Normalizes entity names into deduplication keys.

LLM extractors return cosmetic variants of the same real-world entity across chunks ("Two_Year_Limited_Warranty" vs "two year limited warranty"). GraphBuilder and the graph stores compute their dedup and upsert keys through this module so they agree on which names refer to the same entity. Display names keep their first-seen original form.

Backend differences

The Ecto store computes the same key in SQL (see Arcana.Graph.GraphStore.Ecto) because it has to compare a stored name against an incoming one inside a query. Whitespace folds identically on both sides: each collapses the full Unicode White_Space set, which is wider than Postgres' \s (it doesn't match NBSP) and wider than Elixir's (ASCII only).

Two differences remain, and are deliberate:

• Case folding. The two engines disagree on an open-ended set of codepoints, and the set depends on the database host's libc, not on anything this library controls. Sweeping U+0001..U+2FFFD against PostgreSQL 16.15 on glibc found 56: U+0130, U+1C89, U+A7CB, U+A7CC, U+A7CE, U+A7D2, U+A7D4, U+A7DA, U+A7DC, U+10D50..U+10D65 (Garay) and U+16EA0..U+16EB8. All but the first are glibc lagging Unicode, so a newer or older host shifts the list; U+0130 (İ) is the one that diverges by Unicode's own SpecialCasing rules, and so behaves the same everywhere: String.downcase/1 decomposes it into i plus a combining dot while Postgres' lower() folds it to a bare i, which makes "İstanbul" and "istanbul" one entity in the Ecto store and two in the Memory store.

  Do not treat any of these as a closed list. The Ecto store keys its returned id map on the raw name for exactly this reason, so a spelling pair the two engines disagree about still reaches its own row whatever the host libc thinks.

• Neither backend applies canonical normalization, so NFC "Café" and NFD "Café" are different entities in both.