Evidence-first AI bake-off completion and fallback hardening

[sev7enITA]

Summary

• preserve deterministic evidence-first acquisition and publication gates
• complete provider JSON Schema, privacy-minimized telemetry and frozen golden-set evaluation
• make Qwen3 a real embedding-then-reranking track and add loopback-only local adapters
• replace three polynomial regex paths with bounded linear parsers
• classify network resets as transient fallback events and use monotonic latency
• contain golden-set file access before loopback transmission

Observed bake-off

• BM25: frozen retrieval gates pass
• Gemini 3.5 Flash-Lite: 5/5 completed, all extraction gates pass
• Gemini 3.7 Flash: 0/3 completed in final run, provider returned 503, no promotion
• Qwen3/BGE-M3: not scored because anonymous Hugging Face downloads could not complete; no substitute results
• RAGFlow, LightRAG, Kimi K3 and GraphRAG remain research-only

Verification

• 139 Vitest files / 736 tests
• ESLint and TypeScript pass
• production build passes with 159 pages
• production dependency audit reports zero vulnerabilities
• local Prisma schema is current
• golden-set validation and deterministic baseline pass