Chained multi-hop search

A planner LLM decomposes the query into at most maxHops anchored sub-queries; each later hop is scoped to the running entity set, so the engine never spends compute on candidates already disqualified. Answers questions that join evidence across turns or entities. supportingFactIds is the evidence chain (HotpotQA-style) — the caller can audit exactly which facts drove the answer. synthesize: true adds a grounded answer with citations alongside the per-hop trace. Fans out to the shared embedding / LLM budget, so it sits in the per-credential expensive throttle bucket — 10 requests/min, not the 120/min default. Source: src/multi-hop/multi-hop.controller.ts.

Required scope: brain:read.