v2.0 - Entity Resolver, arc42 Documentation, TurboQuant KV Cache

What's New

Six-stage entity resolver

Academically grounded entity linker (stages 0-5) with a two-tier canonical alias gazetteer that eliminates cross-document fork epidemics. 149-entry curated seed tier for AI/tech entities, plus auto-promoted runtime tier.

Full arc42 architecture documentation

14 sections + academic retrospective following the arc42 template, with C4 Level 1/2/3 diagrams. Includes security audit (7 findings), PII audit (4 findings), 23 quality scenarios and 7 ADRs.

TurboQuant asymmetric KV cache

Asymmetric q8_0 K + turbo4 V configuration via the llama-cpp-turboquant fork. Frees ~3 GB of memory on a 32 GB machine while preserving quality on Gemma 4 Q4_K_M.

Other additions

• cleanup_dedup.py for offline merge of duplicate wiki pages
• Optional bge-m3 embedding server for cross-lingual entity resolution
• 2150 lines of resolver scenario tests
• Comprehensive .gitignore covering all generated content

Breaking Changes

None. Existing wiki content and raw sources are unaffected.