Ingest - archive datasets that haven't been run

33 datasets have an ingest template but have never been archived through ingest. Their latest/ is still entirely library output, so every consumer reads library data despite the template count saying they're migrated.

Running ingest for them is not a bulk operation. Archiving switches latest/ to ingest output, which silently changes which file a consumer resolves to, and the geometry column name often changes with it. That is what broke facilities, cpdb, green_fast_track and knownprojects in nightly QA after dcp_mappluto_wi was archived: nine references to b.wkb_geometry sat several lines above their LEFT JOIN dcp_mappluto_wi AS b, so a grep for the dataset name next to a geometry column found none of them.

7 are pinned by a consumer, so they fail loudly rather than silently. All in products/facilities/recipe.yml with file_type: csv: dcla_culturalinstitutions (see #2623), dcp_sfpsd, dohmh_daycare, fdny_firehouses, hhc_hospitals (see #1431), sbs_workforce1, uscourts_courts (blocked upstream, #1326).

26 have no pin, so archiving switches them to parquet with no warning. 17 of those are geospatial, and the templates disagree about the geometry column:

Several have 3 to 5 consuming SQL files. Each needs its consumers checked before archiving, and reading each join block rather than grepping.

9 are non-geospatial and low risk, with no geometry column to mismatch: council_members, dcp_pop_acs2010_demographic, dcp_pop_acs2010_economic, dcp_pop_acs2010_social, fbop_corrections, nycdoc_corrections, nysdoccs_corrections, nysed_activeinstitutions, sca_enrollment_capacity. The first four have no consuming SQL at all.

Suggested order: the 9 non-geospatial, then the 7 pinned ones (whose pins can come off once an archive exists), then the 17 geospatial one at a time.

admin/ops/clear_library_files_from_latest.py reports the current set as needs_ingest.

Run admin/ops/clear_library_files_from_latest.py to confirm the current needs_ingest set, then inspect products/facilities/recipe.yml and each listed dataset's consuming SQL join block. Archive the datasets in the suggested order, checking geometry columns and pinned consumers one at a time; done means all 33 datasets use ingest output without nightly QA breakage.