak and 0.25 open, 0.4–3 m wide; the phase (model.phase, metres from the wall's start to an axis) the circular mean of their centres modulo the measured spacing where they agree (resultant ≥ 0.8) and at least two windows were found. On Dresden's spawn tile the drives' offsets spread by 0.2 m (robust), the registered pairs close to 1 cm, the mean of three or more drives has a standard error of 0.12 m (median); 66 of 199 modelled walls got a phase (42 of them seen by one drive, placed too — still nearer than a centred axis; a phase needs two images that register). The building bake (lib/city/windows.ts measuredAxes, measuredColumns) lays it on the mesh wall at that wall's own measured spacing — a house's front and side often keep different rhythms (the test wall: 3.88 m on the front, 2.14 m on the side, its object's spec 2.27 m) —: the wall's middle moves to the middle of the lattice's axes that fit 0.7 m and half a window in from its ends, and its length slot holds their count and spacing instead of its length (MEASURED_AXES: from code 16 384 on, 64 counts × 1–6.1 m in 2 cm steps; no wall length reaches 256 m), the window as wide as its object's, never wider than the spacing leaves (the shader holds it there too), and a loose grid's jitter left off. Every mesh wall in the footprint wall's plane that overlaps it keeps the lattice, so a facade of several LoD2 polygons runs on unbroken. Every site measured 2026-11-02 (walls modelled / phased): München 1 697 / 580, Dresden 577 / 157, Leipzig 448 / 79, Grimma 87 / 32, Meißen 18 / 6, Hamburg 12 / 1, Unna none (no panoramas). Elsewhere the axes are centred on the wall at the object's spacing with 0.7 m kept at its ends. The drawing (app/_components/clay-windows.ts, in the clay's fragment pass, walls only): per storey a row of openings from the ground floor's (a top floor's windows end 0.5 m under the eave, under the Traufgesims, shorter if they must, none if less than 0.55 m is left) — a recess the eye looks into, 9–18 cm deep by type (the ray against its box: the back, a side reveal or the head or sill reveal, each with its normal and its share smoothed over the pixel's footprint), its reveals the wall's plaster ×0.93, its back the wall's colour ×(0.40, 0.43, 0.50), lit by half the sun plaster would get, the sun reaching in only through the opening and only where it falls on the wall at a fair angle (receivedShadowNode); a sill 7 cm tall and 7 cm proud, 6 cm past the opening either side, throwing the sun's shadow down the wall; a Fasche 11 cm wide, 1.5 cm proud and 3.5 % lighter, with a hairline of contact shade; a Verdachung 9 cm tall, 10 cm proud; the wall rolling into the opening over a 6 cm quarter round at its sides and head (normals only, where it spans a pixel). Every edge is anti-aliased by the pixel's footprint along the wall and up; where a window is under 6 pixels wide the drawing fades into the facade's mean tone (the wall darker by 0.6 × the openings' share of a storey) and under 2.5 pixels only that tone is left, so a far street does not shimmer. At night 25 % of the windows glow warm, at 0.6 of Abendlicht (a hash per window, NIGHT_WINDOWS; 40 % at full until 2026-10-19). Fenster at full strength (WINDOWS, fixed). Drawn on Dresden's spawn tile (2026-10-09): windows on 308 buildings measured, 1 223 by a neighbour, 1 638 by type; of the window-carrying wall area 70 % has windows, 27 % is party wall, 3 % stands over shops (over all fifteen tiles 891 measured, 4 823 by a neighbour, 31 361 by type; 69–80 % windows). Of the objects with 3.2 m of wall or more, the spawn tile leaves plain only the facades of their own — 122 Wikidata landmarks (churches, palaces, ministries, halls: ALKIS function, or OSM's storeys over 5.5 m) and one glass front — and 11 garages and sheds. The building glTF trades its float roof flag for the snorm16 vec4 that carries it too (4 bytes a vertex more): over the fifteen tiles 6.1 % more to download, 15 % more GPU buffer; no window adds a triangle (the plinth and band pieces, following the street-level storeys, 1.1 % more). Replaces the geometry niches (🗃️ Window niches as geometry). User request 2026-10-09: "Details für die Fassaden von deutlich mehr Häusern (möglichst allen)". Facade heuristics (2026-10-21, G2). Flat-roofed town houses (G10/G13, flatTownHouse): housing (ALKIS 31001 without a commerce 2xxx or public 3xxx use) from 9 m of wall whose storeys are 3.4 m or more (OSM's count, or the neighbours' storey) or that is listed is a town house under its flat roof — windows of type tall, the plinth and the cornices, not a block (a 9 m bound: Meißen had 412 "town houses" without it, four in five under 9 m). Town houses among the flat-roofed houses from 9 m on one sample tile per site: Leipzig 805 of 986, Hamburg 241 of 326, München 296 of 676, Dresden 190 of 483, Unna 61 of 382, Meißen 102 of 183, Grimma 32 of 76. Window frames only on rendered walls (G12): a type's Fasche is not drawn on brick (a measured one stays); Hamburg 656 removed, Dresden 129, München 66, Leipzig 4. A wall's own eave (G22, WALL_EAVE): a wall whose roof edge stands over its object's eave carries the difference in _FACADE's second door slot where it holds no door (code 32 412 + round(Δ / 0.1 m)), and its top row sits under its own eave; no new buffer or varying; 36–62 % of the wall triangles carry one. Party walls of low houses (G16, PARTY_LOW_EAVE_M): a house of at most 9 m of wall whose party wall (over a quarter covered) rises at least 2 m over the neighbour's top gets windows above it; 3.7–12 % of the party-wall triangles reopened. Structures windowless (G17): an object whose ALKIS function lies outside 31001 (a tank, a silo, a crane, a canopy) gets no windows, doors, plinth or shop band; the card says Nutzung ohne Fenster (noWindows: "use"); windowed structures → 0 on every sample tile.

  • Street furniture ❝taken ≈assumed — OSM benches (amenity=bench, points and the ways a bench is sometimes drawn as), picnic tables, litter bins (waste_basket), bicycle stands (bicycle_parking, not wall loops), bollards, post boxes and stop shelters (highway=bus_stop / public_transport=platform with shelter=yes, amenity=shelter of shelter_type=public_transport; one per 8 m) → one point layer per tile (ODbL), about 3 000 objects over the four first tiles when first baked; 11 782 of every kind (plan 030's included) over the fifteen tiles (2026-09-26). OSM seldom says which way a bench looks (direction, on ~3 %), so the bake turns an untagged object to the nearest highway within 25 m (across it when it stands on it); a bench way stands at its midpoint at its mapped length, facing the path side. Only what the tags carry varies: backrest=no benches are stools, a stand's capacity gives its hoops (two bikes each). A bench, stand, bin, post box, column, stop or shelter mapped in a street's lanes (the road class, the nearest way a carriageway, STREETS) moves to the nearest kerb (2026-10-08: Alaunstraße's stands stood in the lane); in a square, a pedestrian zone or on a platform — the road class covers those too — it stays where it is mapped. Dropped: indoors, underground, classes 5 and 8, under a bridge deck (the terrain there is the river or the road it spans); on a bridge's road or path deck a piece stands on the deck (2026-10-20, below). A bollard keeps its tagged height and a metal material (the Stallhof's 1.46 m bronze columns of 1591 are mapped as bollards). pipeline/bake/furniture.py; the viewer instances one small, abstracted model per kind — softened blocks, capsules and single tube strokes, a bench one extruded seat-and-back profile — in the scene's palette at its own brightness (a warm sand-clay for seating, the roads' lavender-grey for metal, the buildings' clay for stone; one matte vertex-coloured material). Low contrast on purpose: a first, deeper palette (honey 211/176/140, slate 157/156/176) stood out from the pale ground and the clay buildings; the pieces now differ from them by form and shadow, not by tone (furniture-layer.ts, lib/city/furniture.ts). Not (yet): bicycle-parking areas, shelters mapped as areas, planters, traffic signs (OSM maps ~100 here) and street-name signs (almost none). Facing the water (2026-10-21): a bench without a direction with water (class 8) within 12 m on the side away from its way (±60°) looks at the water, and so does a playground's piece (Hamburg 200 of 1 413 benches, München 101, Dresden 47, Leipzig 16). Bin colours: a bin's own colour, else the site's when ≥ 20 bins are tagged and ≥ 60 % agree — Hamburg's red (850 of 868); München stays grey (85 % of its 435 tagged), its own-tagged blue and green kept. Five muted tones besides the grey (binRed, binGreen, binBlue, binOrange, binYellow).

  • Signs and fixtures ❝taken ≈assumed (plan 030) — the same bake and layer, eight more kinds (the counts below are the four first tiles', BBBike 2026-09-19; over the fifteen: 175 columns, 786 signals, 24 hydrants and 1 482 sign plates, 20 + 3 clocks, 15 drinking fountains, 231 bus-stop signs): advertising columns (advertising=column, 85; a pale paper drum on a plinth with a darker ring and dome, three poster fields in the palette wrapped round it — colour, no text, turned by the scatter yaw so no two read alike; the 4 lit=yes ones glow softly with the night factor), traffic signals (highway=traffic_signals, 288: a 3.2 m grey pole, a three-lamp head a shade deeper than the metal, the glass barely darker and unlit — the viewer has no traffic to time. The node sits on the carriageway at the stop line, so with traffic_signals:direction the bake walks it to the kerb on the right of the traffic it controls — the kerb from the class raster, ≤ 15 m — and turns the head to face that traffic; without a direction it goes to the nearest kerb, facing the nearest way; kept where it is off the road), fire hydrants (emergency=fire_hydrant: the 6 pillars a 0.8 m red-ochre pillar; the 454 underground ones only their sign plate on a post (the red-bordered white plate abstracted to one soft rose field), moved out of the lane to the nearest kerb — drawn at 70 %, the plan's STOP fallback taken without a GPU plate: the smaller plates were chosen over dropping them), clocks (amenity=clock: 8 on a pole as a double face on a 3.5 m post; 3 of the 8 wall clocks hung on the nearest OSM building outline ≤ 3 m, facing out — the plan asked for a ray against the LoD2 BVH, the bake uses the outlines for the same reason as the tram rosettes; tower clocks and sundials dropped), drinking fountains (amenity=drinking_water, 8: a slim bronze column with a small basin) and the bus stop's "H" sign (highway=bus_stop without shelter=yes, 79, dropped within 8 m of a shelter: the "H" abstracted to colour fields — a soft green disc in a yellow one, no letter — on a 2.6 m pole, a timetable box — shared with the tram stops of plan 024). The clocks' hands show the scene time: two hand quads per face, turned in the vertex shader from one shared uniform that setSun updates on the minute only (clockMinutes, scene-environment.ts); the hands (a soft slate, not black) never cast, so the shadow map is not redrawn for them. All in the furniture's soft pastels: nothing near-black, no fine detail. The re-bake leaves every earlier object byte-identical (checked per tile); the new kinds are appended. Look unverified on a GPU. Pedestrian signals, mast arms, repeaters (2026-10-21; before, no signalled crossing was read — the QA counted Dresden 934, Hamburg 726, München 666, Leipzig 583): a highway=crossing|traffic_signals with crossing=traffic_signals or crossing:signals=yes stands a 2.5 m post with a two-lamp head and a push button at each kerb end of its crossing way (the footway through the node, else straight across), looking across; within 2.5 m of another, one; on a deck by Decks.kerbside (pedsignal: Dresden 1 718, München 1 312, Hamburg 1 270, Leipzig 1 082). A directed signal on a road of ≥ 3 lanes hangs on a 5 m mast arm over the lanes (arm; Dresden 271). An approach whose junction lies within 40 m past its stop line repeats its signal at the far kerb on the traffic's right. A carriageway is its width, else lanes × 3.25 m (≥ 2), else its class's (half for one carriageway of a dual main road); where the road class reaches beyond 15 m round a signal the kerb is searched to 30 m, then taken from the way, instead of the signal being dropped. Signals: Dresden 784 → 1 338, Hamburg 588 → 894, München 571 → 837.

  • Flagpoles ❝taken ≈assumed (2026-11-10) — the same bake and layer: every man_made=flagpole point stands as a slim mast, 7 m or its tagged height (3–30 m, one outside is left out; few are tagged), with a cloth in the wind (k: flag). The flag is drawn in its colours only where the map names it (flag:name, flag:wikidata, subject, country: the first whose name the map fills decides — a Wikidata item the viewer does not know leaves it to the next —, the flag's own name first — Germany, the EU, Saxony, Bavaria, Hamburg), as pastel bands without emblem; every other one — most are a shop's or a firm's (flag:type=advertising, never in a country's colours), some a city's — flies a pale cloth; country counts only on a national flag (no flag:type or national) that names no Wikidata item of its own or of what it stands for (flag:wikidata, subject:wikidata) — on a firm's or a city's it is where they are; a european one flies Europe's. ≈ assumed: the wind (one bearing for the city, from the west-south-west, ±12° a mast by a hash), the cloth's size and its ripple. A mast mapped inside a building outline or on its line stands on its roof and is left out (7 masts: Dresden 2, Hamburg 2, München 3; not under a building's raised floors or over one under the street, not under a canopy, a carport or a shelter, in a stadium, a ruin, a building site or an outline tagged building=no), as is one tagged on a roof (support=roof, location=roof or rooftop); a flag on a facade (support=wall_mounted) is not mapped in any of the seven sites' extracts and would be left out. The cloth moves, so its shadow is the sheet at rest (castShadowPositionNode, ADR 0009). Masts after the fresh BBBike extracts of 2026-11-07 (Last-Modified): see the bake's counts in the PR.

  • Playgrounds ❝taken ◎detected — OSM leisure=playground outlines (≥ 20 m²; 339 over the fifteen tiles, 88 on the four first) → a pale sand floor flush on the ground (a breath warmer than the paving), seated on the ground under each ring vertex (densified to 2 m), and only the equipment OSM maps (playground=swing/basketswing/slide/sandpit/ climbingframe/structure/climbingwall/springy/spring_board/seesaw/ roundabout/playhouse, ~100 pieces) standing on it; a sandpit drawn as an area is its own sand slab, one drawn as a way stands at its midpoint turned along it. A playground mapped without equipment (the QA found 7 of 9 checked had some) stands what the surface model measures there (2026-10-21, src: "dom"): DOM1 − DGM1 blobs 1.2–4 m tall, 2–40 m², ≥ 1 m inside the outline (its hedge or fence), touching nothing taller, NDVI < 0.3 and no canopy tree within 1.5 m, the largest eight — a long narrow one a swing across it, any other a climbing frame. Nothing stands where nothing was measured. Dresden 200 pieces on 122 of its 263 empty playgrounds, Leipzig 54, Hamburg 41. On the Saxon leaf-off flight a bare shrub reads the same. The pieces are not catalogue equipment but soft, single-coloured sculptures (an arch, a wave, a faceted dome, an egg) in five pastels taken from the scene and lifted to the buildings' brightness — a first, literal rendering (A-frames, ladders, a rose safety floor) read as busy and out of place, and saturated pastels stood out as much. Same bake and layer as the street furniture.

  • Fountains, statues, memorial stones, columns ❝taken =computed — the Basis-DLM's monument points (sie03_p, OBJART=51009, BWF 1750/1770/1780, with their official names; GeoSN) conflated with OSM's amenity=fountain points and basin outlines (ODbL; the DLM names none of its monuments a fountain and gives no basin size). A DLM monument on an OSM fountain names it; the other OSM fountains are added. What a monument looks like is in no register, but its bulk is measured: DOM1 − DGM1 (the canopy's nDOM) holds the sculpture groups of the Albertplatz fountains as ~4 × 5 m bodies 3.7 m tall, the Goldener Reiter as 7 m. Where that body stands clear — one connected patch within 6 m (or inside the basin's water), ≤ 60 cells, below 8.5 m and touching nothing taller (a leafless crown reads the same on a 1 m grid) — the bake writes it as relief (27 of 174 monuments). pipeline/bake/monuments.py → monument-layer.ts: a fountain's relief is smoothed (reliefSurface: ×4 bilinear, one binomial pass) into one soft form in the buildings' clay, seated per sample on the terrain (a monument's is composed, below); a monument nothing measured is an abstract clay marker (rounded pillar · slab · shaft, MARKER_SHAPE) — no invented figure. Basins are the OSM outline as a low clay rim (the water its 0.35 m inset) over the highest ground under it, with translucent water bells that grow with the basin, round a measured sculpture when there is one (splash pads flush, reflecting pools still); a point fountain is a 2.2 m round basin. The canopy loses the "trees" its own bake planted on a measured monument (onRelief). The fountains move gently — each bell breathes on its own phase, droplets run down its curtain, light shimmers across the water — and by night the water keeps a faint glow from below and the jets a faint warm white, and a fountain's sculpture is lit softly from its basin (until 2026-10-20 a glow and jets bright enough to outshine the lit streets, "viel zu hell", the maintainer) (setFountainTime and the scene's sceneNight, one shared clock and night factor). Merged/instanced, nine draw calls per tile at most (eight until 2026-10-21). Not walk-blocking (collision is buildings only). What a monument is, from OSM (2026-10-08): the DLM's Denkmal, Standbild and Gedenkstein share one code, so 179 of Dresden's monuments were one rounded pillar. OSM's memorial=* / artwork_type=* tell a statue from a bust, a free sculpture, a stele, a stone or an obelisk: a DLM monument takes the form of the OSM memorial or artwork of its name within 15 m, else the nearest — never a plaque, which is the sign that explains it — and a name that says Obelisk, Stele or Büste wins. The marker is then that form's parts in clay (markerPieces: a body-sized pillar on a 1.6 m pedestal, a head-sized one on a slender plinth, a broad mass without a pedestal, an upright slab, a low rounded block, a four-sided needle on its base), scaled whole to a tagged height; still no figure. A measured relief wins over any form. Free-standing OSM sculptures and memorials the DLM lacks are added (160 on Dresden's fifteen tiles, mostly park and estate sculptures; not inside a LoD2 footprint, not within 3 m of one kept). The card names the form, the artist (artist_name) and the material (material, in German). On Dresden: 274 of 396 non-fountain monuments carry a form. Not used: Mapillary detects no monuments (its map-feature classes are street furniture and signs; checked over the Altstadt, 2026-10-08), Wikidata types more finely (32 Reiterstandbilder round Dresden) but rarely carries a height (7 of 941) and its points match ours worse than OSM's (its materials are used, below); Panoramax and Mapillary photos would need photogrammetry per monument; Wikimedia Commons holds no 3D scan of a Dresden monument. What a monument is made of, as a tone (2026-10-08): where OSM names no material, Wikidata's P186 does for the monuments, sculptures and memorials it knows (fetched per tile, data/_raw/<provider>/wikidata/monuments_<t>.json, Stolpersteine left out; the item of the monument's name within 40 m, else the nearest within 10 m): 15 on Dresden, the Goldener Reiter's copper and gold leaf among them, materialWikidata naming the item on the card. Since 2026-11-05 (plan 081) a monument OSM links to its item (wikidata, which now always means OSM's link) takes that item first, whatever the distance, and no neighbour can claim it. The viewer turns the material into a muted tone in the clay's palette (MATERIAL_TONE, markerTones: warm sandstone, grey granite, bronze's dark green-brown patina, copper's verdigris, a soft gold): the figure takes the most striking metal, the pedestal the stone named (else clay). A marker's pedestal piece and figure piece are tinted apart; a measured relief 3 m or taller is a figure on a pedestal, its upper 45 % the figure's tone (figureShare — the laser scan puts the Goldener Reiter's pedestal top at 4.6 of 8.5 m). Glass, plastic and ceramic stay clay. No texture, no gilding glint: one colour per part. The laser scan (LSC, the same November 2024 flight, ~16 points/m² there) shows the Goldener Reiter as a pedestal and an elongated body with the rider's peak — orientation and mass at 0.5 m, still no horse; not baked (a 380 MB download per tile). Measured monuments as composed solids (2026-10-08): the smoothed 1 m relief read as a heap (the Goldener Reiter a soft gold mound), so a monument's relief is no longer drawn as a surface; it is measured (measuredMarker): its cells' principal axis, footprint and top. A relief of 3 m or more with 6 cells or more becomes a pedestal at its footprint (x0.85) and the cells' median height, with the figure (the cells well above it) on top: lying along the axis where it is long (2.5 m and 1.6 times its width: a body block with an upright pillar on it, the horse and rider, the reclining river gods), else one upright pillar. Else the OSM form at the measured height and axis (an obelisk or stele always), else one mass at the footprint and height. The same clay solids as the markers, tinted per part. A fountain's sculpture keeps its smoothed relief (the winter housings below). What a part is made of, in its surface (2026-10-08): one smooth clay read as a plastic cast. A pedestal is now built as masonry (pedestalCourses: a plinth course a little wider at its foot, the shaft, a projecting cap; sharp-edged blocks, no rounding) and dressed as the bridges' ashlar (courses, running-bond joints, a shade per block). A figure's finish follows its material (partFinish, the instance attribute iFinish): stone tooled (a fine mottle, a touch less matte), bronze and copper a patina (rain streaks, verdigris in soft spots, roughness 0.55, a faint sky sheen at grazing angles), gilding smoother (roughness 0.33) with a warm sheen; no metalness, as no environment map lights the scene. A figure nothing names stays plain clay; a pedestal is masonry, named or not. Caveats, checked against the sources: DOM1 (November 2024) and DOP (March 2024) were both taken while Dresden's fountains are drained and their sculptures boxed for winter — the Albertplatz "bodies" are those housings (flat-topped, ~3.7 m), right in size and place, not the figures; since 2026-10-21 they are recognised (winter_cover) and drawn as a free sculpture's marker. The DOP shows the gilded Goldener Reiter only as glare (its shadow holds the horse's silhouette), so no colour is sampled. The laser point cloud (LSC, the only official source with more form) was unreachable from GeoSN's share when this was built; no openly licensed 3D scan of the landmarks was found. QA corrections =computed (2026-10-21, check-water W5–W9). Outlines: an OSM memorial/artwork outline ≤ 1000 m² stands at its representative point unless a node of it lies inside (Hamburg's Bismarck-Denkmal, 635 m²); larger ones are sites, left out. memorial=war_memorial and a bare historic=monument take the form their measured top says (default_form: < 1.5 m a stone, < 3 m a stele, else a statue; a column where a slim shaft stands; unmeasured a stele or a statue); an OSM name states a form as a DLM one does. Slim columns: a shaft over 8.5 m at a monument named column or obelisk — ≤ 24 cells above 0.3 of its top (12 for an unformed memorial), its ring ≤ 0.45 of the top at the median, under half in a LoD2 footprint — is measured at its height with its pedestal (slim_column), on the laser scan's surface where the provider's LAZ is on disk (hs: "lsc"): the Mariensäule 13.2 m, the Karolinenplatz obelisk 26.4, the Friedensengel 36.6, Dresden's Autobahn-Obelisk 12.9; drawn as a column (pedestal, round shaft, capital) or an obelisk up to 40 / 35 m, its girth at most ×3. Winter housings: on a site flown leaf-off (the canopy's median NDVI < 0.4: the Saxon sites) a box-shaped relief is the housing — cover, no relief: a fountain's figure is then the free sculpture's marker (coveredFigure; the Albertplatz's two fountains, until now drawn as their 3.7 m boxes). Heights in the open: without a relief, the tallest cell within 1.5 m where 80 % of the 2.5–4.5 m ring stays under 0.7 m (hm), raised to the form's least; above the form's most it is a branch or a mast and is dropped (Hamburg's obelisks 2.4–4.5 m, drawn 6 m before). A monument without a relief faces the centre of its square, else its nearest way (a). Small fountains: a drinking fountain or stone block node is ~1 m across (r), not the 2.2 m basin; a point fountain inside a LoD2 footprint moves out within 3 m or is dropped (Dresden 3 / 2, München 6 / 6). The jets run mid-April to October (fountainsRun, one shared uniform, fountainsOn in scene-environment.ts). Scratch bakes, HEAD code → this: forms Dresden 275 → 321, Hamburg 172 → 184, München 197 → 222; reliefs München 34 → 54, Dresden 82 → 88; hm Dresden 25, Hamburg 17, München 11.

