daily-fix: fail-loud CVD assert in issue667_extract _device()

Overview / Motivation

Auto-filed by /daily 2026-07-03 from a recursion-guard-parked workflow-fix candidate.

Goal

_device() asserts CUDA_VISIBLE_DEVICES is set in the environment when gpu_id > 0, raising a self-explaining error naming the env-pin recipe instead of silently returning cuda:0

Workflow gap

• Bug observed: a hand-launch of a launcher-pinned per-cell worker with only --gpu-id silently targets the busy default GPU and crashes at vLLM init_device (#813); only prose-guarded after #880
• Why it is a workflow gap: experiment-code follow-up (out of wf-fix surface); a fail-loud assert binds on ALL launch paths, prose read or not
• Confidence (emitter): medium

Proposed change (candidate sketch — refine in planning)

_device() asserts CUDA_VISIBLE_DEVICES is set in the environment when gpu_id > 0, raising a self-explaining error naming the env-pin recipe instead of silently returning cuda:0

Scope / surfaces

• Primary target: scripts/issue667_extract.py
• Grep the workflow surface for the pattern before editing; list every hit in the plan.

Constraints / invariants

• Workflow-surface only; scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.

Provenance

• workflow_fix_target: scripts/issue667_extract.py
• fingerprint: 2cad43b5152c

parked — running under workflow_fix_target recursion guard, see .claude/rules/workflow-fix-on-bug.md § Recursion guard. LOGGED + surfaced, NOT auto-routed by this session.

target_file: scripts/issue667_extract.py (and _device()-style descendants — NOTE: experiment scripts, OUT of the workflow-fix surface; this is an experiment-code follow-up, not a workflow-fix task) bug_observed: a hand-launch of a launcher-pinned per-cell worker with only --gpu-id silently targets the busy default GPU and crashes at vLLM init_device (#813, 2026-07-02) — the failure is only prose-guarded after #880. why_workflow_gap: (borderline/out-of-scope by target) the worker class has no runtime guard; a fail-loud assert would bind on ALL launch paths, prose read or not. proposed_change: _device() (scripts/issue667_extract.py::_device) asserts CUDA_VISIBLE_DEVICES is set in the environment when gpu_id > 0, raising a self-explaining error naming the env-pin recipe instead of silently returning cuda:0. diff_sketch: | def _device(gpu_id, cpu_only): ...

• if not cpu_only and gpu_id > 0 and "CUDA_VISIBLE_DEVICES" not in os.environ:

•   raise RuntimeError(
  

•       f"--gpu-id {gpu_id} is informational in this worker class; launch with "
  

•       f"env CUDA_VISIBLE_DEVICES={gpu_id} ... --gpu-id {gpu_id} (see gotchas.md manual-launch CVD entry)")
  

  return torch.device("cuda:0") confidence: medium related_task: #880 (raised by both alternatives reviewers + the reconciler as the stronger structural fix, out of the #880 workflow-fix surface)