LocalEvalService crashes with len(None) when inference_result.inferences is None #5876

Summary

LocalEvalService._evaluate_single_inference_result() crashes with TypeError: object of type 'NoneType' has no len() when inference fails and InferenceResult.inferences is left as None.

The failing path is visible in google-adk==1.34.0:

if eval_case.conversation_scenario is None and len(
    inference_result.inferences
) != len(eval_case.conversation):
    ...

InferenceResult.inferences is optional and _perform_inference_single_eval_item() can produce a failure result with status=InferenceStatus.FAILURE, error_message=..., and inferences=None (for example when auth refresh, network, or model calls fail). The later len(inference_result.inferences) aborts the entire eval run instead of returning a structured per-case failure.

Reproduction

Run any eval case where inference fails before invocations are generated. One local repro from an ADK project using Vertex/ADC with expired ADC credentials:

cd idc-runtime
GOOGLE_GENAI_USE_VERTEXAI=True \
GOOGLE_CLOUD_PROJECT=idc-runtime-prod \
GOOGLE_CLOUD_LOCATION=global \
uv run agents-cli eval run --evalset tests/eval/evalsets/build_role.evalset.json

The underlying inference error is:

google.auth.exceptions.RefreshError: Reauthentication is needed. Please run `gcloud auth application-default login` to reauthenticate.

Then ADK crashes in the eval framework:

File ".../google/adk/evaluation/local_eval_service.py", line 273, in _evaluate_single_inference_result
    if eval_case.conversation_scenario is None and len(
                                                   ^^^^
TypeError: object of type 'NoneType' has no len()

The same issue can surface during multi-case eval runs; several tasks can fail inference and then all hit len(None) while asyncio.as_completed(...) is collecting results.

Expected behavior

If inference failed and inference_result.inferences is None, ADK should not call len(None). It should surface a structured failed EvalCaseResult using the existing inference_result.error_message, then allow the rest of the eval run to continue.

Local workaround

We are carrying a narrow in-process guard that checks inference_result.inferences is None before delegating to the original method and returns an EvalCaseResult(final_eval_status=EvalStatus.FAILED, ...) for that case. We also serialize CI eval cases while this failure path is unresolved.

Environment

• google-adk==1.34.0
• Python 3.12
• macOS local repro; also observed from GitHub Actions runners when live Vertex evals reach an inference failure