Railway & bridges

All baked by pipeline/bake/rail.py, built per fine terrain tile in app/_components/rail-layer.ts and, the bridges, bridge-layer.ts (Y-up, part of the tile's dressing) on the cross-tile heightAt over every loaded terrain, so tracks don't truncate at seams (until ADR 0024 it was built once for the fixed block). Replaces the old "brown smear". All geometry is hand-wound to match its supplied normal (pushTri), so every material is FrontSide (halves shadow/fill cost). (Redesigned after the v1 per-line approach z-fought into ragged edges, fragmented, and stacked into "2-story" bridges — see 🗃️ below.)

  • Rails, ballast and decks from OSM where there is no Basis-DLM ❝taken (Hamburg, Berlin; 2026-10-01, ADR 0012) — rail.py run_osm writes the same four files from the site's extract (pipeline/bake/rail_osm.py). Rails: railway=rail|light_rail|subway|narrow_gauge (Hamburg's S-Bahn is light_rail, its U-Bahn subway; the trams stay the tram layer's), out of tunnels and off the underground, with tracks and electrified, clipped to the tile and snap-merged like the DLM's. Ballast: those rails off bridges, buffered by their bed width (4.5 m for one track, 4 m more for each further one), dissolved into one surface. Decks: every bridge-tagged highway/railway way ≥ 4 m (not a proposal or construction), on its man_made=bridge outline where it runs on one; else — OSM maps a carriageway, its pavements and a cycle track as separate ways — the buffers of one bridge's ways on one layer merged into one deck (decks(..., merge=True), merged_decks: the outline, the first name — bridge:name before name —, the weightiest kind, rail over road over path, the longest way as its axis), each at its width, else lanes × 3.25 m + 2 m, else the kind's width; a way on an outline another way already claimed adds nothing. The kind of an outline no way runs on comes from OSM's networks (Ground(tile, networks), the masks the DLM's lines fill otherwise). Deck heights (DOM1 over the DGM ramp), the superstructure, the fairway clearance and Wikidata are the DLM path's, unchanged. Hamburg (four tiles): 348 decks — the Busanbrücke, Brooktorkaibrücke, Kuhmühlenbrücke, the Berlin–Hamburg and Lübeck–Hamburg lines' bridges among them —, 81 km of track (OSM maps each track as its own way where the DLM bundles them, so the count is high against a DLM city's) and 0.27 km² of ballast. Before, the layer was DLM-only and Hamburg, a city of bridges, had none. Marked 🟡 (OSM + DOM1) in Sources by city. Not yet judged on a real GPU.

  • Tracks one by one from OSM ❝taken ≈assumed (2026-10-21, every site; not yet judged on a GPU) — every OSM railway=rail|light_rail|subway|narrow_gauge way above ground, clipped to the tile, with bridge, layer, its gauge g, electrified (1 overhead, 2 third rail, 0 none; an untagged way takes the nearest DLM line's ELK, a DLM U-Bahn line, BKT 1202, its third rail) and its mapped catenary masts m (rail.py rails, rail_osm.rails). The DLM's lines only where OSM maps no track within 6 m (each rest ≥ 20 m; rail_osm.claimed). The DLM gave a line one centreline and at most three tracks: 2 for Dresden Hbf's 15. Track-km Dresden 77.4 → 196.7 (overhead 70.7 → 107.2), Leipzig 25.4 → 54.4, München 24.3 → 35.7, Unna 8.3 → 13.8, Meißen 3.6 → 3.2, Grimma 0.74 → 0.77; Hamburg's 81.2 as before (26.4 km of it third rail). OSM's bridge is the line levels' preference: a deck on a bridge way, the ground off one (railPrefer; 20 % of Hamburg's bridge rails lay on the ground under their own deck). The OSM credit now stands on the DLM sites' rail files too. Covered and below ground: a way covered for more than 30 m (a lid or a building over it) is not drawn; layer < 0 marks a line below, which cuts under a lid up to 800 m at 0.5 % grade (lib/city/levels.ts coverCuts; 832 m of Hamburg's rails had lain on such lids); a cut more than 4.5 m under the ground with no deck over it is buried: no rails, ballast or catenary (BURIED_M: a lid needs room for a train under it). A shallower cut is open, whatever its layer: its trench is burned down to the track, so a drawn track never sinks into the ground. The line was 0.5 m until 2026-11-01, when the Standseilbahn's open cut at its top (layer −1, 0.6–1 m under the DGM) and the cuts before its tunnel portals vanished. Hamburg's passages 17 (405 m) → 41 (1 224 m), a 199 m opening at the Hauptbahnhof among them. Line jumps Dresden 12 → 16 (budget 20), Hamburg 62 → 29, Leipzig 0 → 2, München 5 → 4. Paved track (paved, rail_osm.Paving): embedded, or a service track both of whose sides are paving (a quay, a factory yard): no ballast, no sleepers, the rail head 2 cm up (Hamburg's harbour spurs had stood on 4.5 m of raised ballast). Dresden 6.8 km, Hamburg 3.7, München 0.17, Unna 0.08. Heights: the rail head 0.17 m over the DGM (it holds the ballast's surface) and over a rail deck's top (were 0.34 and 0.08 m); the ballast 0.03 m up (was 0.18); a tram's ballast rail top 0.25 → 0.17 m (lib/city/tram.ts RAIL_TOP_M). Sleepers (app/_components/track-bed.ts, one sceneMaterial): a ribbon per track, a concrete sleeper every 0.6 m drawn by its fragment pass, timber over dark gaps on a truss's open floor; the ballast and sleepers run on across a rail deck, its top painted as ballast between 1 m walkways (bridge-surface.ts railTop). The mesh grows: Dresden Hbf's tiles 341k → 534k rail vertices, 57k → 88k sleeper vertices.

  • Main-line catenary ≈assumed (2026-10-21; not yet judged on a GPU) — over every track with a contact line (electrified 1), drawn by app/_components/catenary-layer.ts from lib/city/catenary.ts: the contact wire 5.5 m over the rail head, staggered ±0.3 m; the messenger 1.4 m above it, sagging ≤ 0.9 m (1 m on a long span); droppers every 8 m; supports where a track crosses a site-wide 50 m grid line (a seam moves no mast); three or more parallel tracks within 7 m on one level share a portal, else a mast 3.2 m out with a cantilever. A mapped power=catenary_mast within 12 m of the track (m) replaces the grid's supports within 30 m of it (Dresden 215, München 83); masts within 15 m of a tram track are left to the tram layer. Under a bridge no support within 10 m and the wires drop under the soffit (the contact wire at least 4.95 m over the rail head). Hamburg's S- and U-Bahn take their current from a third rail: no catenary. Dresden Hbf's tiles carry 94k wire and 24k mast vertices — a phone's memory still to check.

  • Ballast yards ❝taken =computed — Basis-DLM ver03_f (railway AREA, OBJART=42010), dissolved with shapely union_all(make_valid()) in the bake and clipped to the tile (~5 non-overlapping parts) → one merged surface, so dozens of yard tracks can't z-fight. Only where a track's bed lies on it (2026-10-20): the parcel reaches the railway's fences and holds its depots, their yards' tarmac and parking, the embankments' grass and the station forecourts — on the orthophotos half of Dresden's (49 %), two thirds of Leipzig's and three quarters of Meißen's and Grimma's lay off every track, all of it drawn as raised gravel (Leipzig's bus depot east of the Hauptbahnhof, Dresden's Wiener Platz). So the parcel is cut to the beds of the baked rails (by their track count) and of OSM's every track above ground and off a bridge (which maps a yard's sidings the DLM bundles), at rail_osm.py's bed widths (4.5 m, 4 m more per further track), and gaps up to 4 m between two beds closed (a yard's tracks lie 4.5–6 m apart; open, each would stand as its own strip with its own fascia). Ballast then: Dresden 165 → 80 ha, Leipzig 76 → 27, Munich 31 → 14, Unna 8.5 → 4.6, Meißen 2.0 → 0.5. Since 2026-10-21 the beds are the baked tracks' own (rail.py ballast(tile, lines), rail_osm.beds, replacing osm_beds): none on a bridge or a paved track. A track neither the DLM nor OSM maps loses its ballast; outside the parcel no bed is drawn. Draped on the ground (2026-10-01, ADR 0041, addDrapedArea): the outline and its holes triangulated, each triangle split in four while the ground at its edge midpoints and centre strays from its plane by > 1 m (and always above 48 m), down to 6 m, every corner on the ground + BALLAST_RAISE; the fascia (0.45 m) on the ground every 6 m; left out on the rail decks (the tracks ride the deck). ~5·10⁴ triangles and ~0.3 s a tile with a big yard (750 before). (Clamped at the outline's corners alone, the yard spanned whatever lay between them: across the underpass at Dresden's Bahnhof Mitte its sheets sloped from the embankment down to the street corners under the deck — the passage read as a closed wall, from the tracks too.) polygonOffset. The recoloured class-5 splat sits underneath so any gap reads as ballast, not seam.

  • Steel rails ❝taken ◎detected — since 2026-10-21 OSM's every track (Tracks one by one from OSM above), and the Basis-DLM's lines only where OSM maps none within 6 m. Basis-DLM ver03_l, heavy rail only (SPW=1000; the DLM carries trams poorly — SPW=3000/BKT=1201 — so they come from OSM, see Trams below; a line whose Bahnkategorie is tram alone is skipped even at standard gauge, which is how Bavaria files Munich's trams). Not underground: the DLM line has nothing on it to say a subway or an S-Bahn trunk line is in a tunnel, so a stretch that runs within 2 m of a DLM tunnel (ver06_l/ver06_f, BWF=1870 "Tunnel, Unterführung") for more than 15 m is cut out; a shorter overlap is a surface track crossing over one and stays. In Munich this removed 23 km of U-Bahn and the 4.2 km S-Bahn trunk line under the Marienplatz and 0 m of mainline (the rail layer shrank from ≈ 84 km to ≈ 22 km, the trams now drawn by the tram layer alone); it applies to every DLM site (Leipzig's City-Tunnel). (A 6 m reach took Munich Hauptbahnhof's surface tracks, which run straight over the S-Bahn tunnel — 🗃️ below.) Short ATKIS fragments are snap-merged by shared endpoints (1 m) in the bake (≈91→9 lines/tile); at runtime a polyline is split into runs of valid ground (never bridged across a NoData gap) and each track gets a thin rail pair (±GAUGE/2, count from GLS) with a small web, on the level its line runs on (Line levels below): the ground, a rail bridge's deck — riding it with no ballast stacked on top, at the deck's height there (the per-vertex deck profile interpolated along the deck's long axis; until plan 024 the deck's mean), from a lift table of every deck kind that the tram layer shares — or a span over a gap, on a deck of its own. Railway class recoloured dusty-mauve → ballast warm-grey (class 5 in lib/city/landcover.ts).

  • Line levels ◎detected (2026-10-01, ADR 0041) — which level a rail or tram line runs on is decided along the whole line, never per sample (lib/city/levels.ts): the candidates per 4 m (rail) or 2 m (tram) sample are the ground and every deck the line may ride whose outline holds the point or runs within 2 m (a rail only rail decks, a tram any); a Viterbi picks the path the line can drive — the climb beyond its grade (rail 4 %, tram 8 %, 0.35 m slack) is the cost, OSM's bridge tag a 4 m-of-jump preference between two levels as good. Then a grade cone: a sample the line cannot get down to from either side by ≥ 1 m is in a gap — a straight span between its rims (pushed out over the gap's sloped sides), a deck within 1 m of it riding along, ≤ 400 m; one it cannot get up to from either side is in a fill — a straight cut, ≤ 120 m. The build step solves every line of the site with 450 m of context past each end of a tile's piece (the line that runs on through that end in another piece, a switch's straightest branch; lib/city/line-levels.ts, scripts/line-levels.ts) and publishes each line's runs off the ground as lv in the rail and tram files (never committed); a layer without them solves its own piece. Under a span the layer draws a span deck: a slab as wide as the tracks + 0.9 m (a tram's 1.6 m either side), its top in the deck colour, 1.1 m deep, a pier every 26 m where it clears the ground by 2.5 m, drawn by the tile owning its middle (addSpanDeck). Dresden: steps beyond the grade by

    1 m in the drawn lines 230 → 18, Hamburg 330 → 62, Leipzig 41 → 17, Unna 20 → 0 (bun scripts/line-levels-cli.ts <site>; the test holds Dresden to 20). What remains: deck ends above lower ground for good, and deck tops DOM1 put 1.5–2 m off their embankment. (Until then: lifted onto every deck under a point — 🗃️ below.)

  • Line profiles and beds ◎detected (2026-10-24, ADR 0041) — on the ground a line no longer follows the DGM sample by sample: the build step fits each line a profile (Gradiente) within its grade (rail 4 %, tram 8 %) and a vertical curve of at least 2 000 m (rail, EBO/Ril 800.0110) or 500 m (tram, BOStrab) — a robust (Huber 0.1 m) least-squares fit to its levels, a deck's top held — with 150 m of context either side (lib/city/line-profile.ts), published as lz. A line on the ground crossing another a storey (≥ 3 m) below it at ≥ 20° is a span there (the DGM kept its bridge), and so is one standing ≥ 2 m over its own ground beside a line as far below (a viaduct the DGM lacks). The terrain bake lays the ground under every line's bed at its profile (ballast's width 0.1 m under it, a street track's 3 cm), the ground beside it graded at 1 : 1.5 for ≤ 4 m, the nearest bed winning, a cut's only as a trench ≤ 0.5 m deep, never a fill under a deck (lib/city/track-beds.ts, shapeDgm); the fine TIN holds 3 cm under a bed (0.15 m elsewhere). The yard's ballast is draped on the beds' profile to 2 cm. Measured over the seven sites in the PR; Dresden's sleepers under the ballast 14.3 % → 0.05 %, its rail heads under ballast or ground 1.84 % → 0.01 %, its street tram rails in the road 2.6 % → 0.02 %. And nothing lies on a track (2026-10-25): a platform outline is cut back to its edge distance from every drawn track (1.65 m rail, EBO §13; 1.35 m tram; lib/city/platform-clearance.ts clearOfTracks) — OSM's island at Dresden Hbf stood 0.6 m from a DLM track's centre and covered its rails (Dresden 3.5 %, Leipzig 3 %, Munich 1.3 % of the platforms' area cut); a Basis-DLM centreline's stretches over an OSM platform are dropped (outsideAreas, railFeatures: not a track, the axis between a station's tracks; Dresden 503 m, Leipzig 689 m, Munich 55 m); and a flat LoD2 top less than 1.5 m over a track's profile where the track runs on it for 20 m is lowered to the bed before the buildings bake (lib/city/track-formation.ts onFormations: the formation the track lies on — ALKIS 31001_9998 west of Dresden Hbf, 0.4 m over the rails for a kilometre of track; in Hamburg a station building (ALKIS 31001_3091) a track runs on, 1.2 m; a roof on posts — ALKIS Überdachung, drawn open — is never one, the track runs under it).

