workflow-fix: verify_plan heuristic — GPU basis vs routed machine

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #1073 (emitting agent: critic, Methodology lens — prose follow-up).

Goal

Add a verify_plan.py heuristic check that flags any §9 compute-table basis cell naming a GPU (e.g. H100) different from the machine the Spec line / intent routes under auto (GCP INTENT_TO_MACHINE), when the row contains no explicit scaling token (×, "scaled").

Workflow gap

• Bug observed: plan §9 wall-time bases were H100-measured while the auto lane routes A100-80 with no stated cross-GPU scaling; third recurrence of the class (#599, #833, #1073).
• Why it is a workflow gap: .claude/rules/plan-compute-sizing.md § "Cost wall-time against the machine the router will ACTUALLY provision" is unconditional prose, but verify_plan.py has no mechanical check for it — the class keeps sailing through the mechanical pre-pass and is caught only by the (fallible) LM critics.
• Confidence (emitter): low

Proposed change (candidate diff sketch — refine in planning)

• In scripts/verify_plan.py, add check c26_gpu_basis_vs_routed_machine:
• • detect the routed machine: --intent token in the plan (map via backends/gcp.py INTENT_TO_MACHINE import or a static mirror) when backend is auto/absent
• • for each §9 table row, if the basis cell names a GPU model string (H100/A100/L4/H200) differing from the routed machine's GPU

• AND the row lacks a scaling token (`×`, 'scaled', 'per-step rate'), emit FAIL (or WARN initially)
  

• • N/A escape: 'N/A — basis measured on the routed machine' or an explicit backend pin matching the named GPU

Scope / surfaces

• Primary target: scripts/verify_plan.py
• Grep the workflow surface for the pattern before editing (grep -rln 'INTENT_TO_MACHINE\|routed machine' .claude/ scripts/) and update every hit; list them in the plan.

Constraints / invariants

• Workflow-surface only — never experiment code, configs/, or tasks/.
• scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; keep consistent with .claude/rules/plan-compute-sizing.md.
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).

Provenance

• workflow_fix_target: scripts/verify_plan.py
• fingerprint: 2559c56f77b7

Surfaced prose (verbatim, from the Methodology critic's report on #1073): "scripts/verify_plan.py: a heuristic check flagging any §9 compute-table basis cell that names a GPU (e.g. H100) differing from the machine the Spec line / intent routes under auto (GCP INTENT_TO_MACHINE), when the row contains no explicit scaling token (×, "scaled"). This is the third recurrence of the class (#599, #833, this plan); the check is grep-shaped once the intent→machine map is imported."