docs: two-agent research A/B as a short paper

[drewstone]

Rewrites docs/two-agent-research-ab.md into a short research-paper structure (abstract → setup → method → results → discussion → limitations → simpler-loop → reproduce). Method is grounded in the actual code: the worker/driver roles, the driver's mechanical foldGaps steer (not LLM-authored prompts), the hand-wired live worker (router pipeline, no AgentProfile/harness) vs the researcherProfile path, and the equal-compute accounting. Reports both independent runs and is honest that the win is mostly de-duplication.

[tangletools]

✅ Auto-approved PR — d6a93e10

Blanket team auto-approval is enabled for this reviewer service.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-06-24T23:23:13Z}