[ivorbosloper]

es now converts each of its 50 provincial GeoPackages on its own and merges the parts, because the whole of Spain does not fit in memory in one run. Closes #336.

• PerFileConverterMixin (fiboa_cli/conversion/per_file.py) runs the normal convert once per source file into a temporary part, then merges the parts with vecorel-cli's DuckDBBaseConverter.merge_parquet, which checks the ids and Hilbert-sorts over the whole set. With a single source file it is plain convert. It replaces the 450-line PerFileBaseConverter of PerFileBaseConverter: convert file by file, then merge #281.
• The merge runs on a standalone DuckDBBaseConverter rather than mixing that class in, so the converter's own convert chain stays as it is.
• Full 2025 edition (17.8 million rows): peak RSS 10 GB, against more than 27 GB before the one-run attempt was stopped; the output validates.
• Not in this PR: in several provinces about 0.1% of the rows repeat an id (the SIGPAC key plus the declaration line). They get the ~n suffix, as before.

🤖 Generated with Claude Code