daily-fix: GCP intent under-provisions 7B hidden-state capture

kind: infra#wf-fix#wf-fix-fp:6e52a2719981#daily-auto-filed

track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by /daily 2026-06-29 problem sweep. Recurring: #666 (L4 22GB OOM) and #744 (g2-standard-4 OOM) both crashed a 7B forward + hidden-state/Welford capture under the eval intent.

Goal

A 7B-forward-with-activation-capture phase always lands on a GPU with enough HBM; the router never silently under-provisions vs the plan.

Workflow gap

Bug observed: INTENT_TO_MACHINE maps eval/debug -> g2-standard-4 (1x L4). A 7B forward capturing all 28 layers' hidden states needs A100-80; on L4/16GB-RAM it OOMs. #666 and #744 plans requested A100-80 but ran on L4 and crashed mid-run.
Why it is a workflow gap: the intent->machine mapping is the backend router surface; the planner sizing is the planner agent surface.
Confidence (emitter): medium-high; two independent sessions.

Proposed change

Add/route a capture-class intent (or gate eval so a hidden-state-capture 7B forward escalates to an A100-80 machine), OR honor an explicit gpu_type/A100-80 plan request over the intent default.
planner.md: size activation-capture phases (footprint + HBM) at plan time and pick the intent that fits, not the cheap eval default.

Scope / surfaces

src/explore_persona_space/backends/gcp.py (INTENT_TO_MACHINE / selection)
.claude/agents/planner.md (capture-workload sizing)
Grep INTENT_TO_MACHINE / g2-standard-4 callers first.

Constraints / invariants

Workflow surface only. tests/test_gcp_backend.py / tests/test_router*.py pass.
Recursion guard: EPM_WORKFLOW_FIX_SESSION=1.

Provenance

workflow_fix_target: src/explore_persona_space/backends/gcp.py, .claude/agents/planner.md
fingerprint: 6e52a2719981

Sessions: 036936a1 (issue-666), 572d4953 (issue-744).

Activity