  • Funiculars and monorails ❝taken =computed ≈assumed (2026-10-31, ADR 0041) — OSM's railway=funicular is a track like a railway's, marked steep (rail_osm.rails; no catenary, a cable hauls it): its level and profile are solved at 100 % grade with vertical curves of ≥ 100 m (STEEP_GRADE, STEEP_RADIUS), its context only through other steep pieces, and a deck it runs on is a rail deck measured within 100 % grade — Dresden's Standseilbahn climbs 17–29 % on the ground and over its Gerüstviadukt (until then a path deck, its line flattened), Meißen's Panoramaaufzug 65 %. OSM's railway=monorail is no track but a beam (rail_osm.monorails, in the rails' file as mono with bz): its underside is the upper hull of the DGM + 5.5 m for cars hanging under it (monorail=hanging), 4.5 m for cars on top, over the whole chain — straight across a hollow, bent over a crest; the rail layer draws a closed girder (2.6 m wide for two tracks, 1.4 m deep) and every 8 m a steel portal (legs 1.5 m outside the cars' way, a crossbeam on the girder) or, for a beam ridden on top, a column, each foot SINK.box under the lowest ground at its corners and reporting its joins (monorailJoins, scripts/ground-joins.ts: 32 portals, no miss); none over unknown ground or a leg under 2 m (lib/city/monorail.ts, monorail-beam.ts). The hull spans the whole chain, right for a short line (the Schwebebahn's 274 m, the sites' only one); a long line over hills would need its spans limited. In the model Dresden's Schwebebahn runs up to 18 m over its hollow. The clearance, the section and the supports' spacing are assumed; no cars (they would move, ADR 0009). Before: both missing, as neither railway kind was read.

  • Passages under decks ◎detected (2026-10-01, ADR 0041) — a cut lying ≥ 30 % under a drawn deck (any kind, within 2 m) opens the fill in the terrain bake (lib/city/passages.ts, shapeDgm's last pass, both levels): along the line, half-width 2 m per track + 1.5 m, the ground lowered to the line's level − 0.15 m, never raised, its sides where the DGM stood. Where the DGM fills the ground under a bridge (Hamburg's keeps its bridges) the track ran into a hill; a cut under no deck stays closed (a tunnel, a station's cover, a road the bridge data lacks). Hamburg: 15 passages, Munich 1, Dresden none (its tracks pass under their decks on open ground). The fill under a deck that no line passes through stays as measured — under Dresden's Bahnhof Mitte the raised third of the outline is the abutment the DLM outline includes, and the passage beside it reads closed from an oblique view. And the ways under a deck (2026-10-21, B10): every OSM highway (not on a bridge, not in a tunnel, no steps) the DGM closes under a drawn deck (within 2 m of the deck's top) is opened along the way, straight between the lowest open ground either side, where the DGM stands ≥ 0.5 m over that level, with ≥ 2.2 m headroom and ≤ 12 % grade; half-width OSM width/2, else lanes × 1.6 m, a footway 1.25 m, else the class's, + 0.5 m (bridge.py way_passages, rail_osm.passage_half). Where a Land's DGM holds the bridge itself (a hump ≥ 2 m over level ground 2–6 m beside the deck on both sides, or over level water on one side where water runs on under it; ≥ 4 stations of 2 m, not an end alone) the ground under the deck is opened across it, edge to edge, to the lower side's level (hump_passages). Baked as the deck's under, opened by the terrain bake to their own level (no 0.15 m floor) with square ends (lib/city/passages.ts deckPassages; scripts/line-levels.ts appends them to the site's passages). Hamburg's 32564_5932: 2 ways and 16 hump stations, its 32566_5934 2 + 11; Dresden's 33410_5656 1 + 4; München's 32692_5334 1 + 29, 32690_5332 0 + 9.

  • Bridge axis ❝taken =computed — every deck is measured and drawn along a centreline (bridge.py deck_axis, lib/city/bridge.ts axisFrame): the DLM bridge line where there is one, clipped to the outline and run on to its ends (a polyline — the Marienbrücke's road deck curves); else the long axis of the outline's minimum rotated rectangle, through its middle. (The first cut took the outline's two farthest-apart vertices — the ends of a diagonal: across a deck 11 m wide the offsets ran skewed by up to half its width, the Blaues Wunder's truss was measured at −5.5/+3.0 m and drawn beside the deck at its ends.)

  • Bridge decks ❝taken ◎detected — driven by the complete ver06_l (BWF=1800) centreline set (carries every road/rail/path bridge + NAM), each snapped to a clean ver06_f deck AREA footprint the line runs on (half its length within 1 m of it) else buffered as wide as DOM1 sees its deck (Deck width below; the mapped width, else the kind's, where it cannot) — so the parallel Marienbrücke rail + road decks are separate single-volume slabs (top face + one continuous fascia) AND road/path bridges without an area polygon still render. (The match used to be the nearest footprint centre within 50 m: the Marienbrücke's road line, 16 m beside the rail bridge's footprint, took that footprint — the road was laid on the tracks, the tracks got no rail deck and the road bridge was not drawn at all.) Deck height = the roadway DOM1 measures: per 1 m station along the deck's axis the lower third of the surface across the deck (parapets, cars and lamps stand above it), held within −6/+4 m of the abutment ramp from DGM1 (a train or a canopy over the deck is not the deck), smoothed over 15 m and then grade-limited from the middle of the span outwards — ≤ 8 % on a road or path deck, ≤ 4 % on a rail deck — so where the scan loses the deck near an abutment it ramps down instead of dropping off a cliff (the rail bridge by Neustadt station fell 5 m at one end) (bridge.py measured_deck, limit_grade). What comes out is then reduced to the simplest deck it can be (camber_fit): a straight ramp between its two ends plus one upward camber fitted to it, never sagging, no end steeper than the grade — the raster's dips (the river seen through a gap) and waves are not the roadway. (Without it the Waldschlößchenbrücke and the Blaues Wunder bowed inwards.) Without DOM1: the ramp + camber. The ground is a mosaic of the tile and its neighbours (700 m around it), so a deck across a seam comes out the same in both tiles' files, and only the tile that owns its centre draws it (every copy still lifts that tile's rails). Closed from below (the soffit: a deck seen from under it was hollow). Flush parapet walls (no floating cap) along the sides of a stone bridge, an open steel railing on every other (Bridge parapets below) — not where a frame stands on the edge: the frame is the railing. Kind (rail/road/path) from the network that runs on the deck (2026-10-20, Ground.classify, Ground.carried in rail.py): of the DLM's rail (heavy rail, its trams aside — Bavaria files them as standard-gauge railways), road and path lines within 8 m (paths 5 m) of the axis, the one with the most metres following it within 30° (along), rail before road before path on a tie; a footprint no centreline runs on carries the network whose lines leave it onto ground within 2 m of the deck's top, 4 m past its edge, and its axis follows that network's longest line across it. Where nothing runs along or leaves at deck height, the networks under its points decide, rail first. (Counted point by point, the tracks under a road or footbridge made it a rail deck — Leipzig's Bennigsenbrücke, Munich's Reichenbach- and Maximiliansbrücke with their trams —, and a footprint's long axis ran along the road under a wide railway bridge, so the rule by direction alone turned Dresden's Weißeritz rail bridges into road decks. Against OSM's bridge ways, six cities: 215 of 241 decks right before, 229 of 239 after; what is left is road against path — a 3 m service way the DLM files as a path.) (Tried ver06_f-only — it dropped the road/path bridges, which lack area polygons; the DGM ramp alone — it put the Waldschlößchenbrücke 3 m low and rail decks up to 6 m high; per-tile ground — seam bridges drawn twice at two heights; see 🗃️.) The deck on its measured line =computed (2026-10-20, bridge-layer.ts onLine, splitDeckTop; lib/city/decks.ts deckLine): the bake writes the deck's top every 2 m along its axis (line), but the viewer drew it from the ring's few corners — a long outline has four to a dozen — as flat triangles between them, which cut a chord under the camber, and a ring vertex took its own height. So the ring's edges are now divided into pieces of at most 4.8 m, every corner at the line's height at its station (the bake's own corner heights lie up to 12 cm off it, 3 cm at the 95th percentile over the seven cities' 1 240 corners: kept, the fascia met the slab's top with a step there; since 2026-10-22); the top and the soffit, split into triangles of at most 6 m, take the line's height at every vertex; the parapets, the approaches and the masonry run along the same corners. The lift table that stands rails, trams, lamps and benches on a deck read its profile from the ring's corners too: it now reads the line. Before, the drawn top missed the measured one by up to 4.39 m (the Waldschlößchenbrücke, mid-span), 3.08 m (the Albertbrücke), 2.47 m (the Waltherstraßenbrücke), 1.63 m (Meißen's Elbtalbrücke), 1.55 m (Hamburg's Nordkanalbrücke), 1.38 m (Munich's Reichenbachbrücke), 1.25 m (Leipzig's Hermann-Liebmann-Brücke) — 64 of the six cities' 491 road and path decks by more than 0.25 m, so a lamp or bench set on the measured top floated over the drawn one or stood in it —, and the lift table by up to 2.38 m (Hamburg's Güterumgehungsbahn) and 1.14 m (Dresden's Hafenbrücke). Now every drawn deck lies within 2 cm of its line (Hamburg's within 8 cm), the lift table within 7 cm. Approaches (Rampen, 2026-09-30): the deck's ends are measured in DOM1, the terrain is the bare-earth DGM, which runs on under an approach as the ground it was built on — every deck end stood as a slab edge a metre or more above the road running up to it. So from each abutment end (the ring's edges across the axis) that stands ≥ 0.5 m above the ground, a ramp runs out at the kind's grade (road 8 %, path 10 %, rail 3 %) to where it meets the ground, column by column across the end (≤ 2 m apart), within 40 m (lib/city/bridge.ts approachLanding, bridge-deck.ts approaches): the deck's own top colour, the fascia's sides and soffit as deep as the deck (Deck depth below; 1.1 m until 2026-10-21). None where the deck runs on onto another deck, where most of the end finds no ground in reach (a deck over a street, a viaduct onto an embankment the DGM has), or where the end is flush. The approaches join the lift table: a line on the ground (Line levels above) rides one that stands less than 1.5 m over the ground under it (not a street passing below), a rail a rail deck's.

  • Deck width ◎detected (2026-10-21, photo QA V10) — a bridge line on no footprint (DLM ver06_l, OSM ways) is buffered as wide as DOM1 sees its deck (bridge.py measured_halves, rail.py band): every 2 m over the middle half of the line, a cross-section out from the line to where the surface drops 1.2 m under the deck line and then ≥ 3 m within 2.5 m (the deck's side, the water or the street below), up to 22 m; another bridge's footprint or a point nearer another bridge line ends a side first (Beside: twin bridges with no gap); the median per side, a side DOM1 lost mirrored, the total 3–40 m (deck_halves). A side seen in under 30 % of the sections is not measured: the mapped width (DLM BRF of the road + 2 m of pavement per side, OSM width, lanes) or the kind's. (The kind's 11 m drew the Elbe bridges at half their width: Marienbrücke 20.5 m, Albertbrücke 22.5, Augustusbrücke 19.5, Maximiliansbrücke 23.8.) Now the Marienbrücke's road deck 11 → 20.4 m, the Augustusbrücke 11 → 19.4, the Maximiliansbrücke 23.5, the Reichenbachbrücke 24.8, Meißen's Elbtalbrücke 21.9, the Pöppelmannbrücke 9.5; deck area Dresden's 33410_5656 40 340 → 51 059 m², Hamburg's 32564_5932 53 699 → 63 875, München's 32692_5334 11 017 → 16 620, Meißen 13 451 → 19 078. Where DOM1 sees through a deck (Hamburg's U3 viaduct) the OSM outline stays, about 30 m.

  • Decks split by network =computed (2026-10-21, V11) — a footprint or a merged OSM deck that carries rails and a road (or path) is cut into one deck per network: a Voronoi of points every 2 m along each network's lines, its cells joined by network (rail.py split_nearest, shared_decks), each share ≥ 20 m² a deck of that kind along its own lines. (The first line's kind took the whole outline: the Lombardsbrücke's road lay under ballast, both Oberhafenbrücken were railway.) The Lombardsbrücke is now a rail deck of 18.5 m and road decks of 23.7 and 5.8 m.

  • Bridge superstructure ◎detected (Oberbau: truss, pylons, steel arch) — DOM1 above the deck line, per half of the cross-section (the outline grown by 3 m): the highest surface per station, a morphological opening (7 m: lamps, poles, cars go) and closing (11 m: gaps the 1 m raster leaves in a slender member fill), kept where it rises ≥ 3 m for ≥ 25 m and peaks ≥ 5 m. Each half gives one rib (lateral offset + rise per 2 m station; two within 3 m merge). Only a bridge whose class stands above its deck (arch, truss, suspension, cantilever, cable-stayed — Wikidata over OSM) keeps its ribs: over a beam bridge DOM1 sees catenary, trains and trees. At runtime (bridge-layer.ts, lib/city/bridge.ts): on an arch bridge the tightest rib that fits a parabola (rms ≤ 1.5 m, sag ≥ 2 m, crown inside the run) gives its curve to every rib that follows it (the raster sees one side of the Waldschlößchenbrücke patchily), drawn as a steel arch extended below the deck to where it meets the ground (the springing the DOM cannot see), hangers above and posts below the deck; a rib that follows no arch is not drawn (the Marienbrücke's). On a cable-stayed bridge the peak is a pylon on a river pier with stays fanned to the deck by rule (the Molenbrücke). Otherwise the rib is drawn as an open frame on the deck edge in a simple form (ribProfile): where it peaks ≥ 10 m and ≥ 4 m above the dip to anything higher, a tower on a river pier with a portal between the two frames; a straight chord from the deck at the first run's start up to the first tower, a curve sagging between two towers to the lowest the rib was measured there but no lower than 2.6 m (lower, it lay along the parapet as a second thin strip), a straight chord down to the deck at the last run's end; without towers a level girder at the run's median rise. One 1.6 m chord, a post every 10 m, no diagonals — the Blaues Wunder's silhouette at the abstraction of the clay buildings. (A solid fin, tried first, read as a dark tent; the measured rise itself, smoothed, as waves, its side arms ending in mid-air where the raster lost the chord.) The ribs' measured offsets are not trusted across the deck (the DLM centreline can sit metres off the bridge: the Blaues Wunder's came out −5.5 and +3.0 m around an 11 m deck, a truss through the roadway) — two ribs stand on the two deck edges, their outer face flush with the deck's side (towers included), a lone rib near the axis stays central (placeRibs). Pale weathered steel 0xd9dde0, matte; hangers and stays every 16 / 12 m. (Drawn member by member as measured until 2026-09 — chord, a post every 6 m, diagonals in dark 0x8d9ca8: busier and darker than anything else in the scene.) No DOM1: no ribs. pipeline/bake/bridge.py (ADR 0041).

  • Bridge depth and fairway ❝taken =computed ≈assumed — OSM's inland-waterway marks (seamark:type=bridge + seamark:bridge:clearance_height, ODbL) within 30 m of a deck: the navigation clearance over the DGM's water surface gives the soffit there, hence the deck's structural depth (deck − water − clearance, clamped 0.6–5 m), and where the fairway crosses. Beam-bridge piers keep the fairway clear (the main span wide when Wikidata knows it, else 40 m). The clearance's reference level is not NHN-anchored in OSM, so the DGM's water surface at scan time stands in — which is why that depth is now only a cap on the structure's (below).

  • Deck depth ≈assumed (2026-10-21, V16) — by what carries the deck (lib/city/bridge.ts deckCarrier, deckDepth): a beam a twentieth of its span (Wikidata's main span, else the median gap between its supports, beamSpan), 0.8–2.5 m; an arch's deck 1 m; a truss's or cables' floor 1.2 m. The fairway's depth only caps it; the approaches are as deep as their deck. (Taken as the depth, it hit the 5 m clamp on the Meißen bridges, Dresden's Hafen- and Flügelwegbrücke, Hamburg's U3 viaduct: the clearance is against mean high water, not the water the scan saw.) Decks at the 5 m clamp: Dresden 3 → 0, Meißen 3 → 0, Hamburg 6 → 0; the U3 viaduct 5.0 → 0.8 m. At runtime, no re-bake.

  • Bridge type and main span from Wikidata ❝taken ◎detected (CC0) — the ingest fetches the bridges Wikidata knows around each tile (class, main span P2787, length, and since 2026-10-21 the number of spans P1314 — the cache's "version": 2, an older one is fetched again) into data/_raw/<provider>/wikidata/; a deck is matched by an OSM wikidata tag on its outline or fairway mark, else by its DLM name within 150 m. Wikidata's classes (Bogen-, Hänge-, Ketten-, Gerberträger-, Fachwerk-, Balkenbrücke …) override OSM's bridge:structure — OSM calls the Waldschlößchenbrücke a truss; it is an arch. No file: the OSM tag stays.

  • Bridge arches ≈assumed — structure containing arch without a measured steel arch (Augustus-, Albert-, Marienbrücke) → addMasonry: the axis in bays (2026-10-21, V13, masonryBays, masonryArches) from DOM1's piers where the deck bulges over them (bridge.py bastion_piers: the deck's width every metre ≥ 1.5 m over its 41 m running median for 3–16 m, ≥ 8 m from an end, ≥ 10 m apart; kept where Wikidata's span count agrees, piers, and no bay they leave is wider than 70 m or the main span and a fifth), else Wikidata's number of spans (P1314, spans) with its main span (P2787) centred on the fairway, else equal bays of ~26 m; piers a seventh of their span thick (1.5–7 m); each arch a segmental arc rising a fifth of its clear span to 0.2 m under the deck's soffit, springing ≥ 0.5 m over the ground. (Equal 26 m bays drew the Luitpoldbrücke's one 62 m span as three, the Augustusbrücke's 43 m arches as 26 m ones on 3.2 m piers — springing a quarter of the clearance above the higher ground and rising to just under the deck.) The Augustusbrücke now stands on 7 piers 35–46 m apart, the Luitpoldbrücke on bays of 8, 62 and 8 m, the Pöppelmannbrücke on 5 bays. The bastion test is a heuristic: on the Marienbrücke's road deck it finds 14 piers with three 18 m gaps, on the Lombardsbrücke one (bays of 15 and 47 m). The deck's own side edges carried down as spandrel walls (both faces) to the arch, or to the pier top over a pier, and a vault across the deck under each arch — closed from below and from the side, a masonry-viaduct read. A bay whose arch would rise less than 2 m gets none (the banks). (The first cut sprang every arch half a metre over the water from a point, on 0.5 m slabs that read as a second deck under the bridge, and had no vault: from below or through an arch the bridge was hollow.) beam/absent → flat soffit + box piers, or a viaduct's columns (Viaduct columns below). (Until 2026-09 the walls stood on a straight line between the deck's two farthest-apart corners, one deck width apart and at the lowest deck height: on a long deck they stood metres beside it, on a curved one off it entirely, with water showing between wall and deck.) The structure comes from Wikidata when it knows the bridge, else from the nearest OSM man_made=bridge bridge:structure (≤60 m), else from the OSM ways on the deck (half their length on it): their bridge:structure, else an arch when one was built before 1850 (start_date; 2026-10-20, rail.bridge_way) — iron came after, and a bridge that old is masonry, so a measured superstructure is not looked for. (Meißen's Schloßbrücke, 1221, a stone arch over the castle's moat, was a flat slab: neither Wikidata nor an OSM outline knows it, and its way says only its date.) A deck the Basis-DLM leaves unnamed takes the name of its OSM ways — their bridge:name, else a name that names a bridge (…brücke, …steg, …viadukt; a way on a bridge is as often named for its street) — and with it a Wikidata match by name: on Meißen's tile 6 of its 11 unnamed decks.

