Tests whether the agent can run a multi-step polar pipeline (identify Antarctic research stations from a noisy Overture POI layer, draw a 200 km buffer, clip to land, union overlapping zones into coalitions, and separately intersect the over-water portion against water and bathymetry) while picking the right CRS for the region. The hidden gotcha is that a planar 200 km buffer at 70-80 degrees south is severely distorted, so the buffer has to be drawn geodesically (or in a polar projection that preserves area near the pole), and the output CRS has to be the polar stereographic family rather than WGS84 even though the prompt does not name it.
Approach
Fetch Overture `places.place` south of -60 latitude and work out which features are research stations, since the theme has no `research_station` category and a lot of the southern POIs are unrelated businesses that need filtering out.
Fetch `base.land`, `base.water`, and `base.bathymetry` for the same region so the buffers can be clipped to the landmass and the over-water portion attributed.
Draw a 200 km buffer around each station in a way that is honest about polar distortion, either geodesically or in a polar-stereographic projection appropriate for Antarctica.
Clip the buffers to the landmass, then union the overlapping clipped zones into coalitions and tag each feature with its `station_id`, `station_name`, and integer `coalition` id.
Compute the over-water portion of each station's buffer separately and intersect it with the water and bathymetry features, carrying through the water attributes plus a `water_source` column literally equal to `base.water` or `base.bathymetry`.
Write both results as GeoParquet with MultiPolygon geometry in the polar stereographic CRS appropriate for Antarctica.
Pitfalls
Buffering 200 km in WGS84 degrees (or in a non-polar metric CRS) gives a heavily distorted ellipse rather than a proper circle at high southern latitudes, so the sphere areas come out far too small and the `geodesic_buffer_check` and `buffer_area_reasonable` subchecks both fail.
Writing the outputs in EPSG:4326 instead of the polar stereographic CRS for Antarctica fails the CRS, coordinate-range, and area subchecks when the grader sees degree-range coordinates where it expects metres.
Skipping the station filtering and dumping all 2300+ Overture POIs south of -60 (gas stations, shops, hotels, and so on) blows up the station count, fails the name-overlap subcheck, and inflates every area subcheck.
Going the other way and filtering too aggressively, ending up with fewer than five stations, fails the minimum-count subcheck.
Assigning every station its own coalition id (no overlap detection) fails the `coalition_exists` subcheck, which is what trips the planar-buffer broken solution too because tiny buffers do not overlap.
Skipping the land clip and leaving the raw geodesic buffers in the spheres output inflates the total area well past the ±40 % tolerance, since the buffers extend far over the Southern Ocean.
Labelling the `water_source` column with anything other than the literal strings `base.water` and `base.bathymetry` (for example `water` and `bathymetry`, or the input filenames) fails the water-source attribution subcheck.
Producing the water-overlap file without actually intersecting against water or bathymetry geometries leaves either an empty file or wrong areas, so the water subchecks all fail.