The 1000_novels dedup (v2#45) uses the same 5-word-shingle containment, so the code
moves to a module both filters import, with its unit tests in
src/test_shingle_containment.py. It adds Match.cover, the share of a document's
shingles held by any reference (the union over Wolne Lektury tomes).
No output change: rebuilding from the v0.2.5 Parquet and stats gives byte-identical
wolne_lektury.parquet, stats.json and eltec_pairs.tsv.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
wolne_lektury: drop 63 docs that duplicate eltec_pol novels (v2#72)
Option A of slayerlabs/polish-dynaword-v2#72: drop the Wolne Lektury
copies and keep eltec_pol whole (CC BY 4.0). A WL doc is dropped when
>= 0.9 of its 5-word shingles are in one ELTeC novel; no doc scores
between 0.7588 and 0.9636. Three partial overlaps (0.3-0.9) are kept
and listed: both Tredowata tomes (shorter ELTeC text) and Kochanowski's
Piesn IX (quoted in Oko proroka).
6,141 -> 6,078 docs, 103,009,797 -> 90,654,820 tokens. Both input
Parquets are pinned by SHA-256; pairs in wolne_lektury.eltec_pairs.tsv.
WL licence labels unchanged (#57 relabels them).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
