fix(pd-lm): SIGPIPE race in node-setup driver check kills 16-node jobs

[claude-spd1]

The per-node workspace setup runs under set -euo pipefail and gated the CUDA extra on nvidia-smi --query-gpu=driver_version ... | head -n 1 | cut .... On an 8-GPU node nvidia-smi prints 8 lines; head -n 1 exits after the first, and when nvidia-smi loses the race it dies of SIGPIPE (exit 141) → pipefail fails the pipeline → set -e kills the srun task → SLURM cancels all 16 nodes seconds into the job.

Observed failures with this exact signature: jobs 172348 and 172349 (dan.braun, Jul 8) and 174502 and 174513 (pd-user, Jul 13) — roughly a coin flip per 16-node launch, always written off as "node flakiness, retry".

Fix: query only GPU 0 (nvidia-smi --id=0) — single output line, nothing truncates the pipe, no behavior change otherwise.

Verified: job 174520 (16 nodes, memprobe for run p-077fe120) launched with the patched setup script; all 16 nodes passed the driver check and the run completed.

Context: task:profile-llama-32b-cw-east-artifacts

Crew-Address: task/profile-llama-32b-cw-east-artifacts

🤖 Generated with Claude Code

[claude-spd1]

Independent confirmation from task workspace-fetch-sigpipe-flake (the same flake reported as ~1-in-6 on pd-lm launches): on an and-gmi 8×H200 node (driver 570.133.20) I ran the exact unpatched pipeline from _node_workspace_setup 300 times under set -euo pipefail: 57/300 (19%) exited 141. The patched variant from this PR (--id=0): 0/300. That per-invocation rate compounds per node, which explains both the ~coin-flip 16-node failures and the ~1-in-6 smaller launches.

Crew-Address: task/workspace-fetch-sigpipe-flake