Judeo-Arabic to Arabic-script transliteration helper (representations form)

[bsesic]

Summary

Add a Judeo-Arabic ↔ Arabic-script transliteration helper that produces an Arabic-script
representations form for Arabic written in Hebrew script. Small, self-contained utility that
fits TRACE's existing Token.representations model and feeds the embedding step of the
cross-lingual path (#17).

Motivation

Much of the target corpus is Judeo-Arabic — Arabic in Hebrew characters. Two concrete payoffs:

1. For the cross-lingual embedding fill, transliterating Hebrew-script Arabic into Arabic script
   before embedding moves the text closer to what multilingual embedding models (LaBSE etc.)
   have seen and measurably improves matching.
2. It lets Judeo-Arabic text reuse the Arabic pack's (Arabic language pack (ara) — proclitic tokenization + orthographic normalization #16) normalization/scoring once
   mapped, rather than needing a parallel ruleset.

This is the one place the script-conversion trick legitimately applies — it helps when the
underlying wording is the same text in a different script. (Note for users: it does not
help across a translation, where the wording itself differs.)

Scope

• A deterministic mapping function Hebrew-script → Arabic-script for Judeo-Arabic, handling the
  conventional correspondences (and the well-known ambiguous letters), exposed as a normalizer
  that writes an arabic_script entry into Token.representations (diplomatic Hebrew-script
  form preserved in raw/text).
• Configurable/overridable mapping table (conventions vary by period and community), shipped as
  a data file under the relevant pack.
• Documented as best-effort with known ambiguities flagged, not a lossless transliteration.

Acceptance criteria

• A transliteration utility producing an arabic_script representation for Judeo-Arabic tokens.
• Mapping table is data-driven and user-overridable.
• Tests on a small Judeo-Arabic fixture, including documented ambiguous-letter cases.
• Integrates as an optional pre-step feeding the cross-lingual embedding adapter (Cross-lingual alignment path: anchor-based mode + embedding-fill delegation hook #17).
• TDD; suite green on 3.10/3.11/3.12; flake8 clean.

Notes

• Depends conceptually on the Arabic pack (Arabic language pack (ara) — proclitic tokenization + orthographic normalization #16) for the downstream normalization/scoring,
  but the mapping utility itself can be built and tested independently.

Roadmap: supporting utility for Stage 6 cross-lingual work; serves the Judeo-Arabic portion
of the medieval-editions corpus.