feat: review-state — distinguish raw-LLM entities from human-verified

[rajnavakoti]

Part of #82 (calibrated honesty).

Problem

There's no way to tell a raw LLM-extracted entity (unverified, possibly wrong like "pickup-point-manager is a domain") from one a human has actually reviewed and confirmed. Consumers treat both identically. The human-in-the-loop curation loop needs a state to track.

Goal

A review_state on every entity (and ideally relationship) so both curators and consuming agents know what a human has vetted.

Proposed states

unreviewed (raw LLM output, default) → verified (human confirmed) → corrected (human edited) → rejected/deprecated.

Acceptance Criteria

• review_state field on entities, defaulting to unreviewed on extraction
• Curation actions (edit/confirm via viewer or API) transition the state
• Re-extraction (source re-publish) does not silently reset a human-verified/corrected entity to unreviewed without signalling
• Queryable: "show all unreviewed entities" (the curation worklist) and "% of KB human-verified" (a health metric)
• Tests: state defaults to unreviewed; a confirm action sets verified

Out of scope

• The full curation UI/queue (separate) — this is the state model + transitions
• Propagating the state to retrieval/MCP (that issue)