feat(harbor): run a full scenario suite in parallel

[ShuxinLin]

Description

Runs a full scenario suite through Harbor the way benchmarks/run.sh runs it
serially, with the same all profile, the same stirrup-agent and the same
Docker code sandbox, but with scenarios in parallel, each trial on its own
CouchDB.

Stacked on #558 (base: feature/harbor-integration). #558 covers the open
profile, whose scenarios ship in the repo. An external suite such as
AssetOpsBenchScenarioGeneration/scenarios_data did not fit its template for
three reasons:

• Its shared/ data is 2.1 GB (mostly shared/iot), and every manifest
  resolves against it. The runtime image only carries the repo's small copy.
• all.yaml lists a scenario the suite lacks (wosr-62), and the generator
  aborted on the first missing folder.
• Nothing drove it end to end: building images, generating tasks, one job per
  model, and resuming.

Changes

• adapter/generate_tasks.py
  • --runtime-image: the image each task builds FROM.
  • --data-dir: where the suite lives in that image. The per-task layer copies
    the scenario there, and the healthcheck runs
    SCENARIOS_DATA_DIR=<dir> init_data.py <id>. The agent keeps the repo copy,
    as scenario_suite_runner does; it sets that variable for the data load only.
  • --skip-missing: warn and skip profile entries that aren't in the suite.
  • Fix: every task used to get scenario 1's description and the wosr
    keyword.
• suite-image/Dockerfile: bakes the suite into one layer over
  assetopsbench/runtime:dev, shared by all tasks. A bind mount would go
  through Rancher/Docker Desktop's file sharing, which is slow.
• run.sh: the Harbor counterpart of benchmarks/run.sh. It uses the same
  -s and -l flags and the same model list, plus -n for concurrency, -p
  for the profile and a repeatable -m "MODEL EFFORT". It builds the suite and
  sandbox images, regenerates the dataset, and runs one Harbor job per model at
  <leaderboard>/harbor-jobs/stirrup_agent__<model>. Re-running resumes that job,
  the equivalent of --skip-existing. Credentials reach Harbor only through
  uv run --env-file .env.
• README: a section on running a full suite.

Type of Change

• Infrastructure / Tooling Improvement

Industry Relevance

The published leaderboard runs all 232 all-profile scenarios per model, one
at a time against a shared CouchDB. With Harbor each scenario gets an isolated
database, so the same run parallelizes safely.

Related Issues

• Refs: experiment harbour design for AssetOpsBench #557, feat(harbor): run AssetOpsBench scenarios as Harbor tasks #558

Testing & Validation

Run locally: Rancher Desktop with 8 CPUs and 24 GB, against
AssetOpsBenchScenarioGeneration/scenarios_data with the all profile.

• Task generation: 231 tasks from 232 profile entries (wosr-62 skipped).
  17 are scored fmea and 214 static_json, from each scenario_meta.json.
  Every file the manifests reference exists. The largest scenario loads 9.2 MB,
  well inside the 300 s healthcheck.

• Oracle, real containers: one scenario per category, --n-concurrent 4.
  7/7 passed, 0 exceptions, reward 1.000, in 54 s.

• Stirrup smoke via run.sh: litellm_proxy/aws/claude-opus-5 at high,
  code sandbox on, -n 4, one scenario per category. Every trial had its own
  couchdb and dind containers and no host ports. Agents called the MCP tools
  (iot, wo, fmsr, tsfm, utilities) and ran code_exec in the Docker
  sandbox, and spilled tool results were readable through /workspace-share.

  Task	Reward	Passed	Input / output tokens
  car-151	1.0	✅	305k / 3.2k
  fcc-301	0.857	❌	313k / 10.3k
  fmea-9001	0.933	✅	203k / 14.8k
  fmsr-901	0.2	❌	1.04M / 15.2k
  health-401	1.0	✅	193k / 1.4k
  tsfm-1001	1.0	✅	623k / 10.6k
  wosr-1	0.0	❌	462k / 13.5k

  4/7 passed, mean reward 0.713, wall time 18 m 10 s.

  No harness exceptions. The three misses are the model's answers, not
  infrastructure failures. For example, on wosr-1 the agent's sandbox code read
  all 1,076 work orders from the trial's CouchDB and counted 163 failures
  against a ground truth of 211.

• Harbor healthcheck semantics: BaseEnvironment.run_healthcheck returns
  on the first success, so init_data.py (drop and reload) runs once per trial
  and never again during the agent's run.

• Data Integrity: no suite data or credentials are committed. Generated
  tasks, datasets and job output are gitignored or outside the repo.

• Full 231-scenario run: not started yet.

Checklist

• ruff check and ruff format pass on the changed Python file.
• I have performed a self-review of my code.
• I have updated the documentation (benchmarks/harbor/README.md).
• I have signed off my commits (DCO).

[ShuxinLin]

Superseded by #560. Renaming the head branch to feat/harbor-suite-run (corpus → suite) closed this PR; same commits, now under the new branch name.