workflow-fix: RunPod terminal rung maps GCP-only intents to nearest RunPod intent

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from an orchestrator observation on task #841 (round-2 dispatch, 2026-07-03 16:58Z).

Goal

Make the router's RunPod terminal rung map GCP-only intents (e.g. capture-7b) to the nearest RunPod-provisionable intent instead of crashing, so the sanctioned last-rung fallback actually fires when the GCP ladder is exhausted.

Workflow gap

• Bug observed: #841's round-2 launch exhausted GCP (A100 stockout both zones + per-day attempt cap 8) and SLURM (branch gap #653), then the RunPod terminal fallback itself FAILED: pod_lifecycle.py provision --issue 841 --intent capture-7b exited 1 — capture-7b is a GCP-only intent with no RunPod GPU mapping, so the auto-chain ended in NoComputeAvailableError even though RunPod had capacity for an equivalent 1-GPU pod.
• Why it is a workflow gap: the router's documented contract is that RunPod is the terminal rung of the capacity path; an intent-vocabulary mismatch between lanes silently voids that guarantee for every GCP-only intent (capture-7b, and any future GCP-specific machine-type intents).
• Confidence (emitter): high (live epm:backend-selected marker on #841 carries the full attempts evidence).

Proposed change (candidate diff sketch — refine in planning)

• In backends/router.py _runpod_terminal_rung (and the async failover path), translate the requested intent through a small GCP-intent→RunPod-intent map before calling pod_lifecycle.py provision (e.g. capture-7b → eval (1×H100) or the closest single-A100/H100 intent), logging the translation on the epm:backend-selected marker extra.
• Unknown intents with no mapping keep the current fail-loud behavior, but with an error naming the missing map entry.
• Add a router test pinning: GCP-only intent + exhausted ladder → RunPod rung fires with the translated intent.

Scope / surfaces

• Primary target: src/explore_persona_space/backends/router.py
• Also check: src/explore_persona_space/backends/runpod.py, scripts/pod_lifecycle.py (intent table), tests/test_router.py.
• Grep: grep -rn "capture-7b\|INTENT_TO_MACHINE\|RUNPOD.*INTENT" src/explore_persona_space/backends/ scripts/pod_lifecycle.py scripts/gpu_heuristics.py

Constraints / invariants

• RunPod stays the LAST rung only (tests/test_router.py invariants unchanged).
• No dollar caps; the translation must not widen GPU count (map to the same-or-narrower width).
• 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/router.py
• fingerprint: 9a65fe461ab5

Surfaced observation (verbatim, #841 events.jsonl epm:backend-selected 2026-07-03T16:58:33Z): "runpod terminal fallback FAILED (CalledProcessError: ... pod_lifecycle.py provision --issue 841 --intent capture-7b returned non-zero exit status 1.)" following "rung ondemand_a100_80: per-day GCP attempt cap 8 reached" and both-zone ZONE_RESOURCE_POOL_EXHAUSTED.