v0.2.0

What's New

• Standalone eval dashboard served at / (separate from landing page)
• Generic judge providers (OpenAI-compatible + Anthropic-compatible) via unified JUDGE_* env vars
• Removal of fabricated metrics — total_cost: None replaces fake zero-cost reporting
• Heuristic scores labeled with score_method field
• CI now tolerates guardrail fires on adversarial scenarios (relaxed eval gate)

Breaking Changes

• OPENAI_API_KEY / ANTHROPIC_API_KEY replaced by JUDGE_PROVIDER, JUDGE_API_KEY, JUDGE_BASE_URL, JUDGE_MODEL
• total_cost: 0.0 now total_cost: null for unmeasured runs
• Swift runner directory removed