docs(research): precision & efficiency research-agent architecture (#12621)

[mrveiss]

Thinking Path

Research of Maltego / SpiderFoot / Recon-ng / Sherlock, re-scoped to general-topic research (not OSINT), showed AutoBot already owns the needed infrastructure — a KB fact store with provenance/verification (knowledge/facts.py), fetchers, a source registry, and a claim verifier — but they are disconnected, findings never become verified facts, and synthesis is an un-cited template string. This design connects them into one bounded, KB-native loop and adds the missing LLM reasoning layer.

What Changed

Adds docs/architecture/RESEARCH_AGENT_PRECISION_EFFICIENCY_DESIGN.md — the design doc for umbrella #12621 (children #12622–#12625). Docs only; no code.

• KB-native data model (Claim→Fact, Source→Document, Topic→Collection, contradictions→Relation, staleness→Versioning/temporal) — no parallel store.
• Pipeline: /research → Planner (decompose, skip-known, prune) → Source Router → Fetch/Extract → store_fact → N-source Corroborator → saturation stop → grounded cited synthesis.
• Owner decisions locked: /research API first; KB Collections govern scope; quarantine-until-corroborated promotion; both inline + structured contradictions[] output; first PR bundles P0 + synthesis.

Verification

• Doc only — no runtime surface. git show --stat: 1 file, +280.
• Pre-commit hooks passed on commit (trailing-whitespace etc.).
• Table/emphasis/heading style matches existing docs/architecture/* house style (repo has no markdownlint config; compact tables are the norm).

Model Used

Claude Opus 4.8 (1M context)

Closes #12621 (umbrella; tracked, not truly closed — Dev_new_gui merges do not auto-close). Children #12622–#12625 remain open.

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!