Summary

• Switch Claude.generateContent() to use the Anthropic SDK's createStreaming() API when stream=true, replacing the previous non-streaming fallback (which had a TODO: Switch to streaming API comment)
• Text deltas are emitted immediately as partial LlmResponses with partial=true
• Tool use JSON is accumulated across inputJson deltas and emitted as a function call LlmResponse on contentBlockStop
• messageStop emits a turnComplete=true sentinel
• StreamResponse is always closed via Flowable.using() to prevent resource leaks
• Extracted buildMessageCreateParams() to deduplicate param-building logic shared by both streaming and non-streaming paths

Test plan

• testStreaming_textChunksEmittedAsPartialResponses — text deltas become partial=true responses in order
• testStreaming_messageStopEmitsTurnComplete — messageStop event sets turnComplete=true
• testStreaming_toolCallAccumulatedAndEmittedOnBlockStop — tool JSON accumulates across deltas, emits complete function call on block stop
• testStreaming_mixedTextAndToolCall — interleaved text + tool call in same stream produces correct ordered responses
• testStreaming_emptyStream_producesNoResponses — empty event stream produces no output
• testStreaming_toolCallWithInvalidJson_fallsBackToEmptyArgs — malformed JSON logs a warning and falls back to empty args map
• All 1284 existing core module tests still pass