NY: download_bulk_csv output (whole-view + display headers) is pipeline-incompatible; download_resource_csv is the real primary path

Summary

io/ny/acquire.download_bulk_csv is documented as "the primary path" for NY acquisition, but its output cannot feed the Phase-2 pipeline as-built. Discovered during the Phase-3 real pull (ny-disclosure-explore).

Two concrete problems

1. Whole-view, not year-scoped. It hits /api/views/<id>/rows.csv?accessType=DOWNLOAD, which dumps the entire view. qym9-xzj6 ("Client Semi-Annual Report Beginning 2019") is 66.9M rows across 2019–2026 — 2025 alone is 11.2M. There is no $where on that endpoint, so "pull 2025" is impossible through it. The whole-view export is 41 columns wide and streams ~0.9–3.2 MB/s → roughly ~55 GB / multi-hour, and _build_grain does a single in-memory pd.read_csv (infeasible on a 16 GB machine).

2. Display-name headers, not field names. The bulk export header row is human-readable display names (Form Submission ID, Type of Lobbying Focus, Current Period Compensation), not the SODA field names (form_submission_id, type_of_lobbying_focus, current_period_compensation) that columns.COLUMN_MAPS and grain.collapse_to_filing_grain require. Feeding bulk-export output to the pipeline would KeyError in grain-collapse. The NY test fixtures were captured from /resource/<id>.json (field names), so the suite is green while this real-data path was never wired end-to-end — the egress block hid it.

Resolution (this branch)

Added acquire.download_resource_csv + resource_csv_url (SODA /resource/<id>.csv endpoint, supports $select/$where/$order/$limit, returns field-name headers), with 7 behavior tests. Phase 3 pulls 2025 via that path: 9 projected columns (only what the pipeline consumes), $where=reporting_year='2025' → ~3.17 GB, pipeline-compatible headers.

Follow-up to decide

• Deprecate or repurpose download_bulk_csv? It still works as a generic whole-view dumper, but it's mislabeled as "primary" and its output is pipeline-incompatible. Options: (a) demote its docstring to "raw whole-view archival dump only, not pipeline input"; (b) remove it if nothing needs a whole-view dump; (c) add a display-name→field-name header remap shim if a whole-view path is ever wanted.
• The single-request download_resource_csv stream has no mid-download checkpoint (resume-skip is whole-file). Fine for a re-fetchable static dataset, but if larger pulls prove flaky, add keyset/range pagination (note: form_submission_id is non-unique ~1,300×, so any paging must slice by id range, never split a filing across a page boundary).

Start in io/ny/acquire at download_bulk_csv, download_resource_csv, and resource_csv_url, then inspect columns.COLUMN_MAPS and grain.collapse_to_filing_grain. Review the seven resource CSV behavior tests and the Phase-3 pull configuration. Done means deciding and documenting whether the bulk path is deprecated, removed, or adapted without breaking the pipeline.