Previously install.sh short-circuited with "Image ... already present -
skipping load" whenever a tag with the same version already existed in
Docker. That optimization was actively harmful: if an earlier deploy
attempt had loaded a wrong-arch (e.g. arm64) image at the same tag, we'd
never load the corrected tarball in the current ZIP and the arch check
would keep failing against the stale image forever.
docker load reassigns the tag atomically to whatever is in the tarball and
is a fast no-op when the layers are already present, so unconditional load
is both safe and self-healing.
install.sh now also prints the loaded image id + arch and, on mismatch,
tells the operator to `git pull` on the build host before re-running
package.sh so they pick up the buildx/binfmt fixes.
Co-authored-by: Cursor <cursoragent@cursor.com>
0506f39bf3