Pin Tower-compatible CUDA runtime

docs/CODEX_EXECUTION_LOG.md

@@ -12230,3 +12230,4 @@ Open:
 ### Verificatie
 - TypeScript- en Vite-productiebuild geslaagd.
 - Gerichte frontend-, backend-, browser- en productieverificatie volgen hieronder na de releasegate.
+- De eerste productieacceptatie vond een echte CUDA-compatibiliteitsregressie: `torch 2.13.0+cu130` zag de RTX 4080 SUPER maar weigerde initialisatie op de Tower-driver met CUDA 12.9-capaciteit. De AI-image is daarom teruggebracht naar de expliciet gepinde, drivercompatibele combinatie `torch 2.11.0` / `torchvision 0.26.0` via de officiële CUDA 12.8-index; promotie vereist opnieuw live `torch.cuda.is_available()` en modelpreflight.

docs/DEPENDENCY_POLICY.md

@@ -46,12 +46,14 @@ not be installed by the default Docker backend image unless an explicit AI image
 or profile is introduced later.
 
 The opt-in Unraid all-in-one AI build is NVIDIA-GPU-oriented. It installs the
-pinned PyTorch/torchvision pair from the CUDA 13.0 wheel index before installing
+pinned PyTorch/torchvision pair from the CUDA 12.8 wheel index before installing
 the `ai` extra. The production runtime exposes the NVIDIA device, selects
 `YOLO_DEVICE=cuda:0` and sets `YOLO_REQUIRE_CUDA=true`, so missing CUDA fails
 closed instead of silently falling back to CPU. The index and versions remain
 explicit Docker build arguments and require live driver/runtime validation on
-Tower before release promotion.
+Tower before release promotion. CUDA 12.8 is deliberately below the Tower
+driver's reported CUDA 12.9 capability; a newer wheel index may not be promoted
+merely because it exists when `torch.cuda.is_available()` fails on that driver.
 
 Docker dependency metadata is copied before application source. Backend source
 changes therefore reuse the dependency layer while changes to `pyproject.toml`