  • Viaduct columns ≈assumed (2026-10-21, V17) — a railway beam deck longer than 100 m stands on paired 0.6 m columns under its edges every ≤ 13 m with a crossbeam under the girder (lib/city/bridge.ts beamSupports, bridge-layer.ts addColumnPair; Hamburg's U3 runs on them every 10–15 m); over the stretches the class raster calls water (water, baked) a pier on the axis at every other station (≤ 26 m), and a fairway it crosses kept open between two piers (fairwayOpening). Every other beam deck: piers on the axis ≤ 26 m apart (none up to 40 m), the fairway kept clear (pierStations). Viaducts: Hamburg 5, München 3, Dresden 1; the U3 on 50 column pairs.

  • Bridge parapets ≈assumed (2026-10-21, V18) — arch, humpback or stone bridges (no measured steel arch) keep the 0.85 m ashlar wall; every other gets an open steel railing: a 0.2 × 0.3 m kerb flush with the fascia, posts every 2.5 m and a top rail 1.1 m over the deck (lib/city/bridge.ts parapetStyle, bridge-layer.ts addRailing); none where a frame stands on the edge. Railing against ashlar: Dresden 117 / 19, Hamburg 305 / 43, München 44 / 16, Leipzig 38 / 3, Meißen 13 / 4, Unna 10 / 1, Grimma 4 / 2.

  • Station platforms ❝taken ≈assumed — OSM railway=platform (ODbL) → triangulated flat slabs (ShapeUtils.triangulateShape). On the level of the track beside them (2026-10-21, lib/city/platform-level.ts): the edge OSM's height (h), else by the trains it serves (k) — 0.76 m a main line's, 0.96 m an S-Bahn's, 1.0 m a U-Bahn's, 0.55 m a narrow gauge's, 0.25 m a tram's — over the rail head of the nearest drawn track's solved level (ground, deck or span), never under the DGM inside its outline. (Drawn 0.55 m over the DGM, they stood too low where the DGM holds the platform, and on the street under Hamburg's U3.) Hamburg's U3 platform top now at 13.7 m over the ground's 5.4 m, Dresden Hbf's at 120.6 m over the street's 113.9 m. The OSM half of the blend (Basis-DLM has no platform geometry); absent/empty when the site has no .osm.pbf extract. A platform below ground (in a tunnel, location=underground, or on a negative layer or level: osm.below_ground) is skipped — Munich's U-Bahn platforms under the Marienplatz and the Odeonsplatz had been drawn at street level.

  • Trams ❝taken ◎detected ≈assumed (plan 024) — OSM railway=tram (ODbL; 401 ways, 59.6 km in the four first tiles — 146.8 km of track drawn over the fifteen —, every one gauge=1450, electrified=contact_line) + power=catenary_mast (321 points on the four first tiles, 167 of them within 15 m of a tram track — the rest are the railway's) + the OSM building outlines. pipeline/bake/tram.py → tram_<tile>.geojson: each track (fragments chained at 1 m, cut at the tile edge) with its bed, read every 2 m — street on the road class of the committed class raster and on the paved squares and yards (the built-up and path classes), grass (Rasengleis) on the meadow class or NDVI > 0.3, else ballast, a majority over 20 m, no stretch shorter than 20 m — the track cut where it changes; and the OSM bridge/layer (per-tile numbers in the table below). (Until 2026-09-30 one bed per whole chain: street only when ≥ 70 % of it lay on the road class, so a kilometre of street track went to ballast for a lawn or a square under a third of it — the Postplatz's tracks lay half in a brown ballast strip, half in the paving. 23.8 km of ballast became 3.2 km; the committed files were re-cut by the same rule from their own track lines, the class raster and the NDVI, without a new OSM extract — the table below is the bake before that.) A way in a tunnel or tagged location=underground (osm.in_tunnel) is skipped: a tram tunnel is not drawn at street level. At the Albertplatz (150 m round) 87 % street / 5 % grass, the Hauptstraße 100 % street — the plan's > 10 % STOP was not hit. The supports are decided on the tracks and masts within 30 m around the tile (so a seam support is the same in both tiles): a span from a mast across the tracks to the nearest mast on the other side (≤ 28 m, each mast in one pair, shortest first), else an arm from the mast over the nearest track (≤ 10 m; over a parallel second track too); where a track is 45 m from any mapped mast, a rosette span every 30 m between the facades either side (the OSM building outlines, ≤ 15 m out; none when either side has none — the plan asked for a ray against the LoD2 BVH at runtime; the bake reads the OSM outlines instead, because a dressing cannot count on its neighbour's buildings being loaded, and the result must not depend on load order). Each track carries s, the distances at which a span or arm holds its wire. Gauge (2026-10-01): each track carries its OSM gauge in metres as g (tram.py gauge_of, 600–1700 mm accepted, else none; fragments of different gauges are not chained) — Dresden 1450, Leipzig 1458, Munich 1435 mm; an untagged track gets standard gauge (TRAM_GAUGE = 1.435 m in lib/city/tram.ts, Dresden's 1.45 until then). Runtime app/_components/tram-layer.ts: two rails per track at ± half its gauge (addTrack; Dresden's ±0.725 m) with the rail layer's profile (addRibbon), in the road's lavender-grey a shade deeper — street: the heads flush with the road (+2 cm), no sleepers, and no groove (the plan's darker groove strip was left out: fewer, calmer lines); grass: rails +15 cm over a 2.6 m meadow strip (two grass tracks ≤ 4.5 m apart share one lawn, the inner edges at the midpoint, lawnHalves, 2026-10-21, V23: 67 % of Dresden's 13.3 km of grass track, 98 % of Leipzig's 3.7 km, 67 % of Munich's 2.6 km); ballast: rails +17 cm over a 2.8 m ballast strip (+25 cm until 2026-10-21, the rail layer's head height). A track the OSM way puts on a bridge rides the deck (the shared lift table, any deck kind — gated on the tag, so a tram under a railway bridge stays on the ground). Contact wire 5.6 m over the rail top, sagging 0.15 m between the supports (lib/city/tram.ts wireStations/wireDrop: both line ends are stations, so the wire meets its neighbour tile at the same height; gaps > 30 m get evenly spaced virtual stations); span wires 7 m up the masts / 6.5 m at the rosettes, lifted to clear the wires they hold by 0.5 m, with a hanger down to each; arms 35 cm over the wire with a stay. All wires are one camera-facing ribbon mesh per tile whose width is max(true width, 0.8 px) in the vertex shader, its alpha the true coverage (≥ 0.2) × 0.6 in a light slate, faded from 150 to 350 m — light and faint, a pencil line rather than ink; no Line2, no MSAA, no new dependency; wires never cast (castShadow = false). Masts: a 7.5 m pale green-grey pole, instanced, casting. Stop signs (phase 3): OSM puts railway=tram_stop on the track (144 of 164 within 0.5 m of it), so the sign stands on the nearest mapped platform (public_transport=platform / railway=platform, ≤ 25 m) at its point nearest the stop, facing the track — the "H" sign of plan 030 (buildFurniture), 107 of them; none within 8 m of a shelter, a bus stop's sign or another tram sign; the 4 stops with no platform mapped get none. Look unverified on a real GPU (the plan's plates at Postplatz/Augustusbrücke, dusk and 150 m fly are open). Stop signs (2026-10-21, V25) never stand on a platform below ground (osm.below_ground) or on one tagged subway/train/bus = yes without tram/light_rail: Dresden 230 → 229; Munich's 9 signs on the U-Bahn platforms under their stops moved to the tram's own. Side and centre poles ≈assumed (2026-10-21, QA V24): supports stand where a track crosses a 30 m grid line along its main axis (grid_stations; both tiles of a seam pick the same places); where no mast is within 45 m and a side has no facade within 15 m, a side pole (k: pole) stands at the street's edge — the first class-raster change beyond the outer track, 0.5 m past the kerb on a road, 0.5 m inside a lawn or paved bed, 3 m beside the track where nothing changes within 14 m (a square, a road so wide the tram has its own bed), never within 1.7 m of a track — a mapped post lamp within 3 m being the pole (lamp: 1; the step reads the lamps files of the tile and its neighbours); spans ≤ 40 m; a centre pole with an arm over each track only where a double track's central strip (its spacing − 2.65 m, the widest car) exceeds 1.5 m. Drawn as the masts are (the same 7.5 m pole). Track more than 25 m from a support: Dresden 78 → 10 %, Leipzig 83 → 8 %, Munich 62 → 8 % (2 644 / 1 357 / 684 poles; Dresden's spans 133 → 1 587). A span from a pole to a facade is still written as a pole's span (its facade end at 7 m, not a rosette's 6.5 m). The tram step is slow: Dresden 12 min, Leipzig 4.6, Munich 7.3. Style (maintainer feedback on the fences, applied before any plate): everything in the scene's soft clay/watercolour idiom — no near-black, no fine detail: rails in the road's lavender-grey a shade deeper, the groove the plan asked for left out, wires a light slate at ≤ 0.6 opacity, masts pale green-grey.

    tiletracksstreet kmgrass kmballast kmmastsspansrosettesarmsstop signs
    33410_56569520.530.764.7311851281357
    33410_56582012.170.540.25331336616
    33412_56565111.340.981.9016727223
    33412_565895.6601.190017011

    (The table was baked when only the road class made a street bed.)

Walls and fences

  • Walls ❝taken ≈assumed (Brühlsche Terrasse &c.) — OSM barrier=retaining_wall|city_wall| wall + man_made=embankment + natural=cliff (kind cliff, default 3 m; read from other_tags, since GDAL has no natural column on lines) (ODbL), with the tagged height (e.g. the 8.5–9 m city walls). pipeline/bake/walls.py → vertical ribbons toned by their material (lib/city/walls.ts), baked into the fine terrain glTF as a walls node (scripts/bake-tiles.ts wallMesh, ADR 0030): base on every tile's fine TIN (so a wall near a seam reads its neighbour's), top on the high shelf; an earth-retaining wall snaps to the step the TIN measures (face at the ramp foot, a coping cap back to where the ground reaches its level — "Terrain TIN" above), any other wall stands on the OSM line nudged slightly onto the low side; wall-layer.ts only gives it its material. A freestanding wall rises its height above the ground (2026-10-20, kinds wall and city_wall, FREESTANDING_M in lib/city/walls.ts): its top at least its tagged (or default) height above the low side, so on level ground it stands from the ground up instead of ending at it, with a body — a cap over its thickness (0.4 m, a city wall 1 m) and a back face to the ground — straddling its OSM line there and moving the usual 1.5 m onto the low side as the step across it grows to 1 m. (The high-shelf rule alone buried every freestanding wall on level ground, its top at the ground: in Dresden 32 km of the 61 km of barrier=wall, of which DOM1 shows four in five standing a metre and more; in Leipzig and Munich half of them too, in Hamburg a quarter. A retaining wall still ends at its high shelf.) The coping cap faces up whichever side its high shelf is on (2026-10-20, pushCap): the walls draw both sides, and three shades a triangle seen from behind with its normal turned over, so a cap wound downwards came out black from above. (OSM draws a retaining wall with its high side to the left; where a way runs the other way, or the measured step lies the other way, the cap was wound against its normal: 628 cap triangles on Meißen's tile, 5.8 % of its wall area, as the black band along the Burgberg's west wall; 99 on Dresden's spawn tile.) Quays are measured where the water meets the land ⚙ derived (2026-10-21, PR #173, walls.py quays, kind quay): every metre along the water's outline in the class raster (the dry shore re-classed first, "Water" below) the DGM is read 0.5–6.5 m onto the land over the water's level 1.5–3.5 m out on it; a station is a quay where the land stands 1.5 m over the water within 2.5 m, its steepest metre climbs 1 m (a revetment's 1:1.5 or a dike's 1:2 does not), three quarters of the climb to 6.5 m is done by then (a wall, not a slope's foot) and the NDVI there is under 0.2 (a grass bank is no wall). A majority over five stations, runs of 8 m or more; then a mapped line takes the stations it draws or keeps open, so no window bridges it: a wall that holds the earth within 7 m (its snap reaches 6 m for the steepest metre, the same step), a freestanding wall within 3 m that rises 70 % of the station's climb (it faces the water itself; a parapet on the quay's top does not) and a flight of steps within 2 m (its foot stays open). The height is the median rise, the material the site's majority, the walls file's credit the provider's beside OSM's. The terrain bake treats a quay as a retaining wall: it snaps to the measured step, its face at the ramp's foot in the water, its coping on the quay. Measured on the committed tiles (2026-10-21, with the vetoes): 22.1 km on Hamburg's four tiles (11.2 km on the Landungsbrücken's, 10.0 km on the Speicherstadt's), 3.5 km in Dresden's Alberthafen and along the Weißeritz, 100–200 m each on Leipzig's, Meißen's, Grimma's and Munich's rivers, a few metres in Unna. (A first bake left the mapped walls on Hamburg's quay edges standing — 1.75 km of barrier=wall measured down to the water — and drew a quay a metre from each, and a quay across the foot of three flights of steps.) (OSM maps few quays — man_made=quay on 5 km of Hamburg's harbour, three lines in Dresden — and the DGM rounds every quay into a 1–2 m ramp: the water class reached up it, and with the shore rule the ramp read as sand whose TIN triangles hung down the face in folds, the saw-teeth along the Landungsbrücken.) Until then the browser built the ribbons from the GeoJSON over whichever terrains were loaded. Why OSM: no elevation product has the wall as a vertical face (a 2.5D surface cannot), and it's not a CityJSON building, so it "went missing". OSM has it as explicit vector lines with heights. (Corrected by the terrain study, "Terrain TIN" above: the source data does NOT smooth the wall into a gentle bank. The laser ground returns step within ~0.75 m (median, 7 tall walls) and the native 1 m DGM1 within ~1.7 m at 77° — Jungfernbastei: 108.6 → 117.3 m over 2 m, then two more terrace levels at 119.7 and 121.1 m. It is the 1024² resample that widened the step to ~2.9 m at 69°, the "bank" the viewer showed. The earlier claim "LiDAR-ground ≈ DGM" was right about the level — the two agree to 0.00 ± 0.05 m tile-wide — but not about the edge.) Source: a LOCAL Geofabrik .osm.pbf read via GDAL's OSM driver (both the lines and multipolygons layers — GDAL files closed barrier ways as polygons), with no Overpass rate limits and reproducibly (the bash-era version of this bake was verified feature-for-feature identical to the old Overpass bake: 436 walls, same kinds/lengths/heights). Since ADR 0037 every OSM layer comes from that extract (pipeline/bake/osm.py). Heights measured, materials toned (2026-10-21, PR #173 B5/B6) — an untagged wall/city_wall is measured every metre along its whole way: the highest DOM1 cell (the laser scan's 0.5 m dsm_050 where every tile read has one) within 1 m of the line, a cell the NDVI calls a crown (≥ 0.45) left out, over the lower ground 1.5 m either side; a station counts where one side lies open (the surface within 0.5 m of the ground 2.5 m out), no OSM building within 1 m, under its class's cap (wall 6, noise barrier and flood wall 8, historic city wall 15 m). With ≥ 3 stations and ≥ 30 % counting, the median is its height; else noise barrier 4 m, flood wall 2.5, historic city wall (historic=citywalls|city_wall|castle_wall, heritage) 6, a plain wall the tile's measured median (≥ 10 walls) or 1.5 m. (Every untagged wall stood 1.5 m and every city wall 6 m; the surface model gives 1.8–3.1 m medians, Hamburg's 57 % under 1.2 m, noise barriers 2.8–4.6 m, Grimma's flood walls 2.3–4.2 m.) A wall under 0.3 m is left to the kerb; a measured low garden wall (0.3–0.8 m) stands (KERB_WALL_M, MIN_H 0.7 m in lib/city/walls.ts; it was dropped under 0.8 m). Material (mat: material, wall=*, a historic city wall stone, else the site's majority of tagged walls, ≥ 8, else concrete — Dresden and Meißen stone, Leipzig brick, the rest concrete) is drawn as the wall's vertex colour (WALL_TONES; a cliff or an embankment keeps the sandstone).
  • Fences, railings and gates ❝taken ≈assumed — OSM barrier=fence|handrail (lines and the outer ring of areas) ride the walls file after the walls, as {kind: "fence", type, h}: fence_type railing/metal/bars → railing, wire/chain_link/chain/temporary → mesh, wood/split_rail/pales → picket, untagged → railing (Dresden's wrought iron), a handrail → rail; height when 0.3–4 m, else 1.2 m (a handrail 1.0 m). Every line is clipped to the tile as a line — an area's outer ring first, then the ring cut at the tile edge; clipping the area as a polygon first had closed its ring along the edge, a ruler-straight fence or wall on the seam that stands nowhere (1.52 km of fence in 10 rings and 216 m of wall in 2, e.g. 328.7 m of railing along x = 412000; now 0 m on any tile edge, guarded by features.test.ts). Last come the gates (barrier=gate|lift_gate|swing_gate|cycle_barrier points within 0.5 m of a wall or fence line, snapped onto it, {kind: "gate", w, on}: width, else 1.2 m, a lift gate 4 m); a neighbour's gate whose gap reaches over the seam comes along as {seam: true} and cuts this tile's piece too (none over the fifteen tiles). 189.3 km of fence and 1 452 gates on a line over the fifteen tiles (1 238 on fences, 214 on walls; 2026-09-26); on the four first 97.7 km and 875 gates (15.5 / 45.0 / 14.7 / 22.5 km on 33410_5656 / 33410_5658 / 33412_5656 / 33412_5658; 746 gates on fences, 129 on walls; about 520 gate points stand on no mapped line and are dropped). The committed walls are older than the gates (below): a gate is kept only on a line of its kind the file itself carries, which dropped two on: "wall" gates 70–73 m from any committed wall (33412_5656, 33412_5658) on 2026-09-26. Baked into the fine terrain glTF as a fences node (lib/city/fences.ts, scripts/bake-tiles.ts fenceMesh, ADR 0030): one low, calm band along the line on the fine TIN — a flat, double-sided quad per ≤ 2.5 m (closer where the ground bends between, ADR 0030: a band bridging a dip floated over it) at ¾ of the tagged height (0.4–1.5 m, 1 m until 2026-10-21; a handrail only its top 15 cm) — cut w wide at each gate with a leaf in a lighter tone in the gap (a boom at the band's top for a lift gate, swing gate or cycle barrier); a closed ring is cut across its closing vertex too (14 gates at a ring's closure had cut only half their gap). Gates on a freestanding wall (kind wall) cut that wall's ribbon the same way. Fences never enter the conflation, the stair burn or the step snap: they stand on their OSM line. fence-layer.ts gives the band one muted tone close to the ground's and the hedges' (a warm grey-sage; wood a warm stone), a breath deeper at the foot, lighter toward the top edge (a see-through impression by colour alone), fading toward the pale ground from 40 to 160 m, and lit as the ground it stands on (its normal the world's up), so no side turns dark; opaque, no discard, no dither; it receives shadows and casts none. 13 506 / 38 612 / 12 734 / 19 032 triangles per tile. The walls file keeps its walls verbatim from the Geofabrik bake (the terrain study addresses them by index) but for the two seam artefacts, cut off exactly as the fixed bake cuts them; the fences and gates come from the BBBike extract of 2026-09-19. Not yet judged on a real GPU (plan 029, plan 019). Plan 029. An untagged fence within 15 m of an above-ground rail track, or along a pitch, school or industrial site (grown 3 m), over half its length, is 2 m, others 1.2 m (2026-10-21, B9); the flights' handrails ride the same bands (Flights of steps below).
    • Discontinued first look (🗃️ below): a procedural panel — bars every 12.5 cm, a wire diamond mesh, pickets, a post every 2.5 m and a top rail, alpha-cut, a dithered veil far off and a dithered partial shadow. On a phone it was "zu hart und kleinteilig" and then "stärker stilisiert, mildere Farbwahl, Kleinteiligkeit führt zu Artefakten" (maintainer, 2026-09-25): dark iron read as ink against the pastel scene, and the fine bars and the dither aliased (moiré, shimmer).

