test(evals): isolate required backend preflight fixture

[aviggiano]

The evaluation runner's missing-backend test used the packaged invariant profile. After #1120 made strategy groups continue on failure, that backend became optional and the test reached an unrelated reference-cache error on clean CI runners.

Use a local topology with one required fixture task, an inert prompt, and a deliberately absent command. This preserves the real blocking preflight, exact error, and no-run/no-workflow assertions while removing dependence on reference caches. Production code is unchanged.

Validation: full evaluation suite passed all 450 tests; the focused regression, strict lint, formatting, and diff checks passed. An independent agent reviewed the change with no findings. This resolves the package-gate failure found during v0.0.25 release validation.

Greptile Summary

This PR isolates the missing-backend preflight regression test from packaged strategy groups and reference caches.

• Replaces the packaged invariant-profile dependency with a minimal local topology.
• Adds an inert prompt and deliberately unavailable required command.
• Preserves assertions for the exact diagnostic and absence of run or workflow creation.

Confidence Score: 5/5

The PR appears safe to merge; the isolated fixture preserves the intended blocking preflight behavior without relying on unrelated packaged resources.

The replacement topology is valid, its command is deterministically absent under the test’s local PATH, and the existing assertions continue to verify failure before any run or workflow is created.

Important Files Changed

Filename	Overview
packages/evals/test/runner-publish.test.ts	Reworks the required-command preflight test around a valid, self-contained topology and prompt fixture without changing production behavior.

_{Reviews (1): Last reviewed commit: "test(evals): isolate required backend pr..." | Re-trigger Greptile}

Context used:

• Knowledge Base — Evaluation runs and history
• Knowledge Base — Evaluation and benchmarking

[aviggiano]

Runtime shard 3 passed all 74 selected tests with zero failures. The job failed afterward because GitHub's artifact service timed out (Failed to CreateArtifact: ETIMEDOUT) while uploading its validation report.

No code correction is needed for this failure. A job-only rerun will retain the successful suites once GitHub permits the rerun after the active workflow finishes. The full runtime support/adapter suite already passed.