0.2.0

LLM-as-judge scoring.

• derail judge <report.json>: scores saved A/B reports on three rubric
  constructs — belief_stickiness (contradiction probes), catastrophizing
  and negativity (task turns) — with any OpenAI-compatible judge model
• The judge sees the planted stimulus and the response text only, runs at
  temperature 0; parse failures are counted, never silently dropped
• Self-judging produces a warning; --judge-model scripted is an offline
  dry-run
• Zero new dependencies: the judge rides the existing OpenAI-compatible
  client

Emulation, not diagnosis.