Wall → terrain conflation (breakline burn at build time)

The coarse level only since the fine one became a TIN (ADR 0030); there the wall ribbons snap to the measured step instead (lib/city/wall-snap.ts, "Terrain TIN" above).

  • Stepped ground at walls ◎detected — lib/city/terrain-conflate.ts, applied at build time to the coarse grid (and a fine one whose DGM has holes) before it is meshed (scripts/bake-tiles.ts terrainMesh; it ran in the browser at load until ADR 0024). The DGM blurs a vertical wall into a ramp, so the OSM ribbon used to float over it / get swallowed and the ground never "stepped". The conflation reads the terrain's natural shelf level a short way out on each side of each wall line, then snaps nearby cells toward the high-side level on one side and the low-side level on the other — a sharp step concentrated AT the line, feathering back to the untouched DGM within a ~11 m band (nearest-wall-wins). The wall ribbon then skins a real step. Deterministic + source-portable (any DEM + any OSM wall lines). Gated to earth-retaining kinds (retaining_wall/city_wall/embankment/cliff) and only where the two sides actually differ by ≥1.5 m, so freestanding garden walls and flat fountain rims leave the ground alone; a ≤18 m clamp stops a bad height tag gouging a canyon. Pure + unit-tested (terrain-conflate.test.ts). Measured cost (terrain study, "Terrain TIN"): against held-out laser ground it raises the RMSE in the 20 m band around walls from 0.43 to 0.76 m (p95 0.80 → 1.62 m) — the 11 m probe reads the top of a terraced wall, and the step lands on the OSM line, not the measured edge. Hence the fine level's TIN skips it.

Stairs (OSM steps over a lowered terrain)

  • Flights of steps ❝taken =computed ≈assumed (Freitreppe am Italienischen Dörfchen &c.) — OSM highway=steps (ODbL). pipeline/bake/stairs.py → stairs_<tile>.geojson: the axis oriented bottom → top, w from width (else an area:highway=steps outline: area ÷ axis length; else the gap between the OSM walls either side when the slope fills it, the axis re-centred; else 1.5 m — 2.5 m in a pedestrian area: highway=pedestrian or an area:highway outline within 1 m of its middle; 2026-10-21, it was 2.5 m everywhere), z = the two landing heights from the DGM 1 m beyond each end (3×3 m median; the committed DGMs around the tile mosaicked, so a flight across a seam keeps both landings — 2026-11-04, before that both tiles dropped it), n from step_count when its riser is 8–25 cm, else rise ÷ 16 cm. Indoor, underground, tunnel and bridge flights and anything flatter than 30 cm are left out. Where the DGM lacks the structure a flight climbs (less than half its tagged rise) and its top lands on a raised OSM area (layer ≥ 1), the tagged rise wins and the area goes to terraces_<tile>.geojson at the flight's top level — the Brühlsche Terrasse, 118.3 m over the Schlossplatz's 112.2 m. At build, lib/city/stairs.ts lifts the ground inside each terrace (raiseTerraces, after the wall conflation; holes in the OSM area filled), then burnStairs sets the terrain under each flight to 12 cm below the ramp (on the fine TIN 12 + 15 cm, its tolerance) through the steps' inner corners — lifting it where the DGM runs below, since the player walks on the grid — and lowers every vertex beside it whose triangles reach under it, never across a wall; beside a flight that stands above its surroundings (the ground 3.5 m out below its ramp) it digs the DGM's blurred shoulder — the raster rounds a raised flight's sides into a slope a metre or two wide, which showed as a ragged bank beside the cheeks — down to that ground, up to 2.5 m out. The same build step writes each flight as blocks in its material's tone — treads, darker risers, side cheeks down past the bottom landing; where the bank beside a flight stands ≥ 0.4 m above its treads (a flight cut into a slope, read 3.2 m out on the final ground, never across a wall), a Wange instead: a 3 m block topped at the bank's level, a 0.5 m stone coping along the steps and the rest in the built-up ground's tone, over the ground the burn lowered beside the flight, which had shown as a ragged trench (its outer edge meeting the bank where it ends) — since 2026-10-21 only on a side an OSM wall or building runs along (walled, within 1 m at two of three quarter points); along an open side the bank stays ground (Meißen's Burgberg: grass banks) and a tagged handrail (handrail, handrail:left|right, turned with the axis; not where a fence or handrail is mapped) stands as a rail 0.15 m inside the edge on the nosing line, its top 0.9 m up. The tone follows surface (paving/asphalt concrete, sett stone, metal grid metal, unpaved wood), else the site's majority (mat, STAIR_TONES); a cheek reaches the ground beside it, and a top landing over lower ground drops a face to it (ADR 0030) — into the fine terrain glTF of the tile owning its middle (a stairs node, vertex colours); stair-layer.ts only gives it its material. Why: the DGM1 smooths a staircase into a bank (the flight beside the Italienisches Dörfchen read as a grassy slope), and a terrain built from it cannot hold a 16 cm riser (ADR 0030). Fallback: no .osm.pbf → the step is skipped, the committed file stays; no file → no stairs, the terrain as before. Pure + unit-tested (stairs.test.ts, the bake end-to-end against a synthetic OSM extract in pipeline/tests).

Lighting

  • Soft shadows =computed — PCFShadowMap + raised shadow.radius; terrain castShadow=false; normalBias=0; tight camera-following frustum, whose camera also drives the tile streaming (casters behind the player stay loaded) — on a desktop, at a resolution of its own, not the map's: 128 px, where it still refines the eye-level frustum's fine level but never from the air (SHADOW_STREAM_PX, ADR 0047). A phone's shadow camera streams no tiles at all since 2026-10-10 (shadowTilesStream, plan 071; at its 64 px it had loaded a tile's buildings and coarse ground, tiles in use the memory governor cannot free): at a seam with the sun behind the player, the neighbour's buildings cast no shadow until the view loads that tile; the map has no colour target (RingShadowNode, three r187's depth-only render target). Full recipe and dead-ends in the city-walker skill.
  • Sky-view factor =computed (Himmelslicht, plan 033) — the committed DGM1 with every non-vertical LoD2 surface burned on top (max over its triangles, each on its surface's plane; lowveg.py's one CityJSON walk, lod2_rings) → per ≈2 m cell the horizon angle h within 150 m in 16 azimuths, seen from the bare ground → svf = 1 − mean(sin² h) (svf_<tile>.png, 1024², 8-bit, ≈0.5 MB). The cells under a roof (16–19 % of a tile) take the nearest open cell's value (scipy distance_transform_edt): left at their ≈0, LINEAR filtering and the mipmaps pulled a dark 1–2 m band onto the wall feet and dimmed distant streets (a tile's mean at mip 4 fell 0.03–0.05 below its open mean; now equal). Scales the indirect diffuse only (the hemisphere fill), as the material's aoNode, which three multiplies into the indirect light after the lights: the sun is untouched. Terrain: mix(1, svf, row). Clay facades: the ground's value 2.5 m outside the wall (a margin past the footprint's filled texels), doubled (a vertical face sees at most half the sky; the ground at its foot the wall too) and faded to 1 toward the eaves — a courtyard's ground floor dims, its eaves and every roof do not. The raster is shared by a tile's terrain (both levels) and its buildings, refcounted (tile-rasters.ts); the clay binds a white texel until it lands. Where a LoD2 roof was rebuilt from DOM1 (ADR 0036) the rebuilt parts are burned instead, as the city mesh draws them. The crowns count half (2026-10-21, CROWN_TRANSMISSION 0.5): the crowns the viewer draws (canopy, laser scan, cadastre, the DLM's rows — their committed files, as wide as lib/city/tree-placement.ts draws them) are a half-transparent occluder, blocking half the sky they cover above the solid horizon. Left out, an avenue under its crowns saw 0.73–0.94 of the sky, solid 0.01–0.03; at half, Dresden's carriageways under crowns (12.8 % of them) fall from 0.90 to 0.52, about halved on every site, open ground by 0.04–0.08. The horizon (below) still leaves them out — they cast real shadows inside the frustum, and a horizon angle cannot hold half a crown. Beyond the site the field is Copernicus GLO-30 (as the horizon's, below); without it the site's edge continues level. Default 0.5 — not yet judged on a GPU: the plan's plates (a Neustadt courtyard, the Prager Straße, the Elbwiesen) and the double-darkening check against the contact shadows (GTAO since plan 020) are open; Boden-Verlauf (the clay's 5 m ground darkening) is kept as it was, not retuned against it yet. Fallback: no raster → the light as before. pipeline/bake/skyview.py, app/_components/sky-light.ts, lib/city/skyview.ts.
  • Horizon shade =computed (Ferne Schatten, plan 033, ADR 0009) — the same height field at ≈8 m (the roofs burned at 2 m and max-pooled, so a spire still occludes) → per cell and per 16 azimuths the elevation angle of the skyline in two bands: occluders 80 m–10 km away (the far band, 0–45° in 8 bits: 80–1 500 m from the 8 m field, 1.5–10 km from Copernicus GLO-30 alone on a 31.25 m ring, ring_horizon, 2026-10-21) and 8–80 m away (the near band, 0–90°), eight RGBA layers stacked in one PNG (horizon_<tile>.png, 256² × 16 azimuths × 2, 1.13–1.23 MB; legend horizon_<tile>.json, not served); cells under a roof take the nearest open cell's angles. The terrain interpolates the two azimuths around the sun and cuts the sun's direct light by smoothstep(h − 0.8°, h + 0.8°, elevation), combined with the shadow map by min (never a product: where both see the same occluder it must not darken twice). The bands split the work with the shadow map by the fragment's distance from the frustum's centre (shadowReach, a shared uniform node kept by the sun rig as the frustum follows and grows): inside, the map has the near occluders with their shapes and only the far band counts; over the frustum's last 20 % the near band fades in, and beyond it the horizon is the higher of the two bands. Why two bands (2026-09-25 review): with the far band alone, ground past the ~110 m frustum lost the shadows of its own neighbours — a 20 m block's 55 m shadow at a 20° sun was missing entirely, at 10° only the detached 80–113 m tail showed. Committed neighbour tiles fill the margin; beyond the site the land is Copernicus GLO-30 (2026-10-21, pipeline/bake/glo30.py): the 30 m surface model, read once per site out of the public cloud-optimised GeoTIFFs (the site plus 10 km, cached as data/_raw/glo30/<site>.tif), smoothed over 3 × 3 cells and moved onto the DGM1 by the median of their difference over the open ground both cover (align: Dresden 2.84 m, Leipzig 3.25, Meißen 2.22, Grimma 1.45, Hamburg 0.46, München 6.36, Unna 1.87 taken off). Without the cached window the site's own edge runs on level outward (extend_edges). Before, the ground beyond was open at the whole edge's mean height (🗃️ below), a wall up to 25° too high near a sloping site's low side. The ring past 1.5 km adds little: up to 2.5°, on 5.8 % of Meißen's point-azimuths. Baked in ≈25–40 s per tile (sky view ≈20–35 s, horizon ≈5 s). The plan's 4 m raster came to 1.86 MB for the far band alone on the spawn tile, past its 1.5 MB cap, so it ships at 8 m with all 16 azimuths (the far band alone was 0.54–0.60 MB; both bands fit the cap). Terrain only; facades wait on plates. Default 0.8 — not yet judged on a GPU (the plan's 21 December plate from the Brühlsche Terrasse and the frustum seam are open). sky-light.ts applyGroundLight (the material's receivedShadowNode, min with the shadow map), hzSunVisible.

Atmosphere & time of day

  • Height-term fog =computed ≈assumed — DGM elevation (per-fragment world height) → extra haze pooling in low ground, a term of the one scene.fogNode every material takes (fog = false opts out: the river mist); Talnebel, fixed at 0.2 (VALLEY_HAZE; Modell fades it out). height-fog.ts. Sets in with distance (2026-10-09): none within 40 m of the camera, all of it from 900 m (VALLEY_NEAR_M); at the full share 10 m off, the default 0.2 greyed every nearby surface by ≈ 18 % (a shop pane at sRGB 102/105/108 against 56/61/65 without it), and full from 90 m it laid one flat veil over the river and the meadows in the lower half of the start view, the same at 200 m as at 2 km. Depth from the site's relief (2026-10-01, ADR 0012): the pool's fade height (heightFalloff) was Dresden's Elbe-to-rim 28 m everywhere, which hazed flat Hamburg and Munich from the river to the rooftops. prepare-data.ts samples each tile's coarse ground on a 32 × 32 grid and writes the site's 2nd and 90th height percentile into the tileset's extras.ground (groundRelief); create-app.ts sets the fade height to 0.6 × that relief, clamped to 8–28 m (lib/city/valley-fog.ts valleyFalloff; 28 m without ground). Dresden, Grimma, Meißen and Unna keep 28 m; Hamburg ≈ 9.6 m, Leipzig ≈ 12 m, Munich ≈ 15 m. The valley's depth from its rim (2026-10-21): the sky-view bake measures per tile its low ground (glo30.tile_floor, the DGM1 averaged over 30 m cells and ≈ 150 m) and, per 22.5° sector within 5 km of it, the rim's height in GLO-30 (glo30.rims; a building's or a copse's height taken out), written as the horizon legend's valley; prepare-data.ts (valleyRims) reads them into the tileset's extras.ground as a third value, the depth: per tile the lower side's rise over its floor across its deepest axis (valleyDepth), the site's deepest tile. The fade height is 0.6 × that depth, 8–28 m; without GLO-30 the site's p90 − p2 as before. Unna 28 → 12.4 m (a one-sided slope below the Haar is no valley, 🗃️ below), Munich 15.1 → 10.9 m, the rest unchanged. Not yet judged on a real GPU.
  • River mist ❝taken ≈assumed — seen along the water only, gone as the eye looks down on it; DLM water mask (the painted splat's alpha) → a drifting, sun-lit mist sheet over the Elbe, always on. None within 120 m of the eye, full from 1.2 km, at most 24 % opaque and only 30 % of the way from the sky's tint to white (it was full from 90 m at up to 62 % and 60 % of the way to white, and the near river in the start view lay under a pale, drifting band that read as a smear on the screen). The strength is the Flussnebel slider's old default (60 % of 0.4) since the slider went (🗃️ below). water-layer.ts createWaterMist.
  • Drifting clouds ≈assumed — SkyMesh's procedural clouds drift on TSL time (slow cloudSpeed), lit by the sun instant. sun-rig.ts.
  • Golden/blue-hour palette stops =computed ≈assumed — sun altitude → sky/fog/hemisphere colours from a palette with stops at −2° (blue hour) and +6° (golden hour). lib/city/atmosphere.ts STOPS.
  • Meadow mottle ≈assumed — DLM class 1 (farmland/meadow) → a low-frequency colour + normal mottle so grass reads as ground, not paint: world-space value noise in two octaves (~8 m and ~2.5 m). Until 2026-11-02 it was the product of two sines, a chequerboard by construction, which from afar read as one. terrain-layer.ts grassMottle/terrainNormal.
  • Contour ink guard =computed — a flat terrain quad lying exactly on a 2 m or 10 m contour has fwidth 0, and 0/0 striped it with NaN ink (a diamond of lines on flat roads). No slope, no contour line. terrain-layer.ts.

Traffic (the data layers)

Overlays on the city, each switched on and off on its own in the HUD (Erkunden → Verkehrsdaten; lib/city/data-layers.ts, one flag per layer in the look store and the snapshot, all off at boot): nothing is drawn, fetched or polled for a layer that is off (the counted sections' features, 37 KB a tile, are fetched with their tile's dressing; the glass bodies are built on the layer's first switch-on per tile, kept hidden when it is switched off and freed with the tile — plan 062, ADR 0040's 2026-10-07 update). They are measurements laid over the poetic city, not part of it: unlit or flat colours, no shadows, no text in the scene (what they say in words is in the sidebar). Judged on SwiftShader plates only so far — unjudged on a GPU.

