Regression: LoadImage / LoadImageMask false error after exiting subgraph in frontend >= 1.43.18

There seems to be a frontend regression affecting LoadImage / LoadImageMask when using subgraphs.

With frontend versions newer than 1.42.15, after entering a subgraph and then returning to the main graph, LoadImage / LoadImageMask can be marked as errored/red in the UI, even though the workflow executes successfully and the output image is generated correctly.

This does not happen with frontend 1.42.15.

Reproduction:

1. Open a workflow that uses a subgraph.
2. In the main graph, use LoadImage and LoadImageMask.
3. Feed IMAGE + MASK into the subgraph.
4. Run the workflow.
5. Enter the subgraph.
6. Exit back to the main graph.
7. LoadImage / LoadImageMask becomes red or shows an error state in the UI.
8. The backend still completes successfully and the final image is generated.

Expected behavior:

Exiting a subgraph should not mark LoadImage / LoadImageMask as errored if the referenced files exist and backend execution succeeds.

Actual behavior:

LoadImage / LoadImageMask is shown as errored/red after returning from the subgraph, even though generation succeeds.

Tested frontend versions:

Working:
- 1.42.15

Broken:
- 1.43.18
- 1.45.5
- 1.45.7
- experimental / nightly frontend

So this does not appear to be fixed in the latest experimental/nightly frontend. The last known working version is 1.42.15.

Possibly related UI change:

In newer frontend versions, the old clear/remove "X" button on LoadImage also disappeared.

I understand this may be an intentional UI change, but the more serious issue is the false error state on LoadImage / LoadImageMask after subgraph navigation.

Backend/system info:

ComfyUI backend:
- ComfyUI version: 0.21.1
- Required frontend version: 1.43.18
- Python: 3.12.10
- PyTorch: 2.12.0+cu130
- GPU: NVIDIA GeForce RTX 5060 Ti

Launch args:
- --normalvram
- --preview-method auto
- --use-sage-attention
- --auto-launch
- --enable-manager

Notes:

The same workflow, same files, and same backend work correctly when downgrading only the frontend to 1.42.15.

This strongly suggests a frontend regression related to subgraph navigation and LoadImage / LoadImageMask state validation.

I cannot currently provide a screenshot because I downgraded the frontend back to 1.42.15 to continue working, and with 1.42.15 the issue no longer appears.

[1FLUX_Inpainting_Controlnet_020526.json](https://github.com/user-attachments/files/27861184/1FLUX_Inpainting_Controlnet_020526.json)

┆Issue is synchronized with this [Notion page](https://www.notion.so/Issue-12313-Regression-LoadImage-LoadImageMask-false-error-after-exiting-subgraph-in-frontend--3626d73d365081dd81e7d92683df4866) by [Unito](https://www.unito.io)

Start by tracing the frontend state handling for LoadImage and LoadImageMask alongside subgraph entry and exit behavior. Reproduce the workflow with frontend 1.42.15 and a broken version such as 1.45.7, then compare when the false error state appears after returning to the main graph. Done means existing files remain valid and the nodes are not shown as errored when backend execution succeeds.

Written by the indexing model from the issue text.