fix(docker): validate VLABench asset download + install hf_xet

The previous HF Hub snapshot_download step could complete silently
without actually populating the asset tree — several files in
lerobot/vlabench-assets are Xet-stored, and the stock huggingface_hub
install in the base image can skip them. The first runtime signal was
an IndexError deep inside VLABench's task builder (random.choice on an
empty XML list in config_manager.py), far from the real cause.

- Install huggingface_hub[hf_xet]>=0.26 before the snapshot download
  so Xet-stored assets are fetched reliably.
- Scope the download with allow_patterns=['obj/**', 'scenes/**'] (the
  only subtrees the env needs), reducing image size and surface area.
- Add a build-time validator: after the download, walk four
  task-critical subtrees (plates, basket, fruit, tray) and fail the
  build loudly with the list of empty dirs if any is missing XMLs.
  Prints XML counts when successful so CI logs confirm completeness.

Made-with: Cursor

docker/Dockerfile.benchmark.vlabench

@@ -42,8 +42,9 @@ p.write_text(t.replace( \
       open3d colorlog scikit-learn openai gdown
 
 # Download VLABench mesh assets. Task configs reference object meshes
-# (obj/meshes/fruit/, containers/basket/, etc.); without them the task builder
-# picks from an empty mesh list and crashes at task-build time.
+# (obj/meshes/fruit/, containers/basket/, tablewares/plates/, etc.); without
+# them the task builder picks from an empty mesh list and crashes with
+# IndexError at task-build time (random.choice([]) in config_manager.py).
 #
 # Preferred source: an HF Hub mirror. Set VLABENCH_ASSETS_REPO at build time
 # (e.g. --build-arg VLABENCH_ASSETS_REPO=lerobot/vlabench-assets) and we'll
@@ -51,20 +52,36 @@ p.write_text(t.replace( \
 # path for CI — Google Drive frequently returns HTTP 429 ("Too many users have
 # viewed or downloaded this file recently") on shared academic files.
 #
-# Fallback: VLABench's own gdown-based script. Best-effort only; the build
-# will NOT fail if gdown hits a Drive quota, letting us ship an image where
-# task-build may still IndexError but the rest of the pipeline is exercised.
+# After download we *validate* that at least one XML exists under each
+# task-critical subtree and fail the build loudly if not. Silent-empty asset
+# dirs are the #1 cause of VLABench runtime crashes in CI, so we surface them
+# here rather than after a 10-minute eval build.
+#
+# Fallback: VLABench's own gdown-based script. Best-effort only.
 ARG VLABENCH_ASSETS_REPO=""
 RUN ASSETS_DIR="$HOME/VLABench/VLABench/assets" && \
     if [ -n "${VLABENCH_ASSETS_REPO}" ]; then \
         echo "Downloading VLABench assets from HF Hub: ${VLABENCH_ASSETS_REPO}" && \
+        uv pip install --no-cache "huggingface_hub[hf_xet]>=0.26" && \
         python -c "from huggingface_hub import snapshot_download; \
-snapshot_download(repo_id='${VLABENCH_ASSETS_REPO}', repo_type='dataset', local_dir='${ASSETS_DIR}')"; \
+p = snapshot_download(repo_id='${VLABENCH_ASSETS_REPO}', repo_type='dataset', \
+    local_dir='${ASSETS_DIR}', allow_patterns=['obj/**', 'scenes/**']); \
+print('snapshot_download returned:', p)"; \
     else \
         echo "No VLABENCH_ASSETS_REPO set — falling back to gdown (best-effort)" && \
         python ~/VLABench/scripts/download_assets.py --choice all || \
-          echo "WARN: VLABench asset download failed (likely Google Drive quota). Env task-build may crash."; \
-    fi
+          echo "WARN: VLABench asset download failed (likely Google Drive quota)."; \
+    fi && \
+    python -c "\
+from pathlib import Path; \
+import sys; \
+root = Path('${ASSETS_DIR}'); \
+checks = ['obj/meshes/tablewares/plates', 'obj/meshes/containers/basket', 'obj/meshes/fruit', 'obj/meshes/containers/tray']; \
+failed = []; \
+print(f'Validating VLABench assets under {root}'); \
+[print(f'  {c}: {len(list((root/c).rglob(\"*.xml\")))} XMLs') for c in checks]; \
+[failed.append(c) for c in checks if not any((root/c).rglob('*.xml'))]; \
+sys.exit(f'Empty asset dirs (no *.xml): {failed}') if failed else print('All asset dirs populated.')"
 
 # Overlay the PR's source code on top of the nightly image.
 COPY --chown=user_lerobot:user_lerobot . .