EPS Dashboard

Overview / Motivation

Auto-filed by /daily 2026-07-16 from transcript mining (#1315, ~05:28Z).

Goal

Stop the 40GB rung from admitting vLLM workloads that cannot initialize on realistic free device memory.

Workflow gap

Bug observed: the GCP ladder's spot A100-40 rung admitted a vLLM workload that crashed at engine init — 'Free memory on device (22.52/39.49 GiB) on startup is less than desired GPU memory utilization (0.6, 23.7 GiB)' (#1315 ~05:28Z) — ~17 GB was already consumed on the fresh instance and the fits-in-40GB eligibility neither checks free (vs total) memory nor threads the workload's declared gpu_memory_utilization; a full boot cycle was lost before the flex-start A100-80 retry ran clean
Why it is a workflow gap: Each false admission burns a boot cycle + ladder walk; the eligibility check exists precisely to avoid this.
Confidence (emitter): medium (single incident; the ~17GB-consumed cause on a fresh instance deserves its own look)
verified-at-filing: incident text in #1315 transcript (engine-init failure + clean 80GB retry); grep -n 'fits-in-40' src/explore_persona_space/backends/gcp.py CLAUDE.md locates the eligibility contract (no free-memory / mem-fraction term — absence claim, planning verifies exact site)

Proposed change (candidate diff sketch — refine in planning)

gate the A100-40 rung on the workload's declared vLLM memory fraction vs realistic FREE memory (or auto-lower gpu_memory_utilization on the 40GB rung, or pre-probe free memory before engine start) — planner picks

Scope / surfaces

Primary target: src/explore_persona_space/backends/gcp.py
Grep the workflow surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

scripts/workflow_lint.py --check-asks passes; ruff on touched files passes.
This session runs under a workflow_fix_target: Provenance line — it MUST NOT auto-route its own subagents' workflow-fix candidates (recursion guard).

Provenance

fingerprint: d256d6e07608