Retrieve evidence-grounded historical colour knowledge across the seven validated sources, preserving each source's distinct type of knowledge.

Read-only. Deterministic: term retrieval, intent weighting from registry metadata, and a rule table -- no LLM call and no writes to any source artefact.

The answer layer (historical_intelligence.answers) was built and tested but never wired in, so this endpoint used to return raw retrieval output: no false-premise correction, and structured values (Vanderpoel's percentages, Hokusai's application fields) reachable only in evidence mode. It now runs through build_packet, which is where that logic already lived.