Fix Qwen3-VL normalization: use 0.5/0.5/0.5 from checkpoint (#105)

The qwen3-vl VLMSpec hardcoded OpenAI-CLIP normalization stats, but
Qwen3-VL checkpoints specify image_mean=image_std=[0.5, 0.5, 0.5].
This caused a silent ~1.86x overscale on every pixel fed to the
vision encoder.

Fixes #82.

Author: stikves

python/src/coreai_models/vlm/export.py

@@ -81,8 +81,8 @@ def num_visual_tokens(self) -> int:
         patch_size=16,
         spatial_merge_size=2,
         temporal_patch_size=2,  # Qwen frames-per-image (single image -> duplicated)
-        image_mean=(0.48145466, 0.4578275, 0.40821073),
-        image_std=(0.26862954, 0.26130258, 0.27577711),
+        image_mean=(0.5, 0.5, 0.5),
+        image_std=(0.5, 0.5, 0.5),
         rescale_factor=1.0,
     ),
 }