fio-l2-cairo-mixedgeom-split

Analyst notes

Description

Tests whether the agent composes four format operations in the right order (stratify by geometry type, explode multi-parts, reproject, write a multi-layer GPKG) without losing the cross-layer site_id link. The prompt names neither the EPSG nor the operations, so the agent has to read 'Egypt's national grid' as Egypt Red Belt (EPSG:22992), understand that 'single-part' means explode the MultiPolygons, and pick a multi-layer GPKG over three separate files.

Approach

1. Read the input GeoJSON and notice it mixes Point, LineString, Polygon, and MultiPolygon features in one FeatureCollection, all sharing site_id.
2. Split the features by geometry type into three groups, routing each kind to its named layer.
3. Explode the MultiPolygons in the polygons group into singleton Polygons, keeping site_id on every part.
4. Reproject every layer from WGS84 to Egypt Red Belt (EPSG:22992), the Egyptian national projected grid for Cairo.
5. Write all three groups as named layers in a single GeoPackage file called heritage.gpkg, carrying site_id and the other attributes through on every feature.

Pitfalls

• Stratifying without exploding ships MultiPolygons into the polygons layer, so the desktop tool rejects the file and both the singletons-only and the heavily weighted post-explode count subchecks fail.
• Picking UTM 36N (EPSG:32636) is defensible for Cairo as a generic metric pick and is accepted by the soft-CRS format gate, but it costs the crs_is_canonical subcheck because the regional canonical is Egypt Red Belt.
• Leaving the output in WGS84 (no reprojection at all) is still accepted by the format gate under the soft-CRS policy, but both CRS subchecks fail.
• Reprojecting with the x/y axes swapped lands every feature several kilometres off while staying in the right coordinate-magnitude range, which slips past the format gate and then fails the three heavily weighted geometry subchecks (polygon IoU plus the per-site point and line agreement).
• Writing three separate files (heritage_points.gpkg etc.) instead of one multi-layer GPKG fails the format gate on missing layer names.
• Dropping the site_id column on any layer fails the format gate outright because the cross-layer link is the deliverable the persona needs.

Map

Recent runs task v3