Warn that VRAM exhaustion masks --gpus fixes

If PaddleOCR (or any GPU sidecar) holds all VRAM, --gpus will still
appear to fail even when configured correctly. Added explicit callout:
check nvidia-smi before concluding the GPU flag is wrong.

Transcodes/Manual-Transcoding.md

@@ -575,6 +575,16 @@ In EmbyCredits plugin settings, set the OCR endpoint to `http://localhost:8884`.
 The accuracy tradeoff is real (PaddleOCR is stronger on non-Latin scripts) but GPU
 starvation is not an acceptable failure mode for a live media server.
 
+**If you just switched from `--runtime=nvidia` to `--gpus` and it still fails:**
+
+Check `nvidia-smi` before concluding `--gpus` is wrong. The flag may be correct and
+something else may be holding all the VRAM. This is exactly what happens when the
+EmbyCredits PaddleOCR container is running — PaddleOCR steals all available VRAM
+while a credit scan runs and never releases it when the scan completes. Emby and
+Jellyfin then have no GPU access regardless of how the GPU is passed to their
+containers. The GPU assignment method is irrelevant when there is no VRAM left to
+assign. Always confirm VRAM is actually free before troubleshooting container GPU flags.
+
 ---
 
 ### Emby With NVIDIA GPU — bind-propagation=shared Fails to Start