EPS Dashboard

Overview / Motivation

Auto-filed by /daily 2026-07-30 Step C (parked workflow-fix-candidate routing) from a candidate parked on task #1896 (emitting source: orchestrator observation during #1896 clarify/plan, seconded by the Methodology critic; parked 2026-07-30T21:15:22Z). #1896 itself added capture-7b to the SLURM tables only, per its fingerprint scope (merged 20c9d904f47c).

Goal

Close the per-intent fellows-first bypass: add the remaining GCP-mapped GPU intents (lora-7b-h100 at minimum) to the SLURM intent tables, and add a completeness test tying gcp.INTENT_TO_MACHINE's gpu_count>0 keys to the SLURM tables so the class cannot silently reopen.

Workflow gap

Bug observed: lora-7b-h100 (a GCP-mapped 1x H100-80 GPU intent) is absent from the SLURM intent tables exactly like capture-7b was (#1896 fixed capture-7b only) — a lora-7b-h100 auto-lane dispatch still ValueErrors off the free fellows/SLURM lane and burns paid GCP; sweep-8g-a100 / sweep-8g-h100 / eval-h100 are in the same absent class (whether SLURM should serve 8-GPU sweeps needs a design call — flag architectural-adjacent design question in the plan, do not silently decide).
Why it is a workflow gap: the standing fellows-first policy (#1609) is silently bypassed per-intent whenever a GCP GPU intent lacks a SLURM row; there is no completeness test tying gcp.INTENT_TO_MACHINE's gpu_count>0 keys to the SLURM tables (the RunPod rung has exactly this via RUNPOD_INTENT_FOR_GCP_INTENT + its translation-map test).
Confidence (emitter): medium
verified-at-filing: grep -n 'lora-7b-h100' src/explore_persona_space/backends/slurm.py src/explore_persona_space/backends/gcp.py src/explore_persona_space/backends/router.py → gcp.py:393 (INTENT_TO_MACHINE row) + router.py:460 (RunPod translation) present, slurm.py 0 hits (absence in the SLURM tables confirmed); slurm.py's tables carry lora-7b/eval/capture-7b (:530-:603) but no H100-family intents (2026-07-31 filing time). #1896's capture-7b fix merged 2026-07-30T22:46:27Z.

Proposed change (candidate diff sketch — refine in planning)

Add lora-7b-h100 (and any other 1-GPU H100-class intents the plan judges SLURM-servable) to _DEFAULT_TIME_BUDGETS_H + the GPU-count table in backends/slurm.py; add a completeness pin test asserting every gcp.INTENT_TO_MACHINE gpu_count>0 key either has a SLURM row or is on an explicit documented-exclusion list (8-GPU sweep intents may land there pending the design call).

Scope / surfaces

Primary target: src/explore_persona_space/backends/slurm.py, src/explore_persona_space/backends/router.py
tests: tests/test_slurm_*.py / tests/test_router.py completeness pin.

Constraints / invariants

The in-scope src/ backends package is workflow surface (CLAUDE.md § workflow-fix scope).
The 8-GPU-sweep design question (should the fellows lane serve them at its 16-GPU/user caps?) is surfaced in the plan, not silently decided; a documented-exclusion list is the acceptable v1 answer.
Existing router/SLURM tests stay green; tests/test_router.py::test_default_auto_lane_order_is_gcp_first durability pin untouched.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

workflow_fix_target: src/explore_persona_space/backends/slurm.py, src/explore_persona_space/backends/router.py
fingerprint: e67598577033 (tag-authoritative; supersedes body-carried fingerprint: 416a551b1ba7)
origin: parked candidate-block on #1896 events.jsonl, ts 2026-07-30T21:15:22Z (routed by /daily 2026-07-30 Step C)