Expand ambiguous surfaces to every lemma they are a form of

[kdroidFilter]

Summary

Fixes kdroidFilter/SeforimMagicIndexer#3 on the reader side.

Some surfaces are forms of several lemmas, but lexical.db tied each surface to one base_id only. Since #116/#117, a token expands to the entries it is truly a surface of, so פנינו expanded to פנה only and never to פנים ("our face").

The v0.5.0 dictionary adds surface_extra_base(surface_id, base_id), with 26,670 extra lemmas judged in kdroidFilter/SeforimMagicIndexer#5. MagicDictionaryIndex now reads it:

• A token matches each lemma it is a form of.
• The surface counts among the forms (inflections) of each of those lemmas, so expansionFor treats them as its own entries.
• The surface's variants stay with its primary lemma, since they hold associations of that reading.
• Dictionaries without the table keep the previous LOOKUP_SQL, so older lexical.db files work unchanged.

Testing

• :search:jvmTest --tests '*MagicDictionaryIndexTest*': 22/22, including a new test for the פנינו → פנה + פנים case.
• Query run on the real v0.5.0 DB: פנינו → פנה, פני, פנים. Timing is unchanged (2–20 ms for frequent words like כל, אמר).

Note

DownloadLexicalDb only picks a release asset named exactly lexical.db. v0.5.0 uses that name, and the v0.4.1 asset was renamed to match.