  • Counted motor traffic ❝taken ≈assumed — in Dresden the city's Verkehrsmengen (WFS cls:L363 "Kfz/Tag", Straßen- und Tiefbauamt, dl-de/by-2-0; 2 082 sections over the fifteen tiles, mostly counted 2023–2026: one-day hand counts scaled to the average day, induction loops and infrared detectors as yearly means, a few estimates) → pipeline/bake/traffic.py → traffic_<tile>.geojson: each section cut to the tile, the vehicles per day along and against its line (f/b, absent where a direction was not counted), the heavy-goods share per direction, the year and method, the street, br on a section that runs over a bridge (2026-10-21, traffic.py on_deck: at least 10 m of it inside one road deck of the tile's or a neighbour's bridge file, each piece within 30° of the deck's axis there — a road passing under a bridge stays on the ground; by the street's name — "…brücke" — only on a tile without a bridge file, 🗃️ below; where the raw counts are no longer cached, --step traffic re-flags only br of the committed sections, reflag). On a bridge: Dresden 23 → 34 of 1 859 sections (25 gained, 14 approaches named after their bridge lost), Hamburg 0 → 33, Meißen 0 → 6, Unna 0 → 3, Grimma 0 → 1. Runtime (app/_components/traffic-layer.ts, lib/city/traffic.ts): one glass body per counted direction on the right of its travel (both: either side of the line; one: on the line; only a total: on the line, not flowing) — a soft dome in cross-section, 1.2–3.6 m wide and 0.6–6 m tall with the root of the vehicles per day, tapering to a round point over the last ≤ 8 m only where the flow really ends: an end within 3 m of another counted section's end, or on the tile's edge (the bake cut it there), runs on full (openEnds — tapering every section end drew the streets as strings of beads, pinched at every junction and seam). Feet 0.2 m under the ground, on a bridge street the deck. Both terrain levels carry the flows: the fine one a cross-section every 2 m with ten segments, the coarse one (the tiles the fine level has not reached) every 8 m with six, about a seventh of the vertices, on the coarse ground; 3D Tiles swaps them with their terrain. Built on the fine level only, the flows ended at the first tile the fine level had not loaded — the Waldschlößchenbrücke stopped at its Johannstadt end, the Käthe-Kollwitz-Ufer showed nothing from across the river. Its tint runs over five stops placed where Dresden's lanes spread (TRAFFIC_SCALE: 400, 2 000, 5 000, 9 000, 16 000 a day — sage, peach, coral, rose, wine; log-spaced between stops; yellow is left to the trams), slate mixed in with the heavy share. One log scale from 300 to 30 000 put every main road in the same coral (half the lanes carry under 2 700 a day, a quarter over 6 000). The flows keep the scene's hour (lib/city/traffic-hours.ts): the counts are per day, so each lane's count is spread over the day on a measured curve — Hamburg's infrared counters on 38 inner-city main roads, hourly, September 2026 (Urban Data Platform, dl-de/by-2-0), per working day, Saturday and Sunday. The hour's multiple of the average hour (≈ 0.08 at 3 h, 1.8 at 16 h on a working day) sets the colour (the count × that multiple on the scale), the body's size (√, from 0.35× to 1.5× of its daily one) and how many of its comets are lit (one in eight at night, all from the average hour on); the light runs at 11 m/s, slowing to about half at the evening peak (integrated on the CPU, so it never jumps). Two shared uniforms, nothing rebuilt; the sidebar says the hour and that the curve is a typical one, not a measurement of the street. Light runs through the body in the direction of travel (soft comets, up to three times as many on a busy lane), brightest along the crown. The glass (glass.ts) refracts the street, trees and houses behind it with a touch of dispersion, a sky sheen and a Fresnel rim, and thickens into its own colour from the air and past ~200 m, so the layer reads as glass up close and as a map from above; the bodies widen up to 5× from the air (map-overlay.ts mapWidenNode). A street without a count draws nothing — no gap filling, no model. A click on a flow asks it (the twin's card: its counts, directions, year, method, the hour's estimate and the source).
  • Bicycle counters, live ❝taken — the city's permanent counters (WFS cls:L1781 "aktuelle Zählwerte", dl-de/by-2-0; 35 counters, the bicycles of the last full hour per direction) read by the browser from the city's server (it answers any origin) when the layer is switched on and every 5 minutes while it is on — not while the page is hidden; back in view it reads at once when the counts are older than one poll (app/_components/bike-layer.ts, lib/city/bike-counts.ts): a pair of glass columns per counter across the street (winkel, its run, degrees counter-clockwise from north), one per direction (teal, lilac), height 1.5 m + 1.6 × √count (482 an hour on the Albertbrücke ≈ 37 m), the top a soft dome; rings of light rise through each at 0.18 m/s × √count, and a soft pool of its colour lies on the ground around its foot; grey and still when the count is older than three hours. The sidebar lists the counts, busiest first; a click flies there, a click on a column asks it. The first runtime request to a server other than the site's own (ADR 0040).
  • Every site, its own sources ❝taken (Site.dataLayers, sites/<id>.ts): a site names the source of each layer it has, and the HUD offers only those (siteDataLayers; the "sources by city" page marks the rest). Motor traffic: Dresden's own per-direction counts; Berlin's Verkehrsmengen 2023 (gdi.berlin.de WFS dtvw2023kfz, vehicles per working day, dl-de/zero-2-0; configured, Berlin's tiles not built yet); Hamburg's Verkehrsmengen 2019 on the main roads (HH_WFS_Verkehrsmengen, dtv + heavy % sv, dl-de/by-2-0; 547 sections on its four tiles); Saxony's road census SVZ 2021 (LASuV, statewide shapefile, dl-de/by-2-0; Grimma 10, Meißen 14 sections: the federal, state and district roads only); NRW's Verkehrswerte (Straßen.NRW WFS, GML only, DTVKFZA + DTVSVA, dl-de/by-2-0; Unna 23). Every source but Dresden's counts both directions together: the bake splits the total evenly (sp 1, m "census" for the road censuses; pipeline/bake/traffic_sources.py), and the HUD says so. A one-way couplet would be drawn as two half flows — the sources do not mark one-way roads. Leipzig and Munich name none: neither city publishes counts per section, and the Land's census (Saxony's SVZ, Bavaria's BAYSIS SVZ 2021) does not reach their centres (0 and 0 sections on their tiles; Munich's are the motorways at the city's edge). Live bicycle counters: Dresden's WFS and Hamburg's SensorThings service (iot.hamburg.de, the "HaRaZäN" infrared counters, hourly per counting point and direction, dl-de/by-2-0; 21 points on its tiles; lib/city/bike-feeds.ts); Leipzig's counters refresh daily without CORS, Munich's monthly, Berlin's yearly — not live, left out. Trams by timetable: Leipzig (LVB; 19 lines with the night and special lines, 2 973 / 2 504 / 1 842 trips) and Munich (MVG; 14 lines, 2 217 / 2 186 / 1 885 trips) from the same feed and their own OSM tracks, Berlin configured (BVG); Hamburg, Grimma, Meißen and Unna have no trams. The feed lives once for every site, data/_raw/gtfs/ (bun run fetch <site> downloads it where the site has trams).
  • Trams by timetable ❝taken ◎detected ≈assumed — in Dresden the DVB's scheduled tram trips (gtfs.de's Germany-wide GTFS from DELFI's NeTEx, CC BY 4.0; ≈ 290 MB, not committed) → pipeline/bake/transit.py, site-wide → data/dresden/transit/trams.json (≈ 680 KB): 12 lines, 127 stop patterns, 264 running-time profiles, 2 598 / 2 002 / 1 629 trips on a working day / Saturday / Sunday (the busiest day of each kind in the three weeks from the feed's date: 3 October 2026, a holiday, lost to the 10th). The feed has no shapes: each leg is laid on the OSM tram tracks (tram_<tile>.geojson) as the shortest way between the tracks near its two platforms, a metre from platform to track weighing three of way, so a car keeps to the track beside its platform; every leg on the site finds one (Postplatz's platforms stand up to 60 m from a track). Runtime (app/_components/tram-cars.ts, lib/city/tram-timetable.ts): every trip of the scene's kind of day (and the night's of the day before) at the scene's clock, which runs on in real time from the HUD's instant; a car's sections each follow the path on its own, standing at its stops from arrival to departure and easing between them; the rail top over the ground, the deck on a bridge track; a slate window band. Each car in its operator's livery and length (2026-10-21, V26–V29): the GTFS agency of its line (agencies, routeAgency in trams.json, from agency.txt; operatorOf, TRAM_LIVERIES in lib/city/tram-timetable.ts) — DVB yellow, 30 / 41 / 45 m by fleet share; LVB silver-grey, a yellow band along the roof, a blue skirt, 37.6 / 45 m; MVG blue, white along the roof, 36 m; BVG yellow, 30.8 / 40 m; any other a pale neutral 30 m — in round(L / 7.5) sections (at least 2) stretched to the length (until then four 7.2 m sections, DVB yellow everywhere; a timetable baked before the agencies draws neutral cars). No shadow (a moving caster would redraw the shadow map every frame). Behind each car a trail of light along its track (deck.gl's TripsLayer in the scene's palette): 140 m, 3.2 m wide, 0.6 m over the rail top (the coarse terrain stands up to half a metre over the TIN), warm white at the car fading to gold and nothing, one mesh for all trams rewritten every frame; it widens from the air. Not live positions: a delay or a diversion is not shown. The sidebar says how many trams run and from which day's timetable.

Planning views (Modell, plan 055 / ADR 0044)

  • Schwarzplan ❝taken (picture style) — the LoD2 clay (userData.figure) drawn with one unlit black, the terrain white (paperGroundOn = 3, its contours off), everything else hidden for the frame; one threshold in the stylize pass cuts the remaining light away (paper-scene.ts, stylize-effect.ts mode 6). No new input.
  • Strich ≈assumed (picture style) — Papier's card on the ground's plan palette (paperGroundOn = 2, terrain-layer.ts paperGround): the painted class colour's grey lightened to a light neutral, green (g > max(r, b) in the painted colour) a plan green, the splat's water a plan blue, the paint grey; the pass keeps each surface's hue and lays the light out as two washes under a one-pixel pen (mode 5).
  • Geländeschnitt =computed — in a Schnitt the terrain's heights along the cut line (ground.ts atWorld, every 3 CSS px, ≥ 0.5 m) → a poché strip down below the picture (lib/city/section.ts groundStrip, model-cuts.ts); unknown ground breaks the strip, never invented. The clay is drawn two-sided with its back faces as poché (visual-style.ts setClaySection).
  • Ausschnitt plinth =computed — the cut-out's four edges sampled on the terrain → poché walls down to 12 m below the lowest corner; the city clipped by the square's planes (ClippingGroup). Buildings on the edge are cut, not kept or dropped whole (that would need a per-object centroid). Switching the group switches every drawable's build (three keys it by the clipping and keeps one render object for both sides), and frees the one it leaves: built in the frame, that was the whole city's programs at once, both ways — the picture froze, a phone's tab died. A new cut now shows once its programs are compiled off the frames on stand-ins, the clipped ones under the group and the plain ones they come back to, and held while it shows (post-stack.ts holdCut); the HUD says Wird vorbereitet … meanwhile. The shadow pass is not clipped (no compile ahead reaches it): a building outside still casts over the cut's edge.
  • Shadow study sheet =computed — the sun (suncalc) at 9, 12, 15, 18 Uhr on 21.3., 21.6., 21.12. of the shown year, one capture each, as a 3 × 4 sheet with the legend (image-export.ts, lib/city/image-export.ts).
  • Trees at every scale ≈assumed — Modell draws the trees of the terrain level under the picture, as every mode does (✅ The coarse level's trees, Vegetation): every tree where the fine level is loaded, a third of them, wider, where the coarse one shows. No scale thins them of its own (🗃️ Modell's trees thinned by scale).
  • Bridges on both terrain levels ❝taken ◎detected — the coarse level draws its tile's bridges as the fine one does (buildRail with the decks alone: deck, parapets, piers or arches, the measured steel; no rails), and they are asked on it too. LoD2's bridge slabs are left out of the buildings, so every river crossing used to vanish where the fine level's reach ended: in the air beyond the fine level's reach (≈ 2.6 km on a viewport 1080 CSS px high), and in Modell across the whole picture from 2.5 m/px on.
  • The bridges' surfaces ≈assumed — a bridge was one flat colour per part (a pale slab over the river). The deck's top now carries its frame — the station and offset on the axis and the outline's edges there, a vertex attribute (aDeck, bridge-layer.ts addDeckFrame; the top split to triangles of ≤ 6 m, splitDeckTop, since a curved deck's triangulation joins vertices up to 190 m apart and the frame is exact only at them) — and its material (bridge-surface.ts, scene-wide, world space) lays a road bridge out as the street it carries: footways along both sides in the pavement's colour with their slabs, a kerb stone and a gutter, the carriageway in the road's colour with a grain and, on two lanes, the centre line's dashes; a path bridge is sand with kerbs, a rail bridge the track bed with its walkways, the rest stone paving. The fascia, parapets, piers and masonry arches are ashlar: courses and running-bond joints on every upright face, a shade per block, a weathering mottle. The fine detail fades by the pixel's footprint; the footways, the kerbs' line and the mottle stay to the smallest scales.
  • The export's legend =computed ❝taken — the scale from the view's metres per pixel (printed at 300 dpi), the north arrow from its turn, the credits from the site's attribution lines (siteAttribution) — drawn into the PNG, never into the scene (ADR 0042).
  • A planner's design in the city ❝taken ◎detected ≈assumed — a glTF or IFC the visitor loads in the sidebar's Entwurf panel, read on the device and never uploaded (design-read.ts; ADR 0054): its nodes and instances flattened into one geometry about its anchor, moved there in doubles. An IFC is tessellated by web-ifc (design-ifc.ts, its .wasm served by the site itself), spaces and openings left out, a material per element colour; its IfcMapConversion (◎detected; the other German UTM zone carried over through lat/lng, other CRSs not read; its values in the CRS's MapUnit, else the project's length unit) or, failing that, its IfcSite's latitude and longitude at the site's own placement, turned by the context's TrueNorth (≈assumed: approximate, the panel says so) put its origin in the Land's coordinates (lib/city/ifc-georef.ts); geometry already in them keeps its place. An OBJ (metres, Y up, its MTL and textures from the picked files, the textures decoded by createImageBitmap in the scene's worker) and a Rhino .3dm (its unit, Z up turned to Y up, its textures and text dots left out — three's loader draws them through the DOM, which the worker has not; the panel says so —, placed ◎detected by its earth anchor where one is set — its elevation as NHN unless it says ground level or the earth's centre, or is unread and Rhino's default 0, then set on the ground; hidden layers left out; breps by their saved render meshes) are read alike (design-formats.ts). A place more than 2 km beyond the site's tiles, or in another place's coordinates, is not taken: ≈assumed in front of the player, the panel saying why. Its scale is ❝taken from the panel's unit and factor (metres unless set; a hint under 4 m or over 3 km) and kept in the placement file. A landscape plan (DXF, Shapefile, GeoPackage, LandXML; design-plan.ts) becomes surfaces ◎detected by their layer's or attributes' names (lib/city/plan-classes.ts, German and English keywords; class-like attributes first, names and addresses last; unnamed: grey) in the city's palette, laid on the loaded ground in triangles of a few metres, at most 150 000 in all (its own heights kept), what is drawn into an area 10 cm above it; trees as low-poly crowns of the plan's size (≈assumed 6 m and 10 m where it gives none), buildings with a height or storeys as blocks, TINs as a planned ground (lib/city/plan-geometry.ts). Its surfaces clear the site but take no house away. Its place is ◎detected from its own coordinates where they are the site's UTM zone (an easting in X, plain or zone-prefixed, a northing in −Z; lib/city/design-placement.ts), its own Y kept as NHN where the ground under it agrees within 40 m; without one it is ≈assumed on the ground ahead of the player. The panel moves, turns and lifts it and saves the placement as a small JSON file. The LoD2 objects whose footprints its plan covers for half or more (every triangle laid flat on 1 m cells, the flat ones at its lowest level left out; lib/city/design-cover.ts) leave their tiles' index and come back when it moves or goes (CityLayer.setHidden). Drawn in a model-maker's white or the file's own colours; casts and takes shadows; walked against. Its bytes come out of the tile cache; past 250 000 triangles / 40 MB on a phone (3 M / 400 MB on a desktop) its textures go, then it is refused (lib/city/design-budget.ts). Its site — the plan with its ground plate — clears what else stands there (design-site.ts, one raster the vertex stages read): every instanced set (trees, hedges, lamps, furniture, monuments) folds an instance whose foot stands in it (and a question no longer finds it, foldedThing), and the terrain with its walls, kerbs, stairs and fences and the water are pressed under the lowest height the design reaches in each cell (a sloping plate's slope, a house's lowest floor); on the tiles under it the ground is also cut where a triangle reaching out of the site lifts it through the plate. Rails, ballast, platforms and tram tracks within 4 m over that floor go under it too (lineUnderDesign; a line on a bridge above it stays). The walked ground follows (Ground.setCap). A building's plinths, cornices, doors and parts go with it (wholeTrees). Asked, a design answers by the part its triangle belongs to (DesignParts, the merged triangles' runs): an IFC element's entity, Name and GlobalId, a plan's class, a glTF's or OBJ's object and material names, and its placement (lib/city/inquiry-design.ts); outlined by that part's own triangles. Fallback: without coordinates, the player's place; without a ground under it, the player's ground. Not yet: the tram's overhead wires and a river's mist over a site stay; sharing by link.

🧪 Experimental

  • DOP-lean caveat (recorded): =computed standard DOP has building lean (tall roofs displaced over facades). Mitigated in the bake by eroding the roof footprint inward + a robust median; revisit with true-orthophotos if available.

📋 Planned (designed, not built)

Ranked roughly by impact-vs-effort (full rationale lives in chat history / the research that produced them):

  1. DOP NDVI tree-placement density — (crown colour + meadow tint now ✅ active) — the remaining NDVI use: thin/thicken canopy placement by local NDVI so sparse/stressed areas get fewer trees. Lower priority than the two shipped.

  2. Stylized DOP ground-drape — posterized, desaturated DOP blended over terrain past ~150 m for far-distance texture. Aesthetic risk — prototype behind a slider, judge on GPU before committing.

  3. ALKIS parcels — plot boundaries → per-parcel ground tint, garden/courtyard vs street, fences along lot lines; richer Gebäudefunktion than CityGML. The same download would carry the surveyed stairs and walls (AX_SonstigesBauwerkOderSonstigeEinrichtung: Treppe, Mauer, Stützmauer) — a second, official source for stairs.py and walls.py where OSM is thin (ADR 0030).

  4. Cartographic minimap — DTK / basemap.de P10 raster tile + Ortsteile labels replacing the math-drawn minimap.

  5. Dappled canopy shadow — a colour-less proxy caster per chunk whose maskNode cuts the gaps (the shadow pass honours it; the old WebGLShadowMap alphaMap-override gotcha left with plan 020).

  6. Real trees from the laser-scan point cloud — segment high-veg returns → per-tree position/height/crown; bake to per-tile GeoJSON. (First step ✅ above: canopyx crown peaks with h + r, outside the canopy mask only, thinned against the cadastre; since 2026-10-07 the renderer sizes such a crown from its r, 2r within 0.4–0.75 × h. Since 2026-10-20 r is the crown's share of the scan's crown mass — the watershed of the smoothed heights from every scan peak and every canopy and cadastre tree — as a disc of the same area; before, it was the distance to the mass's edge at the peak, which read a lone crown from a peak off its centre and a group's at its narrowest.)

  7. Cascaded Shadow Maps — shipped as four concentric rings (ADR 0009): shapes out to 2.2 km on foot and from the air, on a phone in the bytes of its one map. Past the widest ring the baked horizon (✅ above) casts the long shadows onto the ground. What is left of this idea is virtual shadow maps, once three has multiview.

  8. Adaptive resolution while moving — partly shipped: DoF is skipped while the camera moves (lib/city/regression.ts, plan 007; a phone builds no DoF at all, ADR 0047). AO is not — gating it made the contact shadows blink on every step, so the AO pass (GTAO since plan 020) runs permanently at half resolution instead (ADR 0009). A pixel-ratio drop under motion is the open half (needs a ~1 s hold and a real-GPU look).

  9. Stay cables and bridge colour — trusses, pylons and steel arches now come from DOM1 (✅ above). The Pieschener Molenbrücke's pylon measures, but its stay cables are too thin for a 1 m raster: they are fanned from the measured pylon top by rule; tracing them would need the laser point cloud (LSC). The steel is one painted colour; sampling the DOP along a rib's chord could give the Blaues Wunder its blue.

  10. Atmospheric motes — the one unbuilt item of the aesthetic roadmap: 2–4 k camera-local billboards (not Points: WebGPU draws them 1 px wide) drifting in a toroidal volume (R ≈ 30 m), additive, depthWrite: false, fog: false (fog would brighten distant motes), hash-seeded so snapshots reproduce, opacity + drawCount on one slider. +1 draw call. Design notes in plans/README.md.

  11. Far crown LOD tier — ✅ shipped (see Crown shaping above); still to judge on a real GPU with the --headed harness: whether the far tier's pop at 650 m and the thinned forest read well, and whether a detail-0 crown (20 tris) beyond ~1.5 km is worth a fourth tier.

  12. Ground, what plan 023 left — parks, cemeteries and sports grounds split out of the DLM's built-up class by object type (sie02_f OBJART/FKT, needs the raw DLM); the laser-scan intensity (LSC) as a measured surface-material map where OSM is silent (both ranked in the data-streams survey). An idea only, not planned: shell-textured grass near the camera (4–8 shells, meadow only, no shadow casting) — it runs against the calm, nothing-finer-than-a-pixel look the maintainer has asked for twice (🗃️ drawn fence panels, allotment bed bands); revisit only with a real-GPU plate that stays calm. The raised pavement and DGM1 micro-relief are 🗃️ below.

  13. The 2026-09-25 batch — each planned in plans/, now condensed in completed.md with its looks in plan 019: trams with contact wire (024), trees by species and season (025 — built, GPU plates open), road markings (026), shop glow and heritage from OSM on the buildings (027: ✅ above; the era is 🗃️), allotments, orchards and vineyards (028), fences and gates (029: ✅ above), more street furniture (030), Elbe landing stages, groynes and ferries (031), street names (032 — built, then removed: 🗃️ below), sky-view factor and a baked horizon map (033), small structures from DOM − LoD2 (034), a hidden soundscape (035 — built, then removed: 🗃️ below). Built since (✅ above, looks unverified on a GPU): 024 (Trams), 026 (Road markings), 028 (Cultivated land), 030 (Signs and fixtures), 031 (Landing stages, groynes, ferries), 033 (Sky-view factor, Horizon shade), 034 (Small structures from the laser scan).

  14. The traffic data layers, next (✅ Traffic above) — tram delays from the VVO's departure monitor (webapi.vvo-online.de/dm, real time per stop, answers any origin: poll the stops in view, shift each trip by its delay); the bicycle counters' hourly history since 2017 (cls:L1780, > 140 MB unfiltered: bake a typical day per counter and run the columns with the scene's hour); the accidents with injury (Unfallatlas, the statistics offices, dl-de/by-2-0; pedestrian and cyclist involvement per point) as a fourth layer.

