feat: provenance — track source corroboration per entity + relationship

[rajnavakoti]

Part of #82 (calibrated honesty).

Problem

The over-fit errors ("Quarkus underlies Argo for Cleveron") come from the LLM asserting global claims from a single document's local context. Today there's no record of how many / which docs support a given entity or relationship — so "asserted from 1 doc, uncorroborated" is indistinguishable from "corroborated across 5 docs."

Goal

Track and expose provenance — the source documents supporting each entity and each relationship — as the raw material for confidence (#calibration) and a trust signal in its own right.

Acceptance Criteria

• Each entity records the set of source documents it was derived from (entities already have source_documents; ensure it accumulates correctly across the dedup/merge step)
• Each relationship records its source document(s) — a relationship seen in 1 doc vs many is distinguishable
• A corroboration count (how many independent docs assert this) is available per entity + relationship
• Single-source assertions are flaggable (uncorroborated / low-corroboration)
• Tests: an entity/relationship mentioned in 3 docs shows corroboration 3; a single-doc one shows 1

Out of scope

• Using provenance to compute confidence (that's the calibration issue)
• Surfacing it to consumers (that's the retrieval/MCP propagation issue)