geo: acquire the 3DEP 1 m DEM window in EPSG:26911

550x350 m clipped to the block, already in the project CRS, 984 KB, committed
through LFS. Verified against the service's own point query and the earlier
probe: 115.911/119.422/117.593 m, agreement within 6 cm.

Adds Geo/scripts/dem_stats.py, a stdlib tiled-TIFF reader, because GDAL cannot
run on the host and the exported raster is tiled rather than strip-organised.

Geo/verification/2026-11-06-elevation.md

@@ -42,8 +42,15 @@ Consequence for `0.1.0`: the demo cannot use a flat ground plane. The street and
 
 Points inside the hero footprint returned 114.53–117.13 m — a 2.6 m spread within one building outline, which is not ground reference. 3DEP can carry structures and voids there. Those four diagnostic points are excluded from the relief figure; use the street and sidewalk lines, not building interiors, for any ground surface.
 
+## DEM acquired
+
+Downloaded 2026-11-06 from the 3DEP export service, already in EPSG:26911, clipped to E 376 100–376 650 / N 3 774 050–3 774 400 at 1 m. Manifest, checksum, and full request: [`../manifests/2026-11-06-3dep-dem.md`](../manifests/2026-11-06-3dep-dem.md). File is committed — it is 984 KB, public domain, and the blockout needs it.
+
+Cross-check at three points: raster 115.911 / 119.422 / 117.593 m against the service's own point query 115.910 / 119.418 and this document's probe 115.911 / 119.418 / 117.53. Agreement is centimetres, so the georeferencing holds.
+
+Window statistics: min 110.46 m, max 127.73 m, spread 17.28 m — building tops and the ground rising north of the boulevard. The **walkable relief stays 3.94 m**; the demo ground mesh follows the street, not the window envelope.
+
 ## Next
 
-1. Download the 3DEP tile(s) covering the demo bounds, keep the source file (size is larger than the OSM extract; decide commit vs local-only with a manifest and checksum at that point).
-2. Reproject to EPSG:26911, clip to the block plus margin, and record the vertical datum and the clip bounds.
-3. Sample the kerb line and the crossing ramps specifically — the grid above is a street-level average, not a kerb profile.
+1. Sample the kerb line and the crossing ramps specifically — the grid above is a street-level average, not a kerb profile.
+2. Build the ground mesh from the DEM along the street, not from a flat plane, and assert the 3.94 m fall survives the Blender→Unreal path.