feat(data): seed corpus with 6 Bialik letter crops

[shaypal5]

Summary

• Adds the first real writer (chaim_nachman_bialik, verified) and 6 manual single-letter crops drawn from upstream scan commons__bialik_el_hazippor__p0001 — Bialik's autograph manuscript draft of El Hatzippor (upstream df07bd3).
• Letters: lamed (ל), dalet (ד), vav (ו), he (ה), kaf (כ), bet (ב) — all cut from line 4 of the page, one variant per letter.
• Rights inherited verbatim from upstream: license_expression: PDM-1.0, rights_basis: public_domain, no attribution required. verification_status: inherited_from_upstream.
• Regenerated NOTICE.md, CITATION.cff, datapackage.json. datapackage.json::released_at advances to 2026-05-12T22:30:00Z (max of new extraction.extracted_at); CITATION.cff::date-released stays at 2026-05-12 per release_recipe.json::version_released_date, as designed.
• Crop images committed under data/letters/chaim_nachman_bialik/<letter>/<entry_id>.png, all LFS-tracked (verified via git lfs ls-files).

Type of change

• New writer(s) / new per-letter image entries (ingest)
• Schema or validator change
• Release tooling / CI change
• Documentation / policy
• Refactor / chore (no behaviour change)

Pre-merge checklist

• python3 scripts/validate_indexes.py passes locally — ok: 1 writers, 6 entries, 6 files verified.
• python3 scripts/validate_indexes.py --upstream-path ../public-domain-hand-written-hebrew-scans passes — ok: 1 writers, 6 entries, 6 files verified, 6 upstream-cross-checked.
• python3 scripts/generate_release_artifacts.py was re-run; regenerated NOTICE.md / CITATION.cff / datapackage.json are staged.
• python3 scripts/generate_release_artifacts.py --check passes (ok: release artefacts are up to date).
• python3 -m pytest passes locally — 62 passed, 1 skipped.
• git diff --check shows no whitespace issues.
• New image files are LFS-tracked (git lfs ls-files shows all six PNGs).

Rights / licensing

The upstream entry commons__bialik_el_hazippor__p0001 records license_expression: PDM-1.0, rights_basis: public_domain, attribution_required: false. Bialik died in 1934, so the work is public domain in Israel and other life+70 jurisdictions. The same rights propagate to all six derivative crops; per the inheritance table in LICENSE.md, no attribution metadata is required and verification_status: inherited_from_upstream is correct.

license_expression/rights_basis pair is PDM-1.0 → public_domain, which matches LICENSE_BASIS_MAP in scripts/validate_indexes.py. Since no entry carries an attribution-required license, NOTICE.md's attribution section renders _No entries in this release require attribution._, which is the expected output for this corpus state.

Notes for reviewers

• Letter identification confidence. lamed, kaf, bet are unambiguous from their cursive shapes. dalet, vav, he are best-guess identifications from cursive Hebrew — I could not fully read the surrounding words with certainty. Bboxes themselves are clean isolated single-letter crops, so any mislabel can be fixed by relabeling the entry without re-cropping. Cursive style flagged on each entry as cursive_ashkenazi.
• Crop methodology. Coordinates picked manually by inspecting the upstream scan with Pillow (x, y, w, h of each bbox is recorded in upstream.bbox); each crop saved as PNG at native upstream resolution, then sha256/bytes/dims captured from disk. extraction.tool: "manual", extraction.tool_version: "0.0.0-manual" per the SemVer-prerelease regex.
• Upstream pinning. All entries reference upstream commit df07bd3825405ed93c15fd61fe4d7967fc60885e (40-char SHA, not a tag).
• Scope. Deliberately small (6 letters) to validate the end-to-end pipeline before bulk ingest. Out of scope: tagging v0.0.0-rc, hletterscriptgen integration, additional writers/scans.

Validator output:

$ python3 scripts/validate_indexes.py --upstream-path ../public-domain-hand-written-hebrew-scans
ok: 1 writers, 6 entries, 6 files verified, 6 upstream-cross-checked

🤖 Generated with Claude Code