Evaluation pairs every ID-less tool call with the last tool response #8132

Describe the bug

Evaluation transcripts with missing function-call IDs pair every tool call with the last ID-less response. get_all_tool_calls_with_responses constructs a dictionary keyed by response ID, so all missing IDs collide under None. The incorrect pairs reach the rubric tool-use judge prompt.

Minimal reproduction

from google.adk.evaluation.eval_case import IntermediateData, get_all_tool_calls_with_responses
from google.genai import types

data = IntermediateData(
    tool_uses=[types.FunctionCall(name='search'), types.FunctionCall(name='book')],
    tool_responses=[
        types.FunctionResponse(name='search', response={'result': 'found'}),
        types.FunctionResponse(name='book', response={'result': 'booked'}),
    ],
)
print([(call.name, response.name) for call, response in get_all_tool_calls_with_responses(data)])

Actual: [('search', 'book'), ('book', 'book')].
Expected: [('search', 'search'), ('book', 'book')].

Scope and proposed fix

This affects restored or hand-authored legacy IntermediateData and ID-less InvocationEvents, not normal Runner traces that supply call IDs. Match missing-ID records by name and occurrence, consuming each response once; retain explicit-ID matching. For multiple calls of the same name, recorded order is the only available correlation information.

Reproduced on main at 12aab4aaa, Python 3.12.10, Windows. A JSON-round-tripped Invocation was also evaluated through the real RubricBasedToolUseV1Evaluator with a scripted external judge; its prompt contained the incorrect pairs. No external API credentials are needed.

Prepared with AI assistance; reproduction and validation were executed locally.