Summary

• skip sample load smoke-test entries that require real Google Cloud ADC at import time
• keep LiteLLM streaming finish-reason handling typed as non-null before setting response error details
• remove unused imports from tests/unittests/test_samples.py

Why

Recent CI runs on existing PRs hit two unrelated main-branch failures:

• tests/unittests/test_samples.py::test_sample_loads[...] failed for Cloud samples because offline CI has no ADC.
• mypy reported new FinishReason | None errors at the LiteLLM streaming finalizers when passing mapped_finish_reason into _finish_reason_to_error_message.

This keeps the new sample load smoke test hermetic while preserving coverage for samples that can construct their root agents offline.

Verification

RED before fix:

• uv run pytest tests/unittests/test_samples.py::test_sample_loads --tb=short -q -x failed on a2a/a2a_auth/remote_a2a/bigquery_agent with DefaultCredentialsError.
• uv run mypy src/google/adk/models/lite_llm.py reported the targeted _finish_reason_to_error_message FinishReason | None errors.

GREEN after fix:

• uv run pytest tests/unittests/test_samples.py::test_sample_loads --tb=short -q -> 149 passed, 30 skipped, 3 xfailed, 1 xpassed
• uv run pytest tests/unittests/models/test_litellm.py -q -> 317 passed, 1 skipped
• uv run ruff check src/google/adk/models/lite_llm.py tests/unittests/test_samples.py -> passed
• uv run pyink --check src/google/adk/models/lite_llm.py tests/unittests/test_samples.py -> passed
• git diff --check -> passed
• uv run mypy . still reports the existing baseline errors, but the _finish_reason_to_error_message regression is gone from the full output.