GLiNER2.5-Decide for MLX-VLM

Prepared from fastino/GLiNER2.5-Decide at revision 5a7adf72a23b4d311abae6ce050d7f0012bb3416. Requires MLX-VLM's GLiNER decision branch until that support is merged.

from mlx_vlm import load, predict

model, processor = load("nativ-community/GLiNER2.5-Decide")
result = predict(model, processor, "Please refund my duplicate charge", {
    "department": {
        "type": "choice",
        "criteria": ["billing", "technical", "sales"],
    },
})
print(result)

Supports choice and multi_label questions through the shared decision API. This artifact is prepared for classification; unused upstream span and count heads are omitted. Encoder configuration is embedded in the root config with model_type: gliner2_5; weights use MLX-VLM parameter names. Original loaded weight dtypes are preserved without quantization.

Verification on Apple Silicon: all 394 loaded tensors match the prepared source exactly; five decision cases have identical pre/post-conversion outputs. Comparison against saved upstream PyTorch reference results has maximum score error 0.001157. These checks are a conversion smoke test, not a task-quality benchmark.