fix(build): honor loader.model_class on pretrained load path

[xieofxie]

Problem

winml build -c <build_config.json> -m openai/clip-vit-base-patch32 --device cpu fails with:

Error: You have to specify pixel_values

even though the build config explicitly requests the text-only CLIP variant:

"loader": {
  "task": "feature-extraction",
  "model_class": "CLIPTextModelWithProjection",
  "model_type": "clip_text_model"
}

Root cause

Bisected to #836 (3946a01, "Enable static quantization for Qwen3-0.6B decoder").

The pretrained branch of _load_model (build/hf.py) called load_hf_model without forwarding config.loader.model_class, so class resolution depended entirely on the (model_type, task) lookup in MODEL_CLASS_MAPPING.

#836 began threading config.loader.model_type into load_hf_model as a build-variant override. For CLIP this regresses resolution:

• The build config stores the variant tag model_type = "clip_text_model".
• MODEL_CLASS_MAPPING is keyed on the native type: ("clip", "feature-extraction") -> CLIPTextModelWithProjection.
• The override "clip_text_model" is not a key, so resolution falls through to TasksManager → AutoModel → the full CLIPModel, which requires pixel_values. Export with text-only inputs then fails.

The override scheme works for qwen3_transformer_only because that variant tag is a registered key; CLIP's is not, and the explicit model_class that would have saved it was dropped on the pretrained path (the random-init path already honors it).

Fix

Forward config.loader.model_class to load_hf_model on the pretrained path. resolve_task then takes its Stage-0 user-class path, resolving CLIPTextModelWithProjection directly and ignoring the model_type override. Safe for the qwen3 transformer-only variant since Stage 0 only activates when model_class is set.

Verification

• Repro command now completes (exit 0), loading CLIPTextModelWithProjection and exporting text-only inputs through optimize + fp16.
• Added regression test test_pretrained_load_threads_model_class.
• tests/unit/build/, tests/unit/loader/test_load_hf_model.py, tests/unit/commands/test_build.py all pass (192 + new test).