Harden architecture and establish local reasoning qualification

[faresmohamed260]

Summary

• harden narrative grounding, identity canonicalization, provider attribution, database administration, and bounded release qualification
• add a provider-neutral local reasoning mode with native Ollama JSON and tool contracts
• add a reproducible three-book qualification corpus, resumable checkpoints, deterministic evaluators, scorecards, and per-task routing gates
• enforce workstation-safe model admission, context, RAM/VRAM, concurrency, queue, and candidate-fit policies

Why

This consolidates the clean-architecture hardening and establishes local inference as the reliable reasoning baseline without cloud fallback. It also fails closed when model quality, provenance, or workstation headroom is insufficient.

Validation

• local reasoning and LangGraph focused suite: 60 passed
• local model fit policy: passed
• real-book screening completed for Mistral 7B, Llama 3.1 8B, and Qwen2.5 14B
• full qualification currently proves Qwen2.5 relationship extraction (9/9) and Mistral JSON/tool use (18/18)
• safe qwen resource test stayed below the 10 GiB VRAM gate; interactive throughput was insufficient and is documented
• existing backend, architecture, dashboard, migration, and isolated backup/restore checks remain documented in prior commits

Release status

Draft only. Local reasoning is not production-ready yet. Events, entities, character/world modeling, planning, continuity, and narrative routes remain unqualified; true extraction recall still needs human gold labels. Qwen3.5 9B qualification is pending its resumable local download. No production deployment or migration was performed.