test(integration): geometry results in the parity harness; RS_WorldToRasterCoord + RS_RasterToWorldCoord catalogs

[james-willis]

Unblocks geometry-returning raster functions in the spark-parity suite and uses that to catalog the RS_WorldToRasterCoord / RS_RasterToWorldCoord combined forms — including the missing point-geometry overload.

Harness: geometry results (first commit)

Probing showed the two sides fail asymmetrically:

• SedonaDB needed nothing: its geometry results already arrive as geoarrow.wkb and result_to_tuples already renders them as WKT.
• Sedona Spark was the broken side: DataFrame.toArrow() ships a GeometryUDT column as the UDT's internal serialization bytes — not ISO WKB — which the tuple path then tried to cast to string ("Invalid UTF8 payload").

SedonaSpark.result_to_table now converts geometry columns to WKB in the JVM (ST_AsBinary, transport-only — the same pattern decode_raster_result uses with RS_AsGeoTiff; the shared test SQL stays RS-only) and re-tags the bytes as geoarrow.wkb, so both engines render WKT through the same geoarrow formatter. Columns are renamed positionally before the select because Spark's generated names (rs_worldtorastercoord(rast, 104.0, 494.0)) contain dots that every name-based lookup re-parses as a nested path; the original names are restored on the Arrow table.

Everything downstream (result_to_tuples, assert_result, WKT expected= anchors, compare()) works unchanged. Full suite on this branch: 174 passed, 84 xfailed — main's counts plus exactly the 7 new xfails.

RS_WorldToRasterCoord (second commit)

All probed, every case an xfail stating both observed behaviors:

• 0- vs 1-based (RS_RasterToWorldCoord*/RS_WorldToRasterCoord* are 0-based; Sedona Spark, PostGIS, and RS_PixelAs* are 1-based #1235, now in geometry form): the origin answers POINT (0 0) from SedonaDB and POINT (1 1) from Sedona Spark; interior points likewise one pixel apart.
• Outside the grid the offset is not even uniform (new finding): (90, 505) sits −5 columns and −5/3 rows from the origin. SedonaDB truncates the fraction toward zero (POINT (-5 -1)); Spark floors before its 1-based shift (POINT (-4 -1)) — the rows agree by coincidence, the columns differ. A positive-fraction probe ((105, 493) → 2 vs 3) confirms the rules are truncation vs floor+1, not something else.
• Missing point-geometry overload: RS_WorldToRasterCoord(raster, point) raises "No kernel matching arguments" on SedonaDB; Spark accepts it. The xfail notes RS_RasterToWorldCoord*/RS_WorldToRasterCoord* are 0-based; Sedona Spark, PostGIS, and RS_PixelAs* are 1-based #1235 still applies once the overload exists.

RS_RasterToWorldCoord

The input side of #1235: (1, 1) answers POINT (102 497) from SedonaDB (0-based input) and POINT (100 500) — the origin corner — from Spark (1-based input).

Follow-ups this unblocks

The remaining geometry-returning functions can now get parity files the same way: RS_Envelope, RS_ConvexHull, RS_PixelAsPoint / RS_PixelAsCentroid / RS_PixelAsPolygon. (Item-level-CRS geometry outputs may need another look — these fixtures are CRS-less.) The test_rs_worldtorastercoordx/y docstrings still say the combined form is "deferred"; updating that line is left out here because #1262 touches the same sentence — whichever lands second gets the one-line fix.

[paleolimbot]

This one can be a module level import (since pyarrow already is)

Also, isn't there already a "to geoarrow"? I PRed one into Sedona a few years ago but maybe it's gone now.

[james-willis]

Follow-up pushed (f7ebe5c): the hand-rolled ST_AsBinary + re-tag path is gone in favor of sedona.spark.geoarrow.dataframe_to_arrow, which was indeed already there and does the EWKB conversion + geoarrow tagging (with CRS inference as a bonus). The geoarrow.pyarrow import disappears with it.

One wrinkle: dataframe_to_arrow resolves columns by name (list(df) → __getitem__), so it fails on Spark's dotted generated names like rs_worldtorastercoord(rast, 104.0, 494.0) — filed as apache/sedona#3351 with a repro. Until that lands, the harness renames columns positionally around the call and restores the names on the Arrow table. Full suite unchanged: 174 passed, 84 xfailed.