world/regions/south_africa/terrain/SOUTH_AFRICA_TERRAIN_PASS_2026-08-15.md
Point-in-time record of the first real terrain build for region south_africa, the eleventh of the
twelve vertical-slice regions. Generator: Tools/compose_south_africa.py in the game repo
(C:/dev/Humanity/Humanity). The composed raster, its meta and its zone manifest live in the game
repo; this file carries the findings that belong to CANON rather than to the build.
The lane stopped before the editor import by instruction. Nothing here has entered UE.
A composed compressed transect with an L-shaped ocean on its west and south edges: 6631 x 5101 posts
at 2.0 m/vertex, 13.26 x 10.20 km, one Landscape actor at 255 quads x 1 section x 26x20 components.
All eight canon zones are real Copernicus GLO-30 patches placed by CH_13 route order, not by geodesy.
percent land inside its patch.
source centre: 0.86 to 1.11, with zero zones outside the 0.8 to 1.3 band.
Measured before any terrain was sourced, with ESA WorldCover read at each canon centroid
(QA/bench/south_africa_2026-08-15/centroid_landcover.json in the game repo). Three of the eight
zone_bounds_wgs84 cells in the Zone Catalog do not carry the landcover their own terrain column
claims. None of these was silently sampled; each is recorded here for the page's next pass.
The catalog gives this zone terrain water_body and the bounds w=30.90;s=-16.20;e=31.00;n=-16.10.
A 2.1 km WorldCover window on that centroid reads 53 percent shrubland and 42 percent grassland, and
zero permanent water. The nearest real Zambezi reach is about 25 km north-north-east. The terrain
pass sourced the patch at 31.080, -15.960, which carries 24.2 percent permanent water at the 900 m
patch radius, and declared the offset in the manifest. The page-authored anchor should be re-sited
onto a real reach, because a water_body zone whose coordinates contain no water is the same defect
class as the congo_river_mainstem centroid that read 100 percent tree cover.
The catalog gives this zone terrain coastal and a band from -34.30 to -34.10 latitude. That band is
Overberg wheatland: 98 percent cropland, zero water, and 18 to 20 km inland of the shore the page
describes it as. The terrain pass sourced at 20.900, -34.370, which carries 25.6 percent permanent
water lying south. The band's latitude range should move south to the real Overberg shore.
The canon centroid carries 1.9 percent water in a 2.1 km window and 0.1 percent at the 700 m radius
first tried; Table Bay lies 1.0 to 1.4 km north-east. This one was NOT offset, and the reasoning is
recorded because the decision could reasonably have gone the other way: the canon centre is clean
quay-and-street ground with no stranded harbour basin, and the water this zone needs is supplied by
the composed west ocean its patch edge sits on. Every candidate carrying real water carried it to the
north-east, which is the one side this tile's L-ocean cannot reach, and a real basin stranded a
kilometre inland of a composed shoreline is a worse artifact than a composed bay. The note stands
against the page's own area_bounds cell, which describes a quay-and-warehouse frontage.
Both were caught by controls, not by eye, and both are ported fixes other lanes may need.
1. THE SHORE-BLEND BAND WAS EATING A CANON LANDFORM. The shoreline is made by scaling land toward
zero across a blend band, so every metre of that band is real terrain flattened into beach. At the
inherited 260 m the band reached 92 m inside blombos_cave's control window, and the orientation
control scored that zone 0.406 against its own source while failing to beat a mirrored copy of it.
The zone canon describes as a sea-cliff cave was composing as a ramp. At 90 m it scores 0.995.
2. THE PATCH FEATHER WAS FLATTENING THREE ZONES. A patch carries full relief only inside
(1 - FEATHER_FRAC) times its radius. At the inherited 0.60 that is the inner 40 percent, so the
800 m audit box — whose corners sit at 566 m — measured the composition's own attenuation rim at
every zone smaller than 1415 m radius. Measured: great_zimbabwe 0.59x, cape_colony_port 0.62x,
both FLATTENED rather than inflated. At 0.25, with the two smallest patches sized up to 800 m, all
eight zones land inside the band. This is a third distinct cause of a bad relief ratio, after the
base-core fraction and the datum compression, and it is the first that flattens.
OpenTopography's globaldem endpoint is rate limited to 50 calls per rolling 24 hours on a
non-academic free key — the vendor's own figure, not a house guard, so the limiter constant in
Tools/terrain_fetch_preprocess.py is correct and must not be raised. This lane spent 5 live calls
and then hit the ceiling, because 37 of the 50 belonged to the previous day's wave and were still
13.7 hours inside the window. The remaining source ground was read from the AWS Open Data public COG
collection of Copernicus GLO-30 (s3://copernicus-dem-30m), which is the same product,
unauthenticated and unmetered. A rate limit on one door is not an absence of the data, and the
helper is worth promoting into the shared chain so the next lane never waits on the window.
COP30's native ground sample is about 30 m; this tile is authored at 2.0 m/post, so the DEM is
upsampled 15x. The 31 m-scale landform is real measurement and sub-30 m detail is interpolation.
Against the Zone Catalog's own meters_per_vertex column: four zones ask 1 m and were served at
2.0 m, one asks 2 m and was served exactly, two ask 5 m and one asks 10 m and were served finer. The
1 m ask cannot be honestly served from COP30 by any tile, since it is a 30x upsample of the source,
and the page should either accept the 2 m floor or name a finer source.
HL_0007. This region carries five protected zones — drakensberg_ukhahlamba_amphitheatre,
blombos_cave, kalahari_gwi_reaches, cape_colony_port and great_zimbabwe — and six of eight zones at
care_tier elevated. The terrain pass emits ground elevation and landcover class only. No painted
shelter, no Great Enclosure or Hill Complex wall, no stele, no Bo-Kaap frontage, no mission station
and no mine headgear is modelled anywhere in the outputs. Registered and protected cultural property
is referenced by coordinate and name and is never reproduced. No generative asset runner was run.