shared-exports

Date: 2026-09-16. Status: implemented.

Decision. Every build writes a gzipped CSV of each table beside the Parquet, and one contract, simd_ingest/export_contract.yaml, governs both it and the SQL examples. The contract withholds GridReferenceEasting and GridReferenceNorthing from both tables and Latitude and Longitude from the history table, so neither the CSV nor a linkage result carries a coordinate; the Parquet keeps every column and its schemas are unchanged. CSV is comma separated with RFC 4180 quoting, UTF-8 without a byte order mark, LF endings and gzip written with no name or timestamp, so the same rows give the same bytes. Dates are YYYY-MM-DD, is_current is 1 or 0, and all other values are written exactly as stored. A null and a source blank are the same empty cell, resolved by table, column and record type, which the dictionaries state and the contract records. The build also records a row digest per table, computed from the accepted table rather than from a re-read of the file, so a database load can be checked after import. Each CSV is read back cell for cell before publication; results/CSV_README.txt carries every publisher's attribution; the manifest is written last and records every hash; audit recomputes them.

Reason. The CSV is how the table is shown to the team, so it is the copy most likely to leave the building and the one that should carry the least. Withholding coordinates is a project data-minimisation choice first: a linkage result is one row per patient, and a grid reference is a precise location where a data zone is not. It is also conservative about licensing, since NRS restricts "postcode boundaries and grid references" separately from its openly licensed index and lookups but does not say which governs a coordinate column inside an index file. That reading is unconfirmed and is on the list of questions for Research Data Scotland. Excluding the columns from the SQL as well keeps one policy rather than a footgun for anyone who imports the CSV under the table name the queries use.

Evidence. tests/test_exports.py states the renderings independently of the writer and covers the empty-cell rules, leading zeros, quoting, non-ASCII and supplementary characters, determinism, structural damage and the corruptions a load introduces. Reserved framing characters are rejected in source text rather than stripped. On SPD 2026/2 and SSPL 2026/2 the CSVs are 8.3 MB and 7.2 MB gzipped, and both rows-only fingerprints are unchanged; the Parquet file hashes change because this entry changes the decision log.