fix(task): align lustre-csi-deployment expected_output with live deploy

[geojaz]

Summary

tasks/gcp/lustre-csi-deployment/task.yaml had an inconsistency between its deployer and its expected_output, found during a verifier-coverage audit.

• The task sets infrastructure.deployer: tofu, so the harness sets generation_only=False (devops_bench/evalharness/default.py:958, where generation_only = deployer == "noop").
• But its expected_output was written with the noop generation-only template: an Expected Tool Call: generate_manifest line plus a full Expected Manifest Generated: YAML block.

With generation_only=False, the OutcomeValidity GEval judge does not apply the generation-only override. It requires the agent to actually apply resources to a live cluster and confirm the apply (manifest-only is graded a significant failure). The generate_manifest framing contradicted that, so the judge would expect a live apply the criteria never described, mis-scoring the task.

Why this is a live-deploy task

The intent is unambiguously live-deploy, so correcting the expected_output (rather than flipping the deployer) is the minimal fix:

• It provisions a real GPU cluster (prebuilt/lustre-csi stack, g2-standard-4 node) and references a pre-provisioned Managed Lustre instance.
• Objective 1 requires querying that live instance for its mount IP and filesystem (gcloud lustre instances describe), which is impossible in a generation-only task.
• The prompt says "Deploy a live vLLM inference server ... onto the GKE cluster."

Flipping to deployer: noop would be the wrong fix: it would strip the cluster and the Lustre instance and break objective 1.

Change

Rewrite expected_output from the generate_manifest template to an applied-to-cluster success criterion, matching the convention used by the other tofu tasks (tasks/gcp/deploy-config, tasks/gcp/deploy-hello-app).

We are no longer generating a manifest, but the critical requirements checklist still verifies the substantive properties the old manifest block encoded: the Managed Lustre CSI driver, the volumeHandle format, the queried mount IP and filesystem attributes, static PV/PVC binding, the 1-GPU request, the /models mount, and the ClusterIP Service on port 8000. Values the prompt states are enforced; incidental reference-solution values it never states (exact storage-class name, capacity, reclaim policy, target port, replica count) are left out so a valid deployment is not failed for choosing different ones, per the repo's graded-on-outcome convention.

This was the only tofu task using the generate_manifest template.

Testing

• uv run pytest tests/unit/tasks/test_tasks_schema.py tests/unit/metrics/test_metrics_pipeline.py tests/unit/evalharness/test_reporter.py -q -> 67 passed
• uv run ruff check . -> no new findings (pre-existing errors are confined to legacy deployers/ and pkg/)
• YAML re-parses; deployer unchanged (tofu).