Summary

This PR improves OpenResponses-based pixel-generation serving by adding file-backed media responses and response-format selection for generated outputs.

Specifically, it:

• adds file-backed image/video response handling for /v1/responses
• supports response_format: "url" and response_format: "b64_json" for generated media
• validates that the requested model matches the currently served model and returns 404 on mismatch
• adds video-output handling in the shared pixel-generation pipeline so video responses can be serialized correctly

Example local flow validated during development:

1. Start a server:

   MAX_SERVE_API_TYPES='["responses"]' \
   ./bazelw run //max/python/max/entrypoints:pipelines -- serve \
     --model-path black-forest-labs/FLUX.2-klein-4B \
     --task pixel_generation \
     --port 8000 \
     --devices gpu \
     --prefer-module-v3

2. Send a T2I request:

   cat >/tmp/flux_t2i_request.json <<'EOF'
   {
     "model": "black-forest-labs/FLUX.2-klein-4B",
     "input": "A studio portrait of a tabby cat with dramatic lighting.",
     "seed": 42,
     "provider_options": {
       "image": {
         "guidance_scale": 4.0,
         "output_format": "png",
         "response_format": "url",
         "width": 512,
         "height": 512,
         "steps": 4
       }
     }
   }
   EOF
   
   curl -sS http://127.0.0.1:8000/v1/responses \
     -H 'Content-Type: application/json' \
     --data @/tmp/flux_t2i_request.json \
     > /tmp/flux_t2i_response.json

With response_format: "url", the response includes an image_url that can
be fetched from /v1/images/{image_id}/content. With
response_format: "b64_json", the response includes inline image_data.

Testing

• ./bazelw test //max/tests/tests/serve:test_openresponses_routes

//max/tests/tests/serve:test_openresponses_routes specifically verifies that:

• basic /v1/responses requests still succeed
• requests are rejected when the requested model does not match the served model
• video responses can be returned as downloadable file-backed URLs
• video responses can be returned as inline base64 payloads when response_format: "b64_json" is requested

I also manually verified:

• T2I with black-forest-labs/FLUX.2-klein-4B
• response_format: "url" returns image_url
• response_format: "b64_json" returns inline image_data
• requesting a different model name than the served model returns 404

Checklist

• PR is small and focused — consider splitting larger changes into a
  sequence of smaller PRs
• I ran ./bazelw run format to format my changes
• I added or updated tests to cover my changes
• If AI tools assisted with this contribution, I have included an
  Assisted-by: trailer in my commit message or this PR description
  (see AI Tool Use Policy)

Assisted-by: OpenAI Codex