workflow-fix: verify_plan c26/c27 routed-machine mirror is GCP-era, inverts under runpod-first auto default

Goal

Rebuild verify_plan.py's routed-machine mirror (_C26_INTENT_GPU, and the c27 sets derived from it) so c26/c27 judge a plan's §9 basis GPU against the machine the live router will ACTUALLY provision under the runpod-first auto default (#2054/#2059), instead of the GCP-era A100/L4 families. Then remove the interim staleness caveat the prose surface now carries.

The drift

scripts/verify_plan.py:4826 — _C26_INTENT_GPU maps intents to GPU families:

"lora-7b": "A100", "lora": "A100", "capture-7b": "A100", "ft-7b": "A100",
"eval": "L4", "debug": "L4", ...

Those are the GCP INTENT_TO_MACHINE families (a2-ultragpu A100-80, L4) — correct when GCP led the auto chain, wrong now. GCP provisioning is DISABLED (GCP_PROVISIONING_DISABLED = True, #2028) and the live head is RunPod:

• DEFAULT_AUTO_LANE_ORDER == ('runpod', 'fellows', 'nibi', 'fir', 'mila') (src/explore_persona_space/backends/router.py, #2054/#2059) — verified live.
• RunPod's intent table (.claude/rules/pods.md): lora-7b → 1× H100, eval/debug → 1× H100, ft-7b → 4× H100.

So for a bare-auto plan, c26 expects an A100/L4 basis while the router provisions H100. A plan correctly costed on RunPod H100 trips a c26 WARN; a plan wrongly costed on A100 reads as clean. The check is inverted, the same way critic-lens-reference.md item 13's prose was before #2298 fixed it.

Downstream consumer — not a one-dict edit. verify_plan.py:5113-5116 derives _C27_L4_INTENTS and the non-L4/non-CPU set FROM _C26_INTENT_GPU, so c27 inherits the same GCP-era premise. Scope both.

Why it matters

c26 is WARN-only, which makes this worse, not better: it emits an EXPECTED-FALSE warning on every runpod-costed bare-auto plan, and the prose surface now tells readers to ignore it (see the caveat below). A gate whose documented guidance is "this warning is expected" is a gate reviewers stop reading — and the day a plan is genuinely mis-costed, the real WARN is indistinguishable from the noise.

Fix shape (as specified by #2298's implementer when it raised the concern)

1. Derive the routed GPU family from the live lane head rather than hardcoding a provider's table — the same AST-read-the-source approach workflow_lint.py --check-lane-order-adjective already uses to stay in sync (_default_auto_lane_order), so the mirror cannot silently invert on the next lane-order change. Keep it hermetic (no network, no live API call at lint time).
2. Re-calibrate the corpus: c26/c27 are heuristics tuned against the existing plan corpus, so re-scan after the mapping changes and re-baseline expected findings.
3. Update tests/test_verify_plan.py c26/c27 coverage (and any c26-derived assertions in tests/test_issue2476_gates.py).
4. Remove the interim caveat at .claude/rules/plan-compute-sizing.md:804-806 ("c26's routed-machine mirror is GCP-era and tracked separately — a c26 WARN on a runpod-costed bare-auto plan is expected until that mirror is updated"). Once the mirror is rebuilt that sentence becomes stale prose on the prescriptive surface — i.e. it becomes the exact defect class #2298/#2331 existed to close. Landing the code fix without deleting the caveat just relocates the drift.

Provenance

Raised as concern c26-gcp-era-mirror-stale (severity CONCERN, round 1, by the implementer) on #2298, and deliberately left out of that diff per its plan's Part 5: #2298 was itself a workflow-fix task carrying workflow_fix_target:, so its recursion guard required the companion be NAMED, never auto-filed. It has sat as an open concern on a completed task since 2026-08-22 with no execution path — concerns on terminal tasks are never dispatched.

Surfaced again 2026-08-24 while verifying #2331 (an independent duplicate filing of the same prose drift, archived after #2298 was confirmed to cover its whole scope). #2331 was not mechanically workflow-fix-bound