Please ensure you have read the contribution guide before creating a pull request.

Link to Issue or Description of Change

• Closes: LiteLlm streaming stores one thought part per reasoning delta #6895

Problem:
LiteLlm streaming builds the aggregated (non-partial) LlmResponse by joining the buffered text into a single part, but passes the buffered reasoning parts straight through unjoined. So for a provider that streams reasoning token-by-token (e.g. xai/grok-4.6, OpenAI reasoning models via LiteLLM), the final response — and therefore the persisted session event — holds one types.Part(thought=True) per streamed reasoning delta instead of one part per thinking block. _aggregate_streaming_thought_parts already does this joining (splitting on thought_signature so Anthropic's per-block boundaries are preserved), but it was only wired into the Anthropic message-building path (_content_to_message_param), not into the two stream finalizers that build the response object itself.

Solution:
Call _aggregate_streaming_thought_parts(reasoning_parts) from both _finalize_tool_call_response and _finalize_text_response in src/google/adk/models/lite_llm.py, instead of list(reasoning_parts). This makes every LiteLlm provider produce the same shape the non-streaming path already produces, and matches the aggregator's own docstring ("produces clean parts for session history and outbound requests").

Testing Plan

Unit Tests:

• I have added or updated unit tests for my change.
• All unit tests pass locally.

Added test_generate_content_async_stream_aggregates_reasoning_deltas in tests/unittests/models/test_litellm.py, which streams three separate reasoning_content deltas (no thought_signature, matching how a non-Anthropic provider like xAI streams) followed by a text delta, and asserts the final non-partial response contains exactly one thought=True part with the joined text, instead of three.

$ pytest tests/unittests/models/test_litellm.py -q
404 passed, 1 warning in 5.27s

Also ran the full reasoning/thought-focused subset in isolation:

$ pytest tests/unittests/models/test_litellm.py -k "reasoning or thought" -q
44 passed, 360 deselected in 8.85s

Manual End-to-End (E2E) Tests:

Not run — this is a pure data-shape bug in response aggregation, fully exercised by the unit test above (constructs the same per-token ModelResponseStream deltas a live xAI/OpenAI-reasoning stream would produce and asserts on the resulting LlmResponse.content.parts). No live provider credentials were available to additionally verify against a real streaming API call.

Checklist

• I have read the CONTRIBUTING.md document.
• I have performed a self-review of my own code.
• I have commented my code, particularly in hard-to-understand areas.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• I have manually tested my changes end-to-end.
• Any dependent changes have been merged and published in downstream modules.

Additional context

Both call sites previously read thought_parts=list(reasoning_parts) if reasoning_parts else None; the fix replaces list(reasoning_parts) with _aggregate_streaming_thought_parts(reasoning_parts). No behavior changes for the Anthropic path, which already routed through the aggregator via a separate code path (_content_to_message_param) for outbound requests — this PR fixes the inbound/session-persisted shape for all providers, including Anthropic's own aggregated response object.