Ask SLURM for what the job holds when the fit is launched from the login node (#642) - #645

Fixes #642.

What was wrong

On a cluster where salloc returns a shell on the login node while the allocation is held on a compute node — which is the shell PyBNF is meant to run in, since it is the one that holds the allocation — the srun launcher stopped before a single worker started:

srun: error: Unable to create step for job NNNNN: More processors requested than permitted

Cluster.cpus_per_node prefers $SLURM_CPUS_ON_NODE, which SLURM sets only inside a job step running on an allocated node. On the login node it is simply absent, so the count fell through to dask.system.CPU_COUNT and then to the machine's processor count — both of which describe the login node, which is not in the allocation and is usually several times larger than what the job holds. A job granted 20 CPUs was sized as though it held 128, and --cpus-per-task 128 is a request SLURM refuses outright.

The fix

Two parts, both in pybnf/cluster.py. ADR-0125 records the reasoning.

1. The job's own per-node list is a source of the single count. cpus_per_node now consults $SLURM_JOB_CPUS_PER_NODE — the per-node list SLURM publishes for the job, which is set correctly in a login-node shell — between $SLURM_CPUS_ON_NODE and the two machine-level numbers. Its smallest entry is taken, because this single number both sizes a pool started on every machine and is what one srun step asks for on every machine in it: asking for less than a machine holds costs speed, asking for more than the smallest holds is refused. Inside the allocation nothing changes — $SLURM_CPUS_ON_NODE is still preferred where SLURM sets it. This also fixes the same login-node exposure on the SSH launcher (-t slurm), which reads the same count.

2. The layout hands its counts on rather than having them re-derived (the fix suggested in the issue). srun_worker_layout already reads the per-node list before choosing which command to build; the equal-size path now passes that count — and the phrase naming its source, for the log — into srun_worker_command, which derives one itself only when no caller supplied it. So a $SLURM_CPUS_ON_NODE that does not describe this job, including one exported by hand as the workaround for this bug, no longer sizes a later run.

The srun argument list is otherwise untouched: a launch from inside the allocation builds the command it built before, and the existing regression test that pins that argv still passes.

Verification

The constructed srun argument list is the oracle, as it was for #614 and #617. Eight new tests; all seven that target the defect fail on main and pass here:

• cpus_per_node — the job list is read when the step variable is empty; a mixed list reduces to its smallest machine; the step variable still wins when both are set; an unreadable or zero list falls through to the local numbers.
• srun_worker_command — a caller-supplied count is used instead of the environment (and its source reaches the log); it still caps an oversubscribed parallel_count.
• setup_srun_cluster, end to end in a stood-up login-node environment (no $SLURM_CPUS_ON_NODE, a job list of 20, machine-level numbers of 128) — the default path asks for 20 workers and 20 CPUs rather than 128; an explicit parallel_count = 64 starts all 64 workers but asks for 20 CPUs, so the step is not refused; and the count the layout read is the count the command uses even when $SLURM_CPUS_ON_NODE is stale.
• per_node_cpus — the two fixes compose: a list that cannot be lined up with the machines still supplies the fallback count rather than the login node's size.

Full local gate green (4564 passed) apart from one unrelated pre-existing failure — test_gradient_sens_fallback.py::test_a_declined_model_is_reported_identically_cold_and_warm, which fails on a clean main too against a bngsim built from upstream main and is filed as #644. Docs updated (docs/cluster.rst, docs/config_keys.rst) and a CHANGELOG entry added.

Also noted

ADR-0124 recorded a limitation on the parallel_count path — on a mixed-size allocation its even split can ask for more CPUs than the smallest machine holds — and asked for its own tracking issue rather than being folded into that change. That is now filed as #643, and this PR deliberately leaves it alone: what changed here is only that the single count it is capped by describes the allocation rather than the machine PyBNF is launched from.