fix(harbor): build the runtime image from a clean clone of HEAD

[ShuxinLin]

Description

base-image/Dockerfile does COPY . ., so building from the working tree baked in whatever was there. That included an untracked reports/results_table.csv with a ground_truth column for 56 private scenarios (8 of mini's 35, 41 of lite's 76). The file was in /opt/aob of assetopsbench/runtime:dev, where any agent running code in main could read it, and publish-images.sh would have pushed it to Docker Hub.

This PR applies two guards:

1. Clean build. The new scripts/build-runtime-image.sh builds from a clone of HEAD that contains only that commit, so untracked and git-ignored files never enter the build.
2. .dockerignore. It is the second guard, and the only one for a plain docker build ..

Fix Details

• scripts/build-runtime-image.sh
  • Fetches HEAD at depth 1 into a temp directory, checks the clone is at the expected commit, builds with docker buildx build, and deletes the clone.
  • With no arguments it builds assetopsbench/runtime:dev --load. Any arguments go to buildx unchanged.
  • Warns when tracked files have uncommitted changes, because those are not built.
  • The clone keeps a real one-commit .git, so git rev-parse in StirrupAgent.get_version_command still works. A git worktree wouldn't work, because its .git is a pointer that doesn't resolve inside the container.
• publish-images.sh builds the runtime image through the script. run.sh's error message, README, QUICKSTART, CODE-SANDBOX and the Dockerfile header now point at the script too.
• .dockerignore
  • Uses **/.env / **/.env.*; the bare .env only matched at the top of the repo.
  • Leaves out the open scenarios' groundtruth*, rubric.json, reference_answer.json and scenario_meta.json. The verifier reads the copies under /tests, never these.
  • Leaves out reports/, logs/, src/tmp/, .claude/ and the kdd_tutorial folders.

Depends on #586 for the open profile. On this branch the generator still copies each scenario's groundtruth.txt into the task's build folder, so the per-task layer puts it back. #586 stops that. The two branches merge cleanly (checked with git merge-tree).

Impact on Benchmarking

• No change to baselines: the image keeps the same code and dependencies. Only local files and answers are removed.
• Baseline change

Related Issues

• Found while reviewing feat(harbor): run AssetOpsBench scenarios as Harbor tasks #558 (finding 6; also covers the open-profile half of finding 3 together with fix(harbor): run.sh mounts only the private suite's shared/ #586)

Verification Steps

1. uv run pytest: not run. No Python changed.
2. Manual verification:

   • Rebuilt assetopsbench/runtime:dev with the script from b7a7d84. It is 8.79 GB, down from 8.92 GB, and the temp clone was removed afterwards.

   • Old image vs new image:

     	old image	new image
     reports/results_table.csv	present	absent
     logs/, src/tmp/, .claude/, kdd_tutorial folders	present	absent
     Answer files in /opt/aob	3	0
     git rev-parse --short HEAD	0c45322	b7a7d84
     Commits in .git	1142	1
     .git size	157M	44M

   • harbor run --agent oracle on the open profile with the new image: 3 trials, 0 exceptions, reward 1.000. The verifier scores correctly without answers in the image.

   • With fix(harbor): run.sh mounts only the private suite's shared/ #586's generator and overlay, the tsfm-1019 task container on the new image has no answer-looking file anywhere, and the data load filled all 5 collections.

   • bash -n passes for all three scripts. publish-images.sh was not run (it pushes).

Checklist

• I have added tests that prove my fix is effective.
• My code follows the project's Ruff formatting and linting rules.
• I have signed off my commits (DCO).

[ShuxinLin]

Second review pass, addressed in 47a806e:

• No .git in the image. The one-commit .git still held every committed blob, so git show HEAD:…/groundtruth.txt returned the open scenarios' answers. Its reflog also recorded the builder's email, and its index made COPY . . miss the cache on every build. The script now builds from git archive HEAD and passes AOB_COMMIT; the Dockerfile writes it to /opt/aob/.aob-commit, and StirrupAgent.get_version_command reads that file.
• A plain docker build now fails. Without AOB_COMMIT, a small commit stage fails the build in about a second with a pointer to the script. The ARG lives only in that stage, so a new commit invalidates just the final layer.
• Script defaults: it keeps -t assetopsbench/runtime:dev and --load unless the caller passes a tag or an output. It also warns about untracked files it leaves out, and checks for buildx.

Verification:

• .git absent; git show fails.
• .aob-commit = 47a806e…, and the version command prints 47a806e.
• No answer files in the image.
• Server, evaluation and init_data modules import.
• A second build of the same commit is fully cached (1 s).
• Unit tests: the same 19 pre-existing failures as the base, none new.

Not changed:

• The snap-Docker /tmp and sudo safe.directory edge cases.
• Moving the model preload above COPY . ., which is a separate speed-up.