workflow-fix: critic-lens-reference item 13 still says fellows-first auto default (stale since #2054/#2059)

Goal

Correct the stale auto-lane-order prose in .claude/rules/critic-lens-reference.md item 13 so a plan critic costs §9 compute rows on the machine the router will ACTUALLY provision, and add the mechanical check that keeps the two surfaces from drifting again.

The drift

.claude/rules/critic-lens-reference.md:410 (Methodology lens item 13, "Compute projection costed on the routed machine"):

the plan's §9 compute table must cost each row's planned_wall_h / basis on the machine the backend router will ACTUALLY provision — under the standing fellows-first auto default (#2028: GCP provisioning disabled) that is the fellows H200 cluster, then the free SLURM lanes, with RunPod's H100 intent table as the terminal rung

That describes the PRE-#2054 ordering. The current canonical order is RunPod FIRST:

• src/explore_persona_space/backends/router.py:895 — DEFAULT_AUTO_LANE_ORDER (built by _default_auto_lane_order()); the module docstring at router.py:45 documents the current order, and router.py:899 notes "runpod is a LEGAL lane as of" the inversion.
• CLAUDE.md § Compute backends and .claude/rules/compute-backends.md: DEFAULT_AUTO_LANE_ORDER = ("runpod", "fellows", "nibi", "fir", "mila") — RunPod first (reason: auto_runpod_first), then fellows, then the free DRAC/Mila SLURM lanes, with a terminal RunPod RETRY rung (reason: auto_fallback_runpod).
• Landed by #2054 (standing user directive: RunPod is the first-resort lane, the shared Anthropic fellows/safety org pool) and #2059 (Invert auto lane order: RunPod primary (GPU+CPU), fellows next). Both are completed; neither updated this lens-reference prose.

grep -n "runpod-first\|runpod, fellows\|DEFAULT_AUTO_LANE_ORDER" .claude/rules/critic-lens-reference.md returns nothing — the file has no correct statement of the order anywhere, so this is a plain stale-prose gap, not an inconsistency between two passages.

Why it matters (the failure it causes)

Item 13 is the lens that makes a critic REVISE a §9 compute table costed on the wrong GPU. As written it instructs the critic to expect fellows H200 as the provisioned machine under a bare auto route. A plan whose §9 rows are correctly costed on RunPod H100 — the lane that will actually win — reads as MIS-costed against this lens, and a plan wrongly costed on H200 reads as correct. It inverts the check for exactly the plans it governs (any task with absent/empty backend: frontmatter, i.e. the default).

Moot for the task that surfaced it (#2162 pins backend: runpod and its v7 §9 basis is H100-measured, so the efficiency lens PASSed on recomputed arithmetic), which is why this is filed rather than fixed inline — the next bare-auto plan review is where it bites.

Fix

1. Rewrite item 13's routed-machine clause to state the current order (RunPod H100 intent table first, then fellows H200, then the free DRAC/Mila SLURM lanes, then the terminal RunPod retry rung), keeping the existing #2028 GCP-disabled parenthetical and the INTENT_TO_MACHINE-applies-only-under-rollback caveat.
2. Audit the SAME drift across the other always-on and lens surfaces in one pass — at minimum .claude/rules/critic-lens-reference.md, .claude/rules/clean-result-critic-lens-reference.md, .claude/agents/critic.md, .claude/agents/planner.md, .claude/agents/efficiency-critic.md, .claude/agents/methodology-baselines-critic.md, .claude/rules/plan-compute-sizing.md, .claude/skills/adversarial-planner*/SKILL.md — since #2054/#2059 updated the backend rules but a lens-side sweep evidently did not run. Report which files were clean.
3. Add the mechanical guard so this cannot drift silently again: a scripts/workflow_lint.py check that FAILs when a workflow-surface file asserts a lane-ordering adjective (fellows-first, gcp-first, fellows first) that contradicts the live DEFAULT_AUTO_LANE_ORDER head from backends/router.py. Read the constant, do not hardcode the expected head — the point is that the next inversion updates one place and the lint catches the stragglers. Bundle it into the no-flags default run per the existing convention, and pin it with a test carrying a fixture of the stale line.

Provenance

Surfaced as a prose follow-up by the efficiency-critic (PLAN MODE, verdict PASS) during the #2162 turn-boundary-multipatch follow-up round's post-approval critic panel, 2026-08-14. The critic flagged it as "workflow-surface note for the orchestrator (prose follow-up, not mine to file)" — correct: subagents never file, the orchestrator routes (CLAUDE.md § Workflow-fix-on-bug protocol, surfaced-prose clause). Claim verified independently by the orchestrator before filing: grep -n "fellows-first" returns exactly critic-lens-reference.md:410, and the file contains no correct statement of the current order.

Dedup: no existing task covers this. Registry title scan for lens-reference / lane order / fellows-first / runpod-first / auto lane returned four hits, all completed and all different work (#1302 mbc capsule pin, #1777 /workspace sentinel WARN, #2059 the inversion itself, #783 GCP→RunPod failover).

workflow_fix_target: .claude/rules/critic-lens-reference.md