dockerfile: track latest cadgenbench pre-v1

Install cadgenbench from main so the eval image picks up the Open3D alignment dependency before the v1 release pin is set.

Dockerfile

@@ -46,16 +46,11 @@ RUN python -m ensurepip --upgrade \
     && python -m pip install --no-cache-dir --upgrade pip
 
 # cadgenbench from the Public GitHub repo, same convention and ARG name
-# as the leaderboard Dockerfile. Defaults to `main` (every rebuild picks
-# up latest, pre-v1); lock to a commit SHA at the v1 release for
-# reproducible scores (see space-setup/post-gt-swap.md Stage F).
-ARG CADGENBENCH_SHA=289a2fc
-# Pinned to an explicit commit SHA, so the pip layer below is already
-# deterministic: changing the ARG changes the install command and busts
-# the cached layer on its own. (We previously cache-busted with an
-# `ADD https://api.github.com/.../commits/${SHA}`; GitHub's API now 403s
-# on the build IP, failing the build, and it was redundant for a fixed
-# SHA -- removed. Re-pin by bumping the ARG above, not via that ADD.)
+# as the leaderboard Dockerfile. Defaults to `main` so every pre-v1 image
+# rebuild picks up the latest evaluator dependencies (including Open3D for
+# alignment). Lock to a commit SHA at the v1 release for reproducible scores
+# (see space-setup/post-gt-swap.md Stage F).
+ARG CADGENBENCH_SHA=main
 RUN python -m pip install --no-cache-dir \
         "cadgenbench @ git+https://github.com/huggingface/cadgenbench.git@${CADGENBENCH_SHA}"
 

README.md

@@ -38,7 +38,9 @@ against this image on `a10g-large` and polls for completion.
   leaderboard Dockerfile) plus `libegl1 libegl-mesa0` for the EGL
   context.
 - `cadgenbench @ git+https://github.com/huggingface/cadgenbench@<sha>`,
-  pinned via `ARG CADGENBENCH_SHA`.
+  controlled via `ARG CADGENBENCH_SHA` (pre-v1 defaults to `main`; pin a SHA
+  at release). This installs cadgenbench's Python dependencies, including
+  Open3D for rigid alignment.
 - `vtk-egl` swapped in for the PyPI `vtk` wheel (same swap shape as
   the leaderboard's `vtk-osmesa`; the GPU counterpart). PyVista
   picks up whichever `vtk` dist is installed; no cadgenbench code