docs(harbor): correct stale docs, docstrings and comments

[ShuxinLin]

Description

Fixes text in the Harbor port that later changes left out of date. These were found while reviewing #558 (item 7 of its known issues, plus the watsonx examples in item 5). Nothing changes behaviour except the wording of one error message.

Changes

src/assetops_harbor/stirrup.py

• Module docstring:
  • The example run no longer uses PYTHONPATH=agent, a watsonx model and --n-concurrent 16.
  • It no longer says the MCP servers get their environment through mcphub. The Stirrup runner passes it through mcp_server_env.
  • It no longer says the docker backend is "redundant" under Harbor. The code-sandbox overlay provides the daemon, and its code containers get neither CouchDB's hostname nor the forwarded credentials; local code gets both.
• Docker-backend error: it told users to mount a Docker socket, and now names overlays/code-sandbox.yaml instead. It still contains "no Docker daemon", which test_stirrup.py matches.
• _load_dotenv docstring: SETTING_ENV_VARS (FMSR_MODEL_ID) also reach the container.

src/agent/stirrup_agent/runner.py: the comment no longer mentions FMSR's "standalone watsonx default", which #583 removed.

benchmarks/harbor/README.md

• Step 4 used a watsonx model, which FMSR now rejects, so that run had no generate_failure_modes. It now uses litellm_proxy/azure/gpt-5.6-sol and says why. --n-concurrent goes from 16 to 4.
• Local backend: it said code runs next to the open scenarios' ground truth, which fix(harbor): build the runtime image from a clean clone of HEAD #587 took out of the image. It now says what that code does run next to: CouchDB, the forwarded credentials, and later the verifier.
• Explicit env: the section now names mcp_server_env and says plan-execute uses it too.
• Arms: the code_backend paragraph now describes the overlay path and the difference between the two backends.

benchmarks/harbor/CODE-SANDBOX.md: the run.sh -m example used the same watsonx model.

benchmarks/harbor/base-image/Dockerfile

• It referred to a check_models_list.sh that doesn't exist. It now says nothing checks models.txt against the catalog.
• The build line now matches the script's default tags.

generate_tasks.py and publish-images.sh: "corpus" becomes "suite" or "scenario data". publish-images.sh's example namespace is now quay.io/assetopsbench.

Left alone on purpose: the task template's comments (template/task.toml, template/environment/Dockerfile) are stale in the same way. Editing them changes every generated task's content hash, and Harbor then refuses to resume jobs started on the current tasks. They're better changed together with the next template change that does something.

• Documentation / Tutorial update

Testing

• uv run pytest src/assetops_harbor src/agent/tests/test_stirrup_mcp_env.py gives 24 passed.
• bash -n passes on publish-images.sh.
• ruff check finds nothing in the changed lines. stirrup_agent/runner.py already had an import-order warning and a BLE001, which this PR doesn't touch.

Checklist

• I have signed off my commits (DCO).