daily-fix: default BLAS/OMP thread pins for vectorized VM fi

Overview / Motivation

Auto-filed by /daily 2026-07-05 (route 2: behavior/logic change -> independent review), from the nightly transcript problem sweep.

Goal

(a) set conservative default thread caps (OMP_NUM_THREADS/MKL_NUM_THREADS/torch.set_num_threads) inside the canonical vectorized fit helper so callers cannot forget them; (b) add to the vectorize-many-cell-fits.md detached-launch/Supersede contract: a RELAUNCH must carry the identical env pins (thread caps, choom) as the reviewed original launch command - verify pins in the supervisor before dispatch.

Workflow gap

• Bug observed: #811's vectorized ridge-fit grid ran ~12x slower than its smoke test (first unit ~62 min vs ~5 min projected) from BLAS thread over-subscription on the shared VM; it took two diagnose/relaunch rounds to pin threads and restore ~5 min/unit (~1-3h lost). Later the same day the RELAUNCH supervisor omitted the OMP/MKL pins again: ~55 min at 0/108 checkpoints before the orchestrator killed it and patched i811_fits_supervisor.sh with OMP_NUM_THREADS=8 MKL_NUM_THREADS=8 (~1h more lost). Sessions f2ef929f + 86923fc5, task #811.
• Why it is a workflow gap: the failure originates in the workflow surface / shared helper named below, not in any one experiment.
• Confidence (emitter): medium

Scope / surfaces

• Primary target: src/explore_persona_space/analysis/vectorized_mlp_skill.py, .claude/rules/vectorize-many-cell-fits.md
• Same-incident pair; the helper-side default is the durable half.

Constraints / invariants

• scripts/workflow_lint.py --check-asks + --check-references stay green; ruff clean on touched files; relevant tests pass.

Provenance

• workflow_fix_target: src/explore_persona_space/analysis/vectorized_mlp_skill.py, .claude/rules/vectorize-many-cell-fits.md
• source: /daily 2026-07-05 problem sweep (transcript-mined)