Experiment: dummy_jz

Design

No hypothesis. Plumbing check: sleep 30s on a Jean Zay H100 node (gpu_p6, -C
h100, qos_gpu_h100-t3), write {"dummy": 1}, and verify submit → fetch →
report works on this cluster.

Results

20260726-202904-dummy_jz-2fa7134

commit: 2fa7134

metric	value
dummy	1

Interpretation

The pipeline works on Jean Zay H100. Getting there needed three attempts and
surfaced cluster-specific setup facts worth keeping for real runs:

• H100 is selected by constraint -C h100 + --partition=gpu_p6 +
  --qos=qos_gpu_h100-t3, not by partition alone. (xp gained constraint:
  and qos: config fields for this.)
• Non-login batch shells have no module function; the init path I tried
  (/etc/profile.d/z_modules.sh) does not exist. For anything needing modules
  we must find the correct init or use a login shell. This dummy is pure
  stdlib, so base python3 sufficed and no module load was needed.
• $WORK/$HOME are not expanded by rsync; workdir must be an absolute path.
• Default Slurm account was accepted (no --account needed).
  Nothing here bears on the research question.