fix(engine): preserve Unicode semantic identities

[u235yang]

Summary

• preserve the existing semantic slugs for pure ASCII names while deriving normalized, stable SHA-256 slugs for names containing non-ASCII characters
• use the analyzed concept/entity ID consistently across graph nodes, aggregate pages, source links, backlinks, and candidate paths
• migrate compatible v8 analysis caches to v9 deterministically without another provider/LLM call, while invalid caches still re-analyze during a full compile
• deduplicate aggregate provenance and normalized duplicate terms without dropping distinct legacy ASCII terms such as C++ and C#
• archive obsolete generated pages containing guided-session markers byte-for-byte instead of assigning ambiguous notes to one of several split Unicode pages; compile approvals treat archive creation and old-page deletion as one atomic action

Why

The legacy slugger strips every character outside [a-z0-9]. Chinese concepts and entities therefore collapsed to concept:item / entity:item, causing unrelated terms to share graph nodes, wiki pages, backlinks, and index entries.

Compatibility and migration

• pure ASCII semantic IDs remain unchanged
• NFKC normalization makes equivalent Unicode forms resolve to the same identity
• valid v8 caches migrate locally and preserve producedAt; no provider call is required
• old many-to-one generated pages are removed from the active graph, while any human guided-session notes are retained in a content-addressed archive for manual reconciliation
• codeOnly compilation remains LLM-free for existing invalid non-code caches; a later full compile repairs them

Validation

• pnpm check
• pnpm --filter @swarmvaultai/engine test — 418 passed, 2 skipped
• pnpm --filter @swarmvaultai/engine build
• focused Unicode/cache/path/archive/approval regression coverage
• independent patch review completed with no remaining blockers