v0.3.0

Highlights

• Multimodal tool results across all 4 converters — tools can return text + images + files (#92)
• ConversionContext base class for structured conversion pipelines (#106)
• Runtime IR validation via zero-dependency vendored validator (#91)
• ProviderMetadataStore replacing global metadata cache with TTL/eviction (#112)
• Warnings single-source convergence — eliminates dual-write patterns (#113)
• Shrunk public API surface — internal ops classes no longer in __all__ (#114)

Breaking Changes

• ReasoningConfig.type → ReasoningConfig.enabled (bool)
• candidate_count removed from GenerationConfig (use n)
• ImagePart/FilePart/AudioPart normalized to canonical forms
• IR type fields: Iterable → list; function params → Sequence
• StreamContext now inherits from ConversionContext
• Internal ops classes removed from converter __all__ exports

See the full changelog for details.

PyPI: https://pypi.org/project/llm-rosetta/0.3.0/