infra: default BLAS/torch thread caps for shared-VM python launches

Overview / Motivation

Auto-filed from the 2026-07-02 shared-VM overload incident (load 186-226 on 32 cores). Read-only diagnosis: /tmp/vm_load_diagnosis.md (copy into artifacts/ at planning).

Goal

Cap BLAS/torch intra-op threads for python jobs launched on the SHARED VM so N concurrent jobs cannot oversubscribe the box ~6x (each uncapped torch/numpy job spawns nproc=32 intra-op threads; 5-6 concurrent jobs -> run queue ~160-190 -> load 186-226; PSI cpu some avg10=96.5%).

Workflow gap

• Bug observed: repo-wide, NO VM launch surface sets OMP_NUM_THREADS/MKL_NUM_THREADS/OPENBLAS_NUM_THREADS or torch.set_num_threads; /proc/environ of all 6 heavy PIDs during the incident had zero caps. Each job realizes only ~5-6 cores of throughput while holding 32 runnable threads.
• Why it is a workflow gap: the shared VM hosts every session's CPU work; uncapped defaults make fleet collisions pathological (64-min layers for a 60s job; earlyoom SIGTERM sweeps at <10% mem-avail killed 4-7GB workers silently).
• Confidence (emitter): high

Proposed change (refine in planning)

• src/explore_persona_space/orchestrate/env.py load_dotenv(): os.environ.setdefault OMP/MKL/OPENBLAS_NUM_THREADS to a VM default (e.g. 8) ONLY when running on the shared VM (not RunPod /workspace, not GCE, not SLURM — pods/instances are dedicated and WANT full width). Launch-time explicit values always win (setdefault).
• Add the cap to the VM-side launch guidance in .claude/rules/code-style.md (compute-throughput bullet) so ad-hoc / smoke / free-analysis launches inherit it.
• Consider torch.set_num_threads equivalent where torch is imported before env (document the ordering constraint).

Scope / surfaces

• Primary: src/explore_persona_space/orchestrate/env.py (library; in the #747 spirit), .claude/rules/code-style.md.
• Verify no pod/GCE/SLURM regression: those lanes must keep full threads (grep bootstrap_pod.sh, backends/gcp.py, backends/slurm.py).

Constraints / invariants

• setdefault-only (explicit env always wins). Detection must fail OPEN (no cap) off-VM.
• ruff passes; tests for the detection predicate.

Provenance

• origin: PM-chat VM triage 2026-07-02, load-186 diagnosis (/tmp/vm_load_diagnosis.md)