feat: calibrated entity/relationship confidence (fix uniform 0.9 self-ratings)

[rajnavakoti]

Part of #82 (calibrated honesty).

Problem

The LLM self-rates almost every entity at confidence 0.9+ (a real 3-doc run had 0 entities below 0.6). A confidence signal that's always high is useless — it can't tell a human curator or a consuming agent which entities to doubt.

Goal

Confidence values that actually discriminate trustworthy from shaky entities/relationships, so they're usable as a trust signal.

Approaches to evaluate

• Corroboration-based (leverages #provenance): entity asserted from 1 doc → lower; corroborated across N docs → higher. This is likely the strongest, cheapest signal.
• Evidence-grounding: was the claim explicitly stated vs inferred? (the prompt already asks for hedged_statements — use them.)
• A calibration pass / rubric that down-weights single-source, inference-heavy, or hedged claims.

Acceptance Criteria

• Confidence is derived from signals (corroboration / evidence / hedging), not just the LLM's self-rating
• Distribution actually spreads (not all clustered at 0.9) on a representative corpus
• Single-source, inference-only entities score measurably lower than multi-source, explicitly-stated ones
• Relationships get their own confidence (not just entities)
• Tests: a single-doc inferred claim scores lower than a corroborated explicit one

Out of scope

• Provenance data model itself (separate issue) — this consumes it.