daily-fix: fail loud when --min-ram-gb is inert on a GPU dis

Overview / Motivation

Auto-filed by /daily 2026-08-01 (route 2: behavior/logic change → independent review) from consolidated problem sweep entry C4 (miner 1; session 55419495, #1739, boxes oodhallmax / oodpredshallu).

Goal

Make --min-ram-gb fail loud (or at minimum WARN + refuse silently-undersized provisioning) on GCP GPU dispatches whose resolved machine's RAM is below the requested value — today the flag is a RunPod-CPU-fallback-only knob and GCP GPU machine selection silently ignores it.

Workflow gap

• Bug observed: the #1739 Gap-2 recipe claimed "--min-ram-gb 300 routes Gap 2 to the 340GB a2-ultragpu-2g (the OOM fix)"; the box landed on a2-ultragpu-1g (170GB), the mismatch was rationalized ("170GB is ample here"), and BOTH boxes OOM'd exit 137 ("My earlier 'scoped run won't OOM' call was wrong"). Relaunch with --gpus 2 → 340GB succeeded. The flag's inertness is documented but the dispatch surface gives no runtime signal.
• Why it is a workflow gap: a plan-stated RAM requirement threaded through the canonical dispatch CLI silently provisions an undersized GPU box — exactly the failure class the CPU-side gate (#1010 cpu_fallback_infeasible_for_plan) was built to refuse.
• Confidence: high (premise probed by the miner AND re-verified now).
• verified-at-filing: grep -n 'min_ram' scripts/dispatch_issue.py src/explore_persona_space/backends/{gcp,router}.py → dispatch_issue.py:1471-1478 comment reads verbatim "RunPod-CPU-fallback knob (#1010): read by the feasibility gate in router._runpod_terminal_rung ... GCP machine selection is unchanged (by intent); inert on SLURM lanes" (same text in the --min-ram-gb argparse help at :2713); gcp.py's only min_ram_gb uses are handle/marker-extra THREADING at :4533 and :5798 (reconnect + failover-handle persistence — no machine-selection read); router.py reads it only in _refuse_infeasible_cpu_footprint (:3149, CPU instances). git log --oneline --since='7 days ago' -- scripts/dispatch_issue.py src/explore_persona_space/backends/gcp.py → 9 commits, none touches min_ram_gb GPU semantics (2026-08-01).

Proposed change (refine in planning)

In scripts/dispatch_issue.py (parse-time) or backends/gcp.py (rung-selection time): when --min-ram-gb is set on a GPU dispatch (resolved gpu_count > 0), compare against the resolved GCP machine's RAM (add a RAM column to INTENT_TO_MACHINE / the wide-rung map, or a static machine→RAM table) and:

• default: FAIL LOUD pre-launch naming the mismatch + the satisfying alternatives (--gpus N wide rung, a bigger intent), mirroring --min-gpu-mem-gb's A100-40 rung-skip (#1468) — prefer rung-SKIP semantics (walk to a rung that satisfies the RAM) over hard refusal where the ladder has a satisfying rung;
• at minimum: a loud WARN at dispatch + the requested-vs-resolved RAM on the epm:backend-selected marker. Planning decides skip-vs-refuse per rung; keep RunPod-CPU (#1010) and SLURM behavior byte-unchanged.

Scope / surfaces

• Primary target: scripts/dispatch_issue.py, src/explore_persona_space/backends/gcp.py
• Likely also src/explore_persona_space/backends/router.py (rung walk) + tests/test_gcp_backend.py / tests/test_router.py pins; update the argparse help + the :1471-1478 comment ("by intent" becomes stale).

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff/bash -n on touched files passes.
• Preserve #934 omit-when-absent spec-hash discipline for spec.extra; no behavior change when the flag is unset.
• Recursion guard: this task's session carries the workflow_fix_target Provenance line and MUST NOT auto-route its own subagents' workflow-fix candidates.

Provenance

• fingerprint: d3668290b74d
• workflow_fix_target: scripts/dispatch_issue.py, src/explore_persona_space/backends/gcp.py
• origin: /daily 2026-08-01 problem sweep, CONSOLIDATED.md entry C4.