release: v0.3.0

• feat: implement split architecture (Alpine API + Debian Models)
• feat: proxy file conversion to models container in split mode
• feat: batch BERT inference for 3–5× lower latency on large inputs
• feat: persist pricing data to disk cache (~/.llmzip/prices.json)
• feat: offline-first tiktoken support in Docker (pre-downloaded encodings)
• fix: resolve permission error for logging directory under non-root user
• fix: correct original/compressed token tracking in batch responses
• fix: resolve duplicate LiteLLM fetches under concurrent load
• test: expand test suite covering pricing, i18n, and chunker stress

BREAKING CHANGES

• resolve_prices() and fetch_prices() now return:

  tuple[dict[str, PriceEntry], dict[str, str]]

• The default Dockerfile now builds the lightweight Alpine API image and no longer contains ML dependencies. To run the monolith architecture, you must build from Dockerfile.models or pull the -models image variant.