Give a small GeoParquet file at least ten row groups

[ivorbosloper]

A file under 250,000 rows is now written in at least ten row groups instead of groups of 25,000 rows; larger files keep 25,000. With five or six groups, each one covers most of a small file even when its rows are Hilbert-sorted, so a reader can skip none of them, and Portolan's PTL-DAT-006 rejects the file.

• Found on the canton of Zürich (100k–130k rows): sorted correctly, but its five or six groups each spanned most of the canton, because the canton straddles a quadrant of the CRS-wide Hilbert grid. At ten groups the same rows pass.
• GeoParquet.row_group_size_for(num_rows) is the one rule; both write paths use it. It rounds down to whole 2,048-row groups, as DuckDB writes whole vectors, so a file under 20,480 rows gets fewer than ten groups.
• The DuckDB path learns the row count only after writing, so it applies the size in its final rewrite. It rewrites a small file once even when the file is already sorted or cannot be sorted.

🤖 Generated with Claude Code