v0.5.2

What's Changed

Fixed

• Streaming round-trip event inflation (#157): Fixed multiple scenarios where Provider A → IR → Provider B streaming conversion produced more output events than input events:
  • OpenAI Chat, Anthropic, and Google GenAI converters emitted redundant content_block_end events when no content block was open
  • Google GenAI compound chunks (text + finish in the same SSE frame) triggered duplicate text and finish events. Deferred via StreamContext.pending_text / pending_finish
  • Tool call events generated spurious content_block_start / content_block_end wrappers in non-Anthropic targets

Refactored

• Unified stream_response_to_provider dispatch (#73, #157): Extracted identical dispatch logic (10-entry _TO_P_DISPATCH table + dispatch skeleton) from all 4 provider converters into BaseConverter. Each converter now only implements a provider-specific _post_process_to_provider hook
• StreamContext buffer convenience methods: Added buffer_usage() / pop_pending_usage() / buffer_finish() / pop_pending_finish() to replace manual set-and-clear patterns

Changed

• Pinned dev tooling versions: ty>=0.0.31 and ruff>=0.15.0 in pyproject.toml dev dependencies. CI uses versions from pip install -e ".[all]"
• Converter tests added to CI: tests/converters/ (1086+ tests) now runs in GitHub Actions
• Roundtrip inflation regression test: New pytest-parametrized test suite (15 cases) verifies len(output_events) <= len(input_events) for all 4 providers

Full Changelog: v0.5.1...v0.5.2