Interactions API drops media a tool returns (FunctionResponse.parts) #7884

🔴 Required Information

Describe the Bug:
With Gemini(use_interactions_api=True), media a tool returns never reaches the model. ADK moves a tool's types.Part media (images, PDFs and so on) into FunctionResponse.parts. _convert_part_to_interaction_content builds the function_result step only from FunctionResponse.response, so the parts are dropped silently. The model then answers about a chart or report it never saw.

The other model adapters already forward these parts: LiteLlm, AnthropicLlm, ApigeeLlm and the OCI integration all read function_response.parts. The Interactions API accepts media in a function result as a list of contents. google-genai declares image there in every supported release, and adds audio, video and document in 2.33.

Steps to Reproduce:

1. Use Gemini(model="gemini-3-flash-preview", use_interactions_api=True) with a FunctionTool that returns {"metric": ..., "chart": types.Part.from_bytes(data=png, mime_type="image/png"), "report": types.Part.from_bytes(data=pdf, mime_type="application/pdf")}.
2. Let the model call the tool, then inspect the next request's function_result step.

Expected Behavior:
The function_result carries the JSON response plus the image and the PDF.

Observed Behavior:
With a real google-genai client (2.34.0) and an httpx.MockTransport that records requests:

FunctionResponse.parts: [('image/png', 18), ('application/pdf', 20)]
function_result on the wire: {'call_id': 'call-1', 'result': {'metric': 'latency'}, 'is_error': False, 'name': 'render_report', 'type': 'function_result'}

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: Gemini (gemini-3-flash-preview) through the Interactions API

🟡 Optional Information

Minimal Reproduction Code:

from google.adk.models import interactions_utils
from google.genai import types

part = types.Part(
    function_response=types.FunctionResponse(
        id="call-1",
        name="render_report",
        response={"metric": "latency"},
        parts=[
            types.FunctionResponsePart(
                inline_data=types.FunctionResponseBlob(
                    mime_type="image/png", data=b"\x89PNG..."
                )
            )
        ],
    )
)
print(interactions_utils._convert_part_to_interaction_content(part))
# {'type': 'function_result', ..., 'result': {'metric': 'latency'}}  <- no image

How often has this issue occurred?:

• Always (100%), whenever a tool returns media on the Interactions API