Use language-aware ASCII transliteration for canonical IDs

[ziodave]

Problem

The IRI_UTF_fix branch deletes non-ASCII letters. Distinct non-Latin names can collapse to one canonical entity, and Latin letters such as ß, ł and ı are lost.

Goal

Generate ASCII identifiers using account.language, shared ICU transliteration, and deterministic identity suffixes.

Scope

• Shared normalization for allocator and both generator strategies
• German umlaut rules and explicit supported language routes, with deterministic fallback for unsupported readings
• Account language propagation through cloud, one-shot and persistent contexts
• SHA-256 of normalized original non-ASCII names when no URL hash is available
• PyICU dependency, packaging verification, documentation and regression tests

Non-goals

Automatic migration/deletion of stored entities, universal pronunciation inference, and redesign of existing ASCII-only identity rules.

Acceptance criteria

• Distinct non-Latin names remain distinct within and across callback graphs
• German ü becomes ue; Turkish ü does not use German rules
• Japanese kanji never silently receives Chinese readings
• Allocator and both generator strategies agree and remain stable on repeated runs
• Existing URL hashes, root lookup reuse and parent nesting are preserved
• Language survives both worker modes without cross-job leakage
• Focused tests, kg-build slice checks and Python 3.14 validation pass