Fix the docker workflow and image build

The tag was built from github.repository_owner, which is "Exorust", but
registry names must be lowercase, so buildx rejected
ghcr.io/Exorust/torchleet before it built anything. Lowercase the owner when
composing the image name.

Also install torchvision from the same CPU wheel index as torch. Pulling it
from the default index can produce a build compiled against a different torch,
which fails at import with an obscure symbol error.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: Exorust

.github/workflows/docker.yml

@@ -14,13 +14,20 @@ jobs:
       packages: write
     steps:
       - uses: actions/checkout@v4
+
+      # Registry names must be lowercase, but the GitHub owner is "Exorust".
+      - name: Compute lowercase image name
+        id: img
+        run: echo "name=ghcr.io/${GITHUB_REPOSITORY_OWNER,,}/torchleet" >> "$GITHUB_OUTPUT"
+
       - uses: docker/login-action@v3
         with:
           registry: ghcr.io
           username: ${{ github.actor }}
           password: ${{ secrets.GITHUB_TOKEN }}
+
       - uses: docker/build-push-action@v6
         with:
           context: .
           push: true
-          tags: ghcr.io/${{ github.repository_owner }}/torchleet:latest
+          tags: ${{ steps.img.outputs.name }}:latest

Dockerfile

@@ -1,11 +1,13 @@
-# Zero-install TorchLeet: notebooks + the grader, no local Python setup.
+# Zero-install TorchLeet: notebooks plus the grader, no local Python setup.
 #   docker run --rm -p 8888:8888 ghcr.io/exorust/torchleet
 FROM python:3.12-slim
 
-# CPU-only torch keeps the image small; a GPU box can pip install a CUDA build.
-RUN pip install --no-cache-dir \
-      torch --index-url https://download.pytorch.org/whl/cpu \
- && pip install --no-cache-dir jupyterlab torchvision matplotlib numpy
+# torch AND torchvision both come from the CPU wheel index. Taking torchvision
+# from the default index instead risks a build compiled against a different
+# torch, which fails at import with an obscure symbol error.
+RUN pip install --no-cache-dir --index-url https://download.pytorch.org/whl/cpu \
+      torch torchvision \
+ && pip install --no-cache-dir jupyterlab matplotlib numpy
 
 WORKDIR /torchleet
 COPY . /torchleet