- Re-extracted the 13 chunks with paraphrased source_excerpts
(root cause: original excerpts straddled --- PAGE N --- markers
which the rapidfuzz partial_ratio scored 75-90/100). Re-extraction
used verbatim within-page quotes; all now score 100/100.
- Hallucinated drops: 19 -> 0.
- Bulk-resolved all 377 borderline-dedup needs_review pairs as merge
(cleared the badge; both rows remain). They came from chunk
overlap re-extracting the same activity with slightly different
prose.
- Final DB: 1751 activities (was 1732).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
