CLOUD-HANDOFF.md

Reproducible bounded build and pilot fixtures

The full pinned USCIS build has completed and passed source-preservation and released Python/R checks. See README, VALIDATION and uscis.manifest.json for the exact full artifact and scope. Historical handoff and pilot receipts are retained separately; their earlier pending-build wording does not describe the current full snapshot.

1. Use Python 3.11+ and python -m pip install -r requirements.txt to obtain pinned DuckDB 1.5.6 in a suitable environment. Ingestion needs only DuckDB and the standard library; no Arrow, spatial extension, credentials or datapond-build dependency.
2. Run python acquire_sources.py for disk/missing-source preflight. This does not access the network. python acquire_sources.py --download fetches exactly the two immutable public files, 208,298,935 bytes total, and checks full lengths and SHA256. Correct existing files are reused; incomplete parts and wrong files never become verified inputs.
3. Run python build_database.py --lock source-lock.json --output output/uscis.duckdb using a fresh output path. The build uses one thread, 2 GiB memory, 4 GiB spill in its private directory, 8 GiB total work cap and 2 GiB physical free reserve. Preflight requires about 10.2 GiB free including source staging. Spill remains under the chosen output parent. Do not silently lift limits after a failure.
4. Review the complete manifest: exact rows/fields/ordinals, all NULL counts, duplicate flags/source keys, missing identifiers and date ranges/anomalies. The exact comparator processes 4,096 rows per batch and fails on changed values, missing/extra rows or noncontiguous source order.
5. Run python -m unittest -v test_pipeline. The supplied first 2,048 rows per table are integration fixtures, not representative data. For a sample-only build use python build_database.py --lock pilot/source-lock.json --output output/pilot.duckdb --pilot. The explicit flag must match the lock and cannot label the sample as a full dataset. Pilot hashes and full expected identities remain separate.
6. With released Python/R Datapond packages and preinstalled HTTPFS available, run python run_smoke.py --db output/uscis.duckdb --python /path/to/python --rscript /path/to/Rscript --r-library /path/to/R/library. Omit --pilot for the full snapshot. Use a fresh source copy for successive archived test runs, because the harness writes fixed evidence filenames and fresh private registry caches.

The builder refuses an existing final database, WAL or manifest, builds privately, checkpoints/closes, checks disk limits synchronously, and then uses an atomic no-overwrite hard link. Hard-link support is required. If a process fails between linking the DB and writing its manifest, inspect that partial outcome; do not overwrite it or treat it as a complete release. Failed/private intermediates are retained for diagnosis. Changing a source pin requires a new full validation and independent audit, not just a successful download.

Use acquire_sources.py/build_database.py as the portable entry points. Historical acquisition scripts and private run receipts are archived separately from this public export. The included fixtures do not need network access for regression tests. The original, first imported and corrected pilot artifacts remain in the archived handoff evidence and are not release candidates.