  15. The twin, next — asking trees, bridges and monuments too is ✅ built 2026-10-01 (plan 052 phase 4; see the twin section above). Still planned: a link to an asked building, a Datenstand panel, the ingest writing the provenance (plan 052 phases 4b–7); the day playing, weather as mood, the Elbe at its gauge (plan 053); flood, sun hours, sight lines and a planned building as scenarios (plan 054).


🗃️ Discontinued / rejected (do not retry blindly)

IdeaWhy rejectedCaveat
A depth-of-field switch and manual focus (Tiefenschärfe on/off, Fokus: Automatisch / Manuell with a Fokusdistanz slider, 1–3 000 m, drawn as a ring on the minimap; until 2026-10-10)The maintainer chose fewer settings: the depth of field must stay, but it need not be parametrised.DoF is always on where it is built (desktops; a phone builds none, ADR 0047), focused on the crosshair; it still rests while moving, in the graphic styles and in Modell (two prebuilt pipelines, ADR 0009). An old snapshot's dof, focusMode and focusDistanceM are ignored.
An artistic depth-of-field ramp (sharp to fully blurred over 1.6 × the focus distance, at least 45 m; until 2026-11-06)The maintainer found it too strong, and it snapped on the moment the camera stopped: the half-resolution blur reached its full mix halfway up the ramp, so everything past about twice the focus lost its detail whatever the distance.A thin lens with the view's angle and a 50 mm's opening at f/1.8 (lib/city/lens.ts): a far focus leaves the city nearly sharp and blurs what stands near, a near one softens the city, and the blur grows in over 700 ms after a stop (ADR 0009).
The view's own lens at f/1.8, drawn by three's DepthOfFieldNode (a 23 mm, 13 mm across, the blur capped at 8 px; 2026-11-06 to 2026-11-09)The maintainer saw no depth of field: a tree 6 m ahead blurred by under a pixel with the focus down the street, the node mixed a blur in by its share of the widest (a fifth of the widest showed two fifths blurred), and with nothing under the crosshair the focus stayed on a point the camera had passed (1 m).The opening of a 50 mm (28 mm), a cap of 2.5 % of the frame's height, an own gather that draws each circle as wide as the lens makes it (lens-blur.ts), the sky's focus when nothing is hit.
A river-mist slider (Flussnebel, until 2026-10-10: the sheet's strength 0–100 %, the sheet hidden at 0)The maintainer chose fewer settings: the mist stays, but it is not something to tune.The sheet is always drawn, at the slider's default (60 %: at most 24 % opaque, MIST_OPACITY). An old snapshot's waterMistPct is ignored.
Leaf flutter (Blattflimmern, until 2026-10-10: small world-space value-noise specks, two octaves at ~1–2 m cells drifting with the wind, blending sunlit, sun-facing crowns toward a paler silver-sage "underside" with a faint glint; and its companion slider Windhelligkeit)The maintainer chose fewer settings, and the twinkle was not free: two leafNoise calls — eight sin hashes — on every crown fragment.The sway-coupled brightness stays, fixed at its slider's default (GUST_BRIGHTEN, 0.09 = 50 % of 0.18); it rides on the sway's gust signal, not on the twinkle. An old snapshot's leafFlutterPct and leafBrightPct are ignored.
The look's sliders (25 of the HUD's 26 percent sliders, until 2026-10-10: Talnebel, Tiefenfärbung; the buildings' Boden-Verlauf, Höhenlinien, Streiflicht, Farbvariation, Dachfarbe, Dachsättigung, Traufkante, Gliederung, Fenster, Fassadenbild, Abendlicht, Materialstreuung; Bodendetail, Stadtgrün, Wiesenfärbung, Gegenlicht-Schimmer, Blattdurchscheinen; Kontaktschatten, Himmelslicht, Ferne Schatten, Spiegelung, Papierkorn, Tuschelinien)The maintainer chose fewer settings ("we don't need many of the settings anyway", plan 073). A slider cost the GPU almost nothing — a uniform multiplier; every term ran at any value, 0 included — but each was a table row, a store value, a snapshot key, a writer in its owner and a guide line in two languages.Every term stays, at its slider's default, as a named constant beside it in its owner (TINT, WINDOWS, GROUND_DETAIL, SKY_VIEW, VALLEY_HAZE, AO_EXPONENT, …), so the default picture is unchanged; the depth grading, paper grain, ink and valley haze keep their uniforms, which a picture style or Modell writes. Nebel stays (the multi-tuft switch stayed too, until 2026-11-06: see The near crowns' switch). Tuning a term is a code change, judged by snapshots of place and time (ADR 0018). An old snapshot's keys (tintPct, windowsPct, …: RETIRED_SNAPSHOT_KEYS) are ignored and never reused. The measured follow-up: delete what now runs at a fixed 0 (the glass sheen and the gilding's and fountain water's stand-ins under full Spiegelung; the river's went with its mirror, 2026-10-25) and tier-gate the costly terms (windows, facade reading, ground detail) on phones.
The near crowns' switch (Multi-Tuft-Kronen (nah), look.multiTuft, until 2026-11-06)Maintainer review (2026-11-06): it should always be on; no setting needed. Off, it only swapped the near crowns for the cheap ones — a look nobody chose.Removed: the rich crown is always allowed within the site's rich budget (RICH_TREE_BUDGET, a phone's PHONE_RICH_TREE_BUDGET). With it the cadastre's own silhouettes left their separate distance swap and joined the chunks' plan, so they now count against the same budget (a phone's included). The key is retired (RETIRED_SNAPSHOT_KEYS).
See-through buildings (Transparenz, until 2026-10-10: the clay's opacity hash-dithered, alphaHash, up to 90 %, in the opaque pass)The maintainer chose fewer settings. It was a second build of the clay's graph — compiled in the frame the slider first left 0 — and every crossing of 0 redrew the shadow map, for a look that judges nothing in the city.The clay is always opaque; an old snapshot's transparencyPct is ignored. To see what a building hides, demolish it (R) or cut the model open (Modell's Ausschnitt).
The hidden soundscape (plan 035, 2026-09-26 → 2026-10-10: wind, the city's hum, the Elbe, leaves, birds, crickets, fountains, footsteps by paving, a tram bell and the hour struck from the bell towers, all synthesized with WebAudio behind the L key and the Klang (experimentell) switch; app/_components/soundscape/, lib/city/soundscape.ts; the towers baked by pipeline/bake/soundmarks.py → soundmarks_<tile>.geojson)Removed at the maintainer's decision (2026-10-10) to cut code and maintenance: about 3 000 lines, a bake and a tileset column for a feature that never shipped audibly — none of it was ever listened to (plan 019's listening pass stayed open).Engine, switch, bake and committed files all went; the viewer creates no AudioContext. Revive only with a listening pass first, from git history (5b59b79).
Sports-ground lines beyond a desktop's fine level (until 2026-10-06 on both levels on every device; until 2026-10-10 on a phone's fine level and a desktop's coarse level too: sport_<t>.png read there, sport-ground.ts in that level's colour node)The index raster is a 2048² RGBA texture, 16 MiB on the GPU for a PNG of a few dozen kB, per tile that reads it (14 of Dresden's 15 tiles have grounds) — on a phone against a 336 MiB tile cache (ADR 0047). Beyond the fine level a pitch is a few pixels and its lines a pixel or less; on a phone the lines were the price of a sixth of a fine tile's rasters. A pitch then shows its land-cover class (readsSportGrounds; the level builds the shared variant without the slot); the goals and nets come from the table and stay.A 1024² phone twin of the raster (sport_low, as markings_low) would bring the lines back on a phone at a quarter of the cost; it needs a bake and a look on a GPU.
A visible-band turf test (for a pitch with no tagged surface on a flight without infrared — Bavaria's —, where the GLI cannot tell turf from grass; measured 2026-10-21, not adopted)On München's tagged pitches the luminance parts turf and grass (≤ 109 against ≥ 126, once two tag errors are set aside), but on every infrared flight grass reaches turf's luminance and no visible feature parts them: one flight and 14 pitches are no rule.A green ground on an RGB-only flight keeps its sport's default (✅ Sports grounds). A second RGB-only flight with tagged turf and grass would be the test.
One mean height for the ground beyond the site (the sky view's and the horizon map's ground outside the site's tiles: the mean of the whole edge, until 2026-10-21)On a sloping site one mean for the whole edge stood a wall up to 25° too high over the low side: 6.1 % of Meißen's point-azimuths came out more than 3° off, 3.2 % of Unna's.Beyond the site the ground is Copernicus GLO-30 now (✅ Sky-view factor, Horizon shade); without its cache each unmeasured cell takes the nearest edge cell's height (extend_edges), so nothing beyond stands higher than the edge it continues.
The site's ground relief (p90 − p2) as the valley haze's depth (until 2026-10-21)A one-sided slope read as a valley: Unna's 70 m from its low to its high ground set a haze 28 m deep on a town that lies on a slope, not in a valley.The depth is each tile's rise to its lower rim within 5 km (✅ Height-term fog: Unna 12.4 m); p90 − p2 stays the fallback where no GLO-30 cache is.
The counted traffic's bridge flag by street name (until 2026-10-21)A name says little about a deck: Hamburg's sections carry no names (0 of the 33 on decks flagged), Meißen's carry road numbers, and Dresden's left 25 on decks unflagged and flagged 14 off any deck.The flag comes from the deck geometry now (✅ Counted motor traffic); the name is read only on a tile without a bridge file.
Mapillary's litter bins (object--trash-can, drawn as bins on Dresden until 2026-10-21)The class is any trash can, mostly household wheelie bins put out for collection: 7 of 10 on Dresden's spawn tile and 13 of 15 in the small cities, none clearly public.The bins OSM maps (amenity=waste_basket) stay; Mapillary gives lamps only (✅ Street lamps OSM lacks, from Mapillary).
A whole-sky grid for the roof colour's sun (the shadow fit's sun searched over every elevation and azimuth, until 2026-10-21)Without the flight's clock the fit found suns the day never had: Dresden's 33416_5654 at 29° and Grimma's 33342 at 23° under a 38.5° and a 40° sun.The sun is searched along the flight day's solar path now (✅ Roof colour); the grid (10°/5°, then 2°/1° round its best) remains only for a tile the fetch found no flight days for.
A two-colour roof split from the DOP (a roof's two materials told apart in the orthophoto, measured 2026-10-21, not adopted)242 of the 281 splits found were light, not material: a roof's sunlit and shaded faces, not two coverings.One colour per building is the property table's row; a real split would need a mask texture (ADR 0036).
Missing buildings for outlines LoD2 covers in part (OSM outlines LoD2 covers 20 % or more, measured 2026-10-21, not adopted)Of the 163 candidates most were courtyards and trees, not a building LoD2 lacks.The missing buildings keep the < 20 % rule (✅ Structures beyond LoD2: missing buildings).
The river's mirror and drawn strokes (the sheet's view-angle Fresnel toward the mirrored sky, Spiegelung, sharper between gusts; for a day, 2026-10-25, light sky-coloured arcs on the crests)The maintainer preferred the matte water seen while the sheets were missing ("tatsächlich mag ich den matten Stil gerne", 2026-10-25): the drawn map's water, with only the sun's reflection and a gentle movement on it.The matte river (✅ Water: waves): the body's tint lit through the wave field's normal, the sun's soft path, glints and dashes. The fountains keep their mirror.
Crossing sine ripples on the water (three sines on the river sheet, three swells on a fountain's basin, until 2026-10-23)Their interference repeats every few tens of metres and read as a tiled pattern ("zu repetitiv", the maintainer, 2026-10-23). And on the river they lay flat on every tile without a foam line: their position was first built inside the foam loop (water-layer.ts foamAt), which a tile without a weir never runs, so the Elbe broke at a tile seam into rippling and mirror-flat water.The wind's wave field (✅ Water: waves), the foam loop's position taken before the loop.
Modell's trees thinned by scale (2026-10-04 → 2026-10-05: every tree to 1 : 5 000, fewer to 1 : 9 000, a floor of 30 % beyond, the kept crowns widened by 1/√share — treeShare, treeSpread, the crown shader's crownKept folding the others onto their origin; the coarse level's crowns, the floor's selection, fetched only once Modell thinned. The version before it thinned to none at 1 : 9 000)Two rules for one picture: the scale set the share, the tile renderer the level, and the two disagree. The coarse level shows wherever the fine one is not loaded, not only past 2.5 m/px — at 1 : 3 000 on a phone whose memory governor coarsens the stream, or while fine tiles load — and there the picture showed no tree at all, since its crowns came only once Modell thinned (maintainer, 2026-10-05: trees at large and at small scales, none at 1 : 3 000). The version before showed no tree in an overview of the city.Each terrain level carries its own trees, whatever the mode or scale (✅ The coarse level's trees): the fine level all, the coarse one a fixed third. A generalization by scale, if it comes back, has to follow the level the renderer shows, not run as a second rule beside it.
Pedestrian counts (for the traffic data layers, 2026-10-01)No open measured source for the site: the only pedestrian counter, hystreet.com's laser count on the Prager Straße, is commercial (access on request); the city's "Fußgängerquerung St. Petersburger Straße" is a bicycle counter. A density modelled from stops, shops and census cells would be a model drawn as if measured.Revisit with an open counter network (Telraam sensors in the site, if any, through its API with a key).
Live tram positions (GTFS-RT, the TLMS radio telegrams)gtfs.de's realtime feed is one protobuf for all of Germany, too heavy for a browser to poll; TLMS (wss://socket.tlm.solutions) is a volunteer service whose coverage and uptime the viewer cannot vouch for.The timetable runs instead (✅ Trams by timetable); delays could come from the VVO's departure monitor (webapi.vvo-online.de/dm, answers any origin) per stop — 📋.
Street and square name lettering and the on-foot caption (plan 032: OSM highway names, named squares and the DLM bridge names lettered on the ground from a per-tile Canvas-2D atlas, fading in from 25 m up; on foot, the nearest named street ≤ 25 m in a HUD pill; pipeline/bake/names.py → names_<tile>.geojson, name-layer.ts, street-caption.tsx, lib/city/names.ts)Removed at the maintainer's request after review on a device (2026-09-26): the map look reads better without text. Bake, committed files, layer and caption all went.The DLM bridge name stays in the bridge files. Revive only with a new look decision, from git history (4b08993). Text now appears only on demand, in the inquiry card (ADR 0042).
Every CityJSON attribute as a string column, part ids included (plan 052's first cut)+132 kB gzipped (+9.6 %) on the spawn tile's building glTF, 70 kB of it object ids: Saxon BuildingParts carry random UUID_… ids (41 bytes, incompressible) that no other dataset joins on.The Building's gml:id for every object of its tree and ENUM columns for the code lists: +46 kB (+3.5 %) (ADR 0042). The part UUIDs stay in the committed CityJSON.
Drawn fence panels (plan 029's first look: bars every 12.5 cm, a wire diamond mesh, pickets, posts every 2.5 m and a top rail, alpha-cut in the shader, a dithered veil far off, a dithered partial shadow through a custom depth material)On a real phone "zu hart und kleinteilig", then "stärker stilisiert, mildere Farbwahl, Kleinteiligkeit führt zu Artefakten" (maintainer, 2026-09-25): dark iron and slate read as ink against the pastel scene, and every feature finer than a pixel — bars, mesh, posts, the dithered holes — aliased into moiré and shimmer, near and from the air.A fence is one low band in one muted tone (✅ above): no holes, no dither, nothing finer than its own height. Revisit a pattern only with a real-GPU plate at walking height and from 150 m that stays calm.
A raised pavement behind the kerb (plan 023's leftover: lift the pavement a kerb's height above the road)The DGM1 has no such step to lift, and ADR 0030 makes every part meet the ground at the ground's own level: the kerb's back now runs down to the pavement (meetGround), which took its ground-join misses from 32 % to 2 % (2026-10-01 audit).Only with a measured kerb height per street (none of the sources carries it) and a terrain cut that ADR 0030's join check accepts.
DGM1 micro-relief as a normal texture (plan 023 phase 7: a 1 m normal map over the "2 m mesh")The fine level is a ±0.15 m TIN of the native DGM (ADR 0030), not a 2 m grid, and terrainNormal (terrain-layer.ts) deliberately pulls near-flat normals up: the DGM's ruts, survey wobble and 8-bit normals read as dirty flecks under a low sun. Kerbs are real geometry now.Revisit only if a real-GPU plate shows the ground too flat — and then as a softer calm threshold, not more relief.
A shop window at an OSM shop node (2026-10-08, the shopfronts' first bake: a wall no photo showed as a shopfront, with an OSM shop=* or café node within 3 m, got one 2.4 m bay centred on the node — 1 074 walls over the fifteen Dresden tiles, all but the spawn tile's)A node says a shop is there, not where its windows are or how wide: the lone narrow bay with its riser, head and piers read as a door, not a shop window (the Hauptstraße's GDR pavilions, glass from end to end, showed a row of "doors"; owner's review). Better no window than a wrong one.Every tile's panoramas are measured now; a wall they do not show stays plain. The OSM shop still tints the ground floor (Ladenzone) and lights it at dusk.
Window niches as geometry (PR #143 as first pushed, 2026-10-09: each measured window cut into its wall as a niche, rounded 6 cm into a 14 cm reveal, the wall re-triangulated round it)Too plump: a rounded niche must survive the 2 km tile's 16-bit quantised positions, so its radii were centimetres wide and the openings read as soft lumps (owner's review: "Die Fensterformen sind zu plump"); every window cost triangles (+0.84 % over the site for 4 000 windows — all windows of all houses would have been millions); and only measured walls had them, 527 buildings, about 3 % of Dresden's, so a street of plain houses carried every thirtieth one windowed.The same measurement now drives windows drawn by the clay on every house (Windows on every house, ADR 0010): no triangles, the recess a parallax look into the wall.
Procedural window grid on facadesReads as a modern office block, fights the historic LoD2 silhouette (user veto).Faint storey banding is the only kept remnant. Re-affirmed by plan 050: a glass facade (OSM building:material=glass) gets a cool tint and a sky sheen, never panes, mullions or a texture. Two exceptions, both measured, never procedural: (1, user decision 2026-10-08) shop windows on a shop's ground floor may show glass — dark, slightly reflecting panes where street photos measured the bays, divided only by a shop window's own mullions, never a grid and never above the ground floor (Shopfronts from street photos); (2, Manuel's requests of 2026-10-09) every house may show windows, drawn by the clay as recesses in the wall's own colour, darker and cooler inside — each building's own rhythm (its axis spacing, width and storey where street photos measured it at Spearman ρ ≥ 0.6 between two drives, else borrowed from the nearest measured house of its roof form and eave, else its type's, calibrated to the photos' medians), its storeys from OSM's levels or its height, centred on each wall; no panes, no mullions, no texture (since 2026-11-06, when Manuel lifted the glass veto, their back mirrors a little sky at a grazing view as a shop window's pane does), and never on a party wall, beside a door, over a shopfront, on a church, a hall, or a glass or metal facade — a palace, museum, theatre, parliament or court gets tall windows of its own type on storeys of ~5.5 m (Windows on every house, ADR 0010). What stays vetoed is the uniform grid pasted over facades regardless of the building.
Structures from the surface model alone (plan 050: every tall gap between DOM1 and max(DGM1, LoD2 roof) as a column or block)The gap is mostly not buildings: tree crowns, power pylons, and on the flight day the construction cranes — Leipzig's centre showed dozens of 50–95 m spikes with no mapped structure, Hamburg 90–98 m ones beside a 2023 building.Only where OSM names the structure (man_made=*, a building outline) does the surface model measure it; a landmark's roof relief is the one gap drawn without a mapped outline, and only over its own LoD2 roof.
