Fix table-to-scan alignment and preserve unresolved exports
Repeated, irregular and unlabelled printed-page sequences caused the tables config to attach some exported tables to the wrong scan. This PR matches each existing table against its report's page ALTO, separates tables that had been conflated onto one image, and preserves ambiguous items explicitly.
- 451 table items move to different scans; 3,529 retain their original scan; 106 remain unresolved in a side artifact with their original CSVs.
- Repaired rows: train 2,520, validation 302, test 327 (3,149 total; 3,980 attached table items). All 4,086 original IDs/CSV payloads and report split assignments are conserved across the config and unresolved artifact.
- Three replacement Parquet shards, source/decision ledgers, a local-only UV reproduction script, and corrected card wording. The default corpus and licences are unchanged. This is an alignment repair, not a transcription or image-resolution upgrade.
Validation covers all 288 reports/24,052 page ALTO documents, schema/provenance/split/payload invariants, all embedded JPEGs, and replay of the published scorer. All six previously identified wrong-scan items are corrected and 24 previously aligned items retain their scans. A fresh risk-enriched sample of 58 tables plus eight low-margin train/validation moves was inspected against source images. Two low-margin items were retained through explicitly documented agent visual adjudications; thresholds were not tuned to rescue them.
The card now states the limits clearly: ALTO agreement and agent spot checks do not make these machine-extracted CSVs gold labels, and the attached tables may not exhaustively annotate a page. See audits/table-alignment-2026-09-09/README.md for counts, thresholds, sampling caveats and exact per-table decisions.
Base revision: 1a726bbb3a8dc8f3499cf2db0a983865ab6fd2b6. Open for dataset-owner review; no merge is performed.