Replace segmentation model type heuristic with arch field

[andriiryzhkov]

Summary

• Add arch field to dt_ai_model_info_t (backend layer), parsed from model config.json
• Replace decoder output count heuristic in segmentation.c with explicit arch lookup ("sam2" or "segnext")
• Models must declare "arch" in their config.json — missing arch fails with a clear error message

Motivation

The previous approach detected SAM vs SegNext by counting decoder outputs (3+ = SAM, 1 = SegNext). This is fragile and would break when adding new model architectures with different output counts. Using an explicit arch field makes model type detection reliable and extensible.

[TurboGit]

But then we need an update of data/ai_models.json with the new arch field, no?

Ok, found the PR on darktable-ai adding the arch field. I need to enable the notification for darktable-ai as I continuously fail to see the changes there :(

[TurboGit]

@andriiryzhkov :

In the new models created I don't have the arch in config.json:

$ cat mask-object-sam21-small/config.json 
{
    "id": "mask-object-sam21-small",
    "name": "mask sam2.1 hiera small",
    "description": "Segment Anything 2.1 (Hiera Small) for interactive masking",
    "task": "mask",
    "backend": "onnx",
    "version": "1.0",
    "tiling": false
}

And then creating the mask fails, the message on the console:

14.1144 [segmentation] unknown arch '' for mask-object-sam21-small

If I add manually the arch into config.json all this working.

[andriiryzhkov]

I see, missed to add in conversion script. Will update shortly

[andriiryzhkov]

darktable-org/darktable-ai#9

[TurboGit]

Thanks!