v0.2.0

Added

• Model registry: full catalog sync from LiteLLM plus LiveBench and LMSYS benchmark
  scores, with a Models page for provider management and model testing.
• Discovery endpoints: GET /v1/models, GET /v1/providers, GET /v1/task-types,
  filtered to live providers with tool-call capability flags.
• Custom providers: connect any OpenAI-compatible endpoint from the UI, including a
  LiteLLM proxy type.
• Applications hub: per-app kill switch, routing policy, and model opt-out; Budgets page
  with per-app and per-provider spend caps.
• Analytics: per-user spend and realised-savings visibility (#38); provider prompt-cache
  tokens captured and priced (#37).
• Routing: difficulty-aware classification signal (#36); LLM-based policy generation;
  scoring-engine router replacing the LLM router.
• Bedrock: tool-call support on the Converse API, including streamed tool calls.
• AI governance controls: transparency, safety, audit, and data lifecycle.

Changed

• Repository prepared for open source under project-arbr (#1); ESLint added and CI
  hardened (#33); dependabot PRs grouped per ecosystem (#17).

Fixed

• Completions no longer truncate at 1024 tokens; real finish_reason reported (#34).
• SSE truncation on /v1/chat/completions for native providers; DeepSeek R1 hang and
  ValidationException on Bedrock; crash on null messages with tool-result history.