AgentDetails.tool_declarations come back as dicts after a JSON round trip, breaking judge metrics on saved runs #7876

🔴 Required Information

Describe the Bug:
AgentDetails.tool_declarations is typed list[Any], so the genai_types.Tool objects in it don't survive JSON. Two ordinary flows round-trip through JSON:

• Loading saved eval results (EvalSetResult.model_validate_json).
• The eval service guide's "Score a run again without rerunning it" flow (InferenceResult.model_validate_json).

After either one, every tool comes back as a plain dict. Consumers of the field expect Tool objects:

• rubric_based_multi_turn_trajectory_quality_v1 reads tool.function_declarations and raises AttributeError: 'dict' object has no attribute 'function_declarations'. LocalEvalService catches that, so the metric silently becomes NOT_EVALUATED.
• The other judges serialize the tools through get_tool_declarations_as_json_str. Their prompt now carries the dict's null fields: 785 characters instead of 224 for one tool, including 17 nulls. The same run is therefore judged on different input depending on whether it was saved first.

Steps to Reproduce:

1. Build an AppDetails whose agent has a genai_types.Tool in tool_declarations. Inference records this on every invocation.
2. Round-trip it with AppDetails.model_validate_json(app_details.model_dump_json()).
3. Inspect the type of tool_declarations[0], or score the round-tripped invocations with rubric_based_multi_turn_trajectory_quality_v1.

Expected Behavior:
The round-tripped tool_declarations equal the original Tool objects, and the judges see the same tools text as before saving.

Observed Behavior:

entry type after round trip: dict
judge tools text: original 224 chars, restored 785 chars, nulls in restored: 17

  File ".../rubric_based_multi_turn_trajectory_evaluator.py", line 270, in _assemble_dialogue_history
    for func in tool.function_declarations or []:
AttributeError: 'dict' object has no attribute 'function_declarations'
restored run -> rubric_based_multi_turn_trajectory_quality_v1 NOT_EVALUATED score None

Environment Details:

• ADK Library Version (pip show google-adk): 2.13.0 (also on main at e9e8f61)
• Desktop OS: macOS
• Python Version (python -V): 3.11.16

Model Information:

• Are you using LiteLLM: No
• Which model is being used: any (the round trip doesn't involve a model)

🟡 Optional Information

Minimal Reproduction Code:

from google.genai import types
from google.adk.evaluation.app_details import AgentDetails, AppDetails

tool = types.Tool(function_declarations=[
    types.FunctionDeclaration(name="set_light", description="Turns a light on or off.")])
app = AppDetails(agent_details={"home": AgentDetails(name="home", tool_declarations=[tool])})

restored = AppDetails.model_validate_json(app.model_dump_json())
print(type(restored.agent_details["home"].tool_declarations[0]))  # <class 'dict'>

How often has this issue occurred?:

• Always (100%), for any invocation whose agent has tools, once it has been through JSON