Invented or imported landmark geometry (hand-modelled landmarks per city, glTF models from 3D warehouses, a stock spire or dome by type)Not repeatable for the next city, a licence per model, and a detailed model breaks the clay style next to LoD2 boxes.Geometry beyond LoD2 comes only from a measurement confirmed by a name (ADR 0036).
Landmark roof relief as stacked slabs (plan 050 as first built: the excess over the highest LoD2 roof cut into ≥ 2 m bands, at most 8, each band's smoothed outline extruded as a flat slab — a contour model)A spire LoD2 cuts short became a stepped pyramid (Unna's Stadtkirche: LoD2 stops at 137 m, the surface model at ≈ 178 m) and the Elbphilharmonie's crests a flight of terraces — the plan's own STOP ("reads as a stack of plates"); 508 slabs on 74 objects in the first bake.Replaced by the measured height field on the 1 m grid, lightly smoothed, built as one surface with corner heights averaged (✅ Landmark roof relief). Still no invented smooth roof: every height is a surface-model cell.
Rebuilt roofs as stepped flat blocks only (ADR 0036 as first built, 2026-10-01 → 2026-10-03: every region of 1 m height bands at its median, a prism each)Right for the flat complexes the gate was made for, wrong for every roof it caught that is not flat: the Frauenkirche's dome (its LoD2 part ~5 m under the scan, so rebuilt) became four stacked drums, the Hauptbahnhof's arched train shed a staircase of 52 ledges, the Schauspielhaus's hipped roofs contour lines — the relief's lesson again ("a stack of plates").Replaced by the measured faces (✅ Roofs rebuilt from DOM1): the same objects, flat levels where the scan is flat, its surface where it slopes or curves.
Rebuilt roofs' faces as the smoothed scan (ADR 0036, 2026-10-03 → 2026-11-06: every face the scan's surface, two edge-preserving Gaussian passes)Tiles, dormers, skylights and the metre of blur at every edge read as a crumpled blanket over ordinary pitched roofs, and gable flanks too narrow for a face as a flight of terraces with a wavy step: An der Frauenkirche 5's ring of roofs looked like a dune.Replaced by the plausibility rule (✅ Roofs rebuilt from DOM1): planes, an extruded profile, or the scan only on a landmark.
Rebuilt roofs with strips and grid-edged outlines (ADR 0036, 2026-10-03 → 2026-10-19: any part ≥ 20 m² wider than a cell or two drawn on its own, borders on the 1 m grid simplified at 1 m)The strips a few metres wide (a parapet run, the ledges along a vault's eaves, a mansard's terraces, plant rooms) stood as boxes of their own height, and every border diagonal to the grid as a staircase of 1 m walls: the Hauptbahnhof, the Prager Spitze and An der Frauenkirche 5 read as heaps of blocks.Replaced by the 2026-10-19 update (✅ Roofs rebuilt from DOM1): a part holds a disk 7 m across or joins a neighbour within 3 m; outlines from a 2 m majority, simplified at 2 m.
A lathed column for every mapped tower (plan 050 as first baked: man_made=tower, lighthouse and water_tower measured like masts and chimneys)Where LoD2 has a roof over the mapped foot it draws the tower already; the column stood inside it and doubled it — Meißen's cathedral towers (81/83 m), Grimma's churches, 13 of 14 towers in Munich.BUILT_TOWERS under a LoD2 roof are skipped; a mast or chimney on a roof is not in LoD2 and stays.
Every Basis-DLM rail at street level (the rail layer before 2026-09-30)The DLM files subways and S-Bahn trunk lines as ordinary railway lines: Munich's U-Bahn (23 km) and the S-Bahn trunk line (4.2 km) ran across the Marienplatz and the Odeonsplatz as broken track, and OSM's U-Bahn platforms lay on the squares above them.A stretch within 2 m of a DLM tunnel (BWF=1870) for > 15 m is cut; OSM platforms and tram ways below ground are skipped.
A 6 m tunnel reach for the rail cut (the first cut: a rail within 6 m of a DLM tunnel)Munich Hauptbahnhof's surface tracks run straight over the S-Bahn tunnel, parallel and a few metres beside its axis: the wider reach cut them too.The DLM draws a tunnel on its rail's own axis, so 2 m is enough; plus the 15 m overlap floor so a surface track crossing over a tunnel keeps its crossing.
Street bed for trams on the road class alone (plan 024's rule)A track through a pedestrian zone or a paved square (class path, built-up) got the gravel bed: Munich's trams by the Marienplatz read as broken track (≈ 13 km ballast in Munich).Pavement is road, path or built-up; ballast in Munich ≈ 3 km.
Absolute notability floor for landmarks (a fixed sitelink threshold)Set where a metropolis's landmarks stand out, it left Unna with one.A low floor (≥ 2 sitelinks), then a floor relative to the tile's most notable landmark (below).
A fixed count of landmarks per tile (the most notable 12, plan 050 as first built)On a dense old-town tile the count, not notability, decided: the second rank fell off the Altstadt tile behind the Frauenkirche, the Zwinger and the Semperoper, while a quiet tile listed everything with two sitelinks.Those with ≥ 8 % of the tile's top sitelinks (≥ 2), 40 as a safety cap only (2026-10-01, ADR 0012).
Per-site facade material (Site.facades: "brick" for Hamburg, plaster elsewhere)One palette for a whole site is wrong inside it — Hamburg is clinker in the Speicherstadt and glass and plaster in HafenCity, Dresden has brick quarters too (246 mapped brick walls) — and every new city needed tuning by hand (ADR 0012).Removed 2026-10-01: each building wears what its OSM neighbourhood is mapped as (✅ Facade context). A look that differs between cities is derived from data, not added to Site.
One valley-haze depth for every site (Dresden's 28 m Elbe-to-rim drop)Tuned on Dresden's valley; on flat Hamburg and Munich the pool reached the rooftops and the whole city stood in haze.The depth follows the site's ground relief (✅ Height-term fog), capped at Dresden's 28 m.
Rails, ballast and bridge decks from the Basis-DLM only (until 2026-10-01)Honest, but Hamburg and Berlin publish no open DLM: a city of bridges had no bridges and no track — the U3 viaduct, the Hauptbahnhof's approaches.OSM's rail ways and bridge ways stand in, the deck heights still measured in DOM1 (✅ Rails, ballast and decks from OSM).
Raw OSM colours on facades (building:colour used as given)A mapped red or #ff0000 is a signal colour in a pastel city.Hue kept, saturation and lightness clamped into the clay's register.
Building era (colour by construction year; plan 027 phase 3)Coverage: OSM carries start_date on 52 and year_of_construction on 12 of 8 310 building outlines in the four first tiles, measured 2026-09-25 (0.8 %, far under the plan's 30 % bar). No official source is reachable: the LfD Sachsen heritage layer (INSPIRE WMS iwms_gsz_schutzgebiete, Kulturdenkmale_Flaeche) answers GetFeatureInfo with designation and name but no dating, its WFS paths are refused (403); the Denkmalliste's dating lives only in its web app, per object; Dresden lists its Kulturdenkmale among the themes without an open dataset (2026-09-25).Revisit with an official Baualter dataset (the city's, or ALKIS baujahr where a Land fills it); listed buildings alone would colour only the monuments.
Allotment bed bands (plan 028 as first shipped: 1.2 m soil/green/grass stripes per ≈12 m jittered-Voronoi plot over a NEAREST colony-id raster)Maintainer feedback on a phone (2026-09-25, 33410_5658 from ≈180 m up): the colony's edge and its carved paths showed the 1 m raster's staircase, and the flat pale stripes read as a rendering glitch, not as gardens.Replaced by a baked signed distance (LINEAR, a soft wandering edge) and analytic plots — soft greens, thin soft paths, a few warm beds, flower dots — box-filtered and faded with distance (✅ Cultivated land). Keep cell ids off any boundary the eye can see.
Orthophoto for facade colourNadir DOP only sees roofs — no facade data.DOP for roofs is fine and is now the 🧪 entry above.
Plain foliage translucencyReads as "noise" at instance distance.Superseded by the next row: the shadow-gated version went too.
Backlit crown shimmer and translucency (until 2026-11-06: a warm emissive glow pow(back, 3.6) and a broader green one pow(back, 1.6) on near or large crowns wherever the eye looked towards the sun, gated on the daylight ramp only since plan 020 — a node material cannot sample the sun's shadow map)Maintainer review (2026-11-06): against the light a crown went brighter over its whole face, evenly, so it read flat and odd instead of shining through.Removed, no replacement: a crown against the light is lit as the sun and sky light it (its occlusion, its sunlit rim). The crown material lost its emissive node and a varying. Revisit only with a real light-through term (depth through the crown), never an even glow.
VSM shadows"Corduroy"/grid rings on large ground at grazing sun.Use PCFShadowMap + radius instead.
Large normalBiasBright peter-panning contact strip.Keep normalBias=0, small negative bias.
Bigger shadow frustum / 4096 mapCoarser texels → fraying / cost without gain once radius softens.Concentric rings of 2048 + radius (ADR 0009); until 2026-11 a tight frustum at 3072 (121 m on foot since 2026-10-11, paired with a 30 m dead zone: the same reach, a third fewer redraws, texels 10 % coarser — ADR 0009's update).
Sobel / deferred outlinesHard edges clash with the watercolor look.The asked element's outline (✅ The outline) is not one: a single soft silhouette of one element, on demand, never the scene's edges.
The pencil loop (plan 052 phase 4's first mark: a wavering graphite stroke on a band of paper, once round and a little past, on the ground around an asked tree or monument and along a deck's parapets; lib/city/pencil.ts, pencil-mark.ts)Maintainer review (2026-10-01): which element is selected should read more strongly. A ring on the ground says "here", not "this" — beside a crown among crowns, or along a deck seen from below, it did not name the element; and a building, hatched only, had no line at all.Replaced by one screen-space outline for every kind (✅ The outline), in the same graphite and paper.
Selective bloom, quad leaf billboardsNo payoff yet for the cost.Revisit only with a concrete need.
BatchedMesh for buildingsAlready merged per tile; would break objectid picking/demolish and not cut draw calls.Bottleneck is fill-rate, not draw calls.
Blender texture bakingNo UVs on the source geometry.—
Orthophoto as the only tint sourceLeaves everything identical where imagery is flat; no facade info.Hash carries variation; DOP augments roofs.
A line lifted onto every deck under it (rails and trams until 2026-10-01: a point inside a rail deck's outline rode the deck, a tram on an OSM bridge way any deck)A lower line passing under another's bridge jumped 7–8 m onto it and back (16 of 109 rail–deck crossings in Dresden; the flyovers north of Neustadt), and an upper line dropped into the DGM's gap wherever it ran longer than the deck's outline — 230 steps beyond the grade by > 1 m in Dresden.The level is decided along the whole line within its grade (✅ Line levels, ADR 0041).
Per-line ballast ribbons (rail v1: one ~9.6 m ribbon per ver03_l line)42+ overlapping coplanar ribbons in the yard z-fought into ragged/torn edges.Replaced by the dissolved ver03_f area as one merged surface.
ver06_l centreline-buffered decks (rail v1)Buffered planks stacked deck-top + ballast + parapet-cap → "2-story" bridges, and one plank merged the parallel Marienbrücke spans.Replaced by ver06_f deck polygons (one slab per real footprint); kept as the no-ver06_f portability fallback.
Per-tile rail layer (rail v1)Each tile's own heightAt returned null off-tile → tracks truncated at every seam.Build on the cross-tile heightAt — once for the block until ADR 0024, now per fine terrain tile over every loaded terrain.
LoD2 bridge objects as buildings (ALKIS 53001_* in the CityJSON)Saxony's LoD2 carries its bridges as Buildings (52 on the first four tiles): flat 1 m slabs at one height. The building bake drew them as clay houses (with the facade detail) 1–3 m off the rail layer's decks — every Elbe bridge had a second deck.withoutTrafficStructures drops them from the building mesh; the rail layer draws bridges alone. Their footprint is no better than the DLM's.
Deck height from the DGM abutment ramp aloneTwo endpoints + camber: the Waldschlößchenbrücke's deck came out 3 m low (below its own navigation clearance), and rail decks whose ends sit under catenaries or trees up to 6 m high.The roadway measured in DOM1 along the axis, held near the ramp (✅ above); the ramp stays the no-DOM fallback.
Per-tile ground for bridge decksThe rasters stopped at the tile edge (sampling clamped there), so a deck across a seam got two different profiles — and both tiles drew it: two decks metres apart (Schlachthofbrücke).A DGM/DOM mosaic around the tile; the owner draws, every copy lifts rails.
ver06_f-only bridge decks (rail v2 first cut)ver06_f has area polygons only for (mostly rail) major spans → road/path bridges (Augustusbrücke etc.) vanished + everything mis-classified rail.Drive from the complete ver06_l set, footprint from ver06_f where matched.
Motion-gated SSAO (plan 007 as first shipped)The contact shadows blinked on every footstep — reads as a bug, not a saving.The AO pass runs permanently at half resolution (N8AO's halfRes then, GTAO since plan 020); only DoF is dropped while moving (ADR 0009).
Cloud shadows / per-frame shadow updates for wind swayWould force the rings' depth passes every frame over tens of thousands of trees, undoing the on-demand shadow map.Sway and cloud drift run in the main pass only; the cast shadow stays static (ADR 0009).
Multi-echo ratio as the low-vegetation cue (LSC, low vegetation)Measured: only 26 % of OSM-hedge pixels reach echo ≥ 0.5, fences 56 % (AUC hedge-vs-fence 0.28); recall 25 % of observable hedge length vs 58 % for the NDVI + intensity rule.Keep it as the tall-vegetation cue (100 % of > 5 m forest vs 3 % of roofs) and as a qualifier of intensity.
Blob-wide "multi-echo ≈ 0" for small structures (plan 034 as planned)A shed's roof edge splits the pulse (part roof, part ground), so no blob's multi-echo share is ≈ 0: 4 blobs on the spawn tile at ≤ 0.05, 43 at ≤ 0.10 (median over all band blobs 0.93); the interior share admits 377, 112 of them against tree crowns.The rule works per cell: no multi-echo return in a cell's 3 × 3 window, the rim grown back after (the ✅ "Small structures from the laser scan").
OSM scrub polygons filled with shrubs (OSM-only tiles)A jittered 3–4 m grid gave 4–9 k shrubs per neighbour tile — more than the laser scan finds on the primary — mostly under existing crowns.OSM-only tiles get the OSM hedges only.
Laser-scan-only hedges (LSC, the 478 unmapped "hedges" of the low-vegetation bake)~30 % of the low-vegetation mask lies 1–2 m from a > 3 m crown: crown-rim false positives a first-surface model cannot tell from understory, and they read as stray hedges along tree rows.Shipped since 2026-10-21 where they pass the filters of Hedges above (long, slim, clear of crowns, green); the rest (short, wide, along crown rims, not green) stay research-only (--research).
Shrubs (LSC blobs + OSM natural=shrub nodes, 3 226 on the primary)Same crown-rim false positives, and the lobed dome reads as a faceted grey "boulder" at arm's length.Kept in the bake behind --research; a better shrub shape is shape polish, not data.
Camera-follow grass tuft ringShadow-casting instances rewritten every frame; reads as confetti.Meadow mottle + normal perturbation in the terrain shader (✅ above).
Per-lamp real point lightsthree bakes the light count into every program (since plan 020: every light into every lit node build's key) → a recompile storm on every add/remove, plus per-light cost.A fixed pool of 3 real lights (1 on a phone) retargeted to the nearest heads; every other lamp is emissive + sprite (ADR 0009).
Plain (non-shadow-gated) foliage translucency, quad leaf billboards, selective bloomNoise at instance distance / no payoff for the cost.None; the gated shimmer and translucency went too (Backlit crown shimmer and translucency).
Baked RGB splatmap (landcover_rgb_<tile>.png: RGB = pastel palette, A = water coverage, plus a 2048² variant)The look lived in the bake: a colour change meant re-baking every tile, and the palette was spelled three times (bake, minimap, shader fallback). Its alpha was data, so every resize had to split colour from alpha — sharp premultiplies alpha across a resize, which once turned every land texel black (ADR 0010, superseding ADR 0010).The bake writes class ids only; the one palette (lib/city/landcover.ts) is painted on the GPU at load. A per-fragment palette lookup was rejected too: class ids cannot be mipmapped, so far boundaries would alias.
Overpass-based OSM bakes (lamps, platforms, bridge structure)Live queries: rate-limited and not reproducible, and one more way of reading OSM next to the local extract the walls already used (ADR 0037).Every OSM layer comes from one local Geofabrik .osm.pbf via GDAL's OSM driver (pipeline/bake/osm.py). The committed lamp/platform/bridge-structure files still predate this; the next re-bake moves them (data/dresden/provenance.json).
Bash bakes (scripts/extract-*.sh + Python/Pillow heredocs)Three languages, string-built paths, tile names and CRS spelled per script; numpy and gdal_calc.py were missing, so raster maths was written around Pillow (ADR 0037).One package, pipeline/bake/, on numpy/rasterio/pyogrio/shapely in a uv environment, driven per site tile by bun run bake.
uint16 heightfield + custom building vertex-stream codecs (<tile>.heightfield-<n>.json + .u16.gz; vertex stream + meta JSON)Private formats with a codec on each side that no other tool could open, and the browser still burned the wall breaklines into the grid at load (ADR 0024).Standard glTF 2.0 (meshopt + quantisation) with an EXT_structural_metadata table; the bake does the breaklines. Costs wire size (≈1.0 + 1.1 MB → ≈1.4 + 1.5 + 0.4 MB per tile); a heightmap-PNG custom content type is the fallback if that ever matters more than tooling.
The terrain TIN's own payload (dgm1_<tile>.tin-<cm>cm.json + .bin.gz: byte-split delta planes + varint triangles; the prototype, ±0.25 m on the neighbours)A private format again, next to the glTF every other content is (ADR 0024), and two terrain paths in the viewer.The fine level's glTF is the TIN (reordered for meshopt, extras.tin), ±0.15 m on every tile — 20–27 % smaller than the 1024² grid it replaced (ADR 0030).
Fixed 2×2 block loaded at boot (one primary tile + three neighbours, never unloaded)The world was bounded by boot cost, nothing ever unloaded, and collision/demolish stopped at the primary tile's edge. The hand-written tile manager proposed instead (ADR 0024, never accepted) would have re-built the schedule, LRU and worker 3DTilesRendererJS already has (ADR 0024).The site streams as a 3D Tiles tileset; distance, not a role, decides which tile is detailed.
Inserted building (a glTF model, else an orange marker box, dropped at a fixed Dresden spot; once on the B key)No caller since the B key was disabled — dead plumbing through the HUD, the scene and a Dresden-only constant; removed with the site config (ADR 0037).Revisit only with a concrete use (e.g. a planned building to preview).

Art-direction north star: a soft, illustrative watercolor / contour-map look (paper grain, depth grading, vignette, contour-line terrain). Every transformation above is judged against it — additions that read as hard-edged, photoreal, or "modern office" get rejected here.

Previous
Data → feature provenance
Next
Portability — rendering any location, with whatever data it has
Source on GitHub: docs/transformations.md
On this page
  • ✅ Active
  • Geometry & ground
  • Building detailing (all keyed off CityJSON attrs + the loader's surfacetype)
  • The twin: identity, facts and provenance (asked on demand)
  • Vegetation
  • Railway & bridges
  • Walls and fences
  • Wall → terrain conflation (breakline burn at build time)
  • Stairs (OSM steps over a lowered terrain)
  • Lighting
  • Atmosphere & time of day
  • Traffic (the data layers)
  • Planning views (Modell, plan 055 / ADR 0044)
  • 🧪 Experimental
  • 📋 Planned (designed, not built)
  • 🗃️ Discontinued / rejected (do not retry blindly)