response_evaluation_score (COHERENCE) fails with INVALID_ARGUMENT: Error parsing JSON in all regions — vertexai 1.153.1 #6853

Bug description

When using PrebuiltMetric.COHERENCE via client.evals.evaluate() (which backs response_evaluation_score in google-adk), the call consistently fails with:

400 INVALID_ARGUMENT: Error parsing JSON. Last error: Expecting property name
enclosed in double quotes: line 1 column 2 (char 1), Input: {* **STEP 1: ...}

The autorater LLM generates a free-text response starting with {* (bullet points), and the :evaluateInstances endpoint attempts to parse it as JSON and fails.

Reproduction

import vertexai
import pandas as pd

client = vertexai.Client(project="<project>", location="us-central1")

df = pd.DataFrame([{
    "prompt": "What slabs are in the model?",
    "response": "The model has 17 C25 concrete slabs."
}])

result = client.evals.evaluate(
    dataset=df,
    metrics=[vertexai.types.PrebuiltMetric.COHERENCE],
)

Environment

• google-cloud-aiplatform (vertexai): 1.153.1
• Regions tested: us-central1, europe-west1, europe-west3 — all fail
• The COHERENCE metric YAML (gs://vertex-ai-generative-ai-eval-sdk-resources/metrics/coherence/v1.yaml) produces free-text output, but the :evaluateInstances server endpoint attempts to parse the autorater response as JSON and fails

Expected behavior

Metric evaluation completes and returns a coherence score.

Actual behavior

num_cases_error=1, num_cases_valid=0 on every call. The full error from the server:

google.genai.errors.ClientError: 400 INVALID_ARGUMENT.
Error parsing JSON. Last error: Expecting property name enclosed in double quotes:
line 1 column 2 (char 1), Input: {* **STEP 1: Identify the purpose and audience:...

The autorater generates a valid free-text evaluation but the server-side JSON parser rejects it.