Seedream 5 Pro Partner Node disconnects at ~120 seconds for 2K reference edits despite a longer local client timeout

## Summary

Seedream 5 Pro reference-edit requests sent through the official ComfyUI Partner Node frequently terminate at approximately 120.8 seconds when the requested output is around 1920x1080, 2048x1152, or 2560x1440.

The same input class and prompt are much more likely to complete at 1K, and a larger Seedream 5 Lite request completed successfully. This makes the failures look like a fixed deadline in the `api.comfy.org` Partner proxy path rather than an unsupported resolution or a general local-network outage.

The affected endpoint observed in the local API logs is:

```text
POST https://api.comfy.org/proxy/byteplus/api/v3/images/generations
```

## Environment

- OS: Windows 11 (10.0.26100)
- ComfyUI: v0.33.1, commit `72865f4f27eaf5396f8f36370e0a2be3a9a090ee`
- Frontend package: 1.48.7
- Python: 3.13.14 embedded
- aiohttp: 3.14.3
- Node: `ByteDanceSeedreamNodeV2`
- Model: `seedream 5.0 pro`
- Authentication: Comfy account API key with sufficient prepaid credits
- Reference-image editing, `thinking=true`

For diagnosis, the local HTTP client timeout was increased to 180 seconds and automatic retries were disabled. The remote connection still ended at approximately 120.8 seconds. No provider endpoint, request payload, or model behavior was changed by that diagnostic modification.

## Reproduction steps

1. Log in to local ComfyUI using a Comfy Account API key and ensure the account has sufficient credits.
2. Add the official ByteDance Seedream Partner Node.
3. Select `seedream 5.0 pro`.
4. Supply one reference image and a normal, non-sensitive image-refinement prompt.
5. Request a 16:9 output around 1920x1080, 2048x1152, or 2560x1440.
6. Use the provider-default prompt optimization (standard mode) and run the workflow.
7. Observe that a slow request is disconnected at approximately 120 seconds even though the local client is prepared to wait longer.

## Observations

| Model / output | Prompt optimization | Observed result |
|---|---|---|
| Seedream 5 Pro, 1920x1080 reference edit | Provider default / standard | Disconnected at approximately 120.8 s on two observed runs |
| Seedream 5 Pro, 2560x1440 reference edit | Provider default / standard | Still waiting at exactly 120 s; the cloud request subsequently failed/skipped |
| Seedream 5 Pro, 1024x1024 face/reference edit | Provider default / standard | Four successful runs in approximately 32-108 s; two slow runs disconnected at approximately 121 s |
| Seedream 5 Lite, 2848x1600 reference edit | Standard | Completed successfully in approximately 48.4 s |

Representative exception:

```text
aiohttp.client_exceptions.ServerDisconnectedError: Server disconnected
```

The local workflow has a soft fallback that returns the input image after a cloud failure, so the ComfyUI task can appear successful or `SKIPPED` even though the paid Partner operation did not return an edited image. The underlying API log is what shows the disconnect.

## Actual behavior

Slow Seedream 5 Pro requests are terminated near the same 120-second boundary. Increasing the local client timeout does not extend the lifetime of the upstream connection.

## Expected behavior

All resolutions exposed by the official node should be able to complete reliably in standard mode. Long-running work should either:

- use an asynchronous submit-and-poll operation, or
- keep the Partner proxy connection alive long enough for the provider's documented resolutions and modes to complete.

The client should also receive an unambiguous final status so it does not need to guess whether a timed-out request may still have consumed credits.

## Related upstream work

Fast Mode was added in #15750 and backported in #15758. It is a useful speed workaround, but it does not make the documented standard mode reliable for larger Seedream 5 Pro reference edits. The provider documents standard mode as the higher-quality option, so users should not have to switch to Fast Mode solely to stay below an infrastructure deadline.

There are also reports of `Server disconnected` errors on other `api.comfy.org` Partner proxy endpoints, for example #14094.

## Attachments

- Screenshot showing the Seedream node still waiting at 120 seconds.
- Screenshot showing the cloud request being skipped/failed after the deadline.

Note: the second screenshot has an old user-customized node title containing the word `LITE`, but the model selector visible inside the node is `seedream 5.0 pro`.

No API key, authorization token, prompt containing personal information, or private source image is included.

Begin with the ByteDanceSeedreamNodeV2 entry point and the logged POST to https://api.comfy.org/proxy/byteplus/api/v3/images/generations. Trace how the Partner request lifetime and final status are handled around the approximately 120-second disconnect, using the reproduction steps and observations as the baseline. Done means standard Seedream 5 Pro requests receive an unambiguous result without being cut off at that boundary.

Written by the indexing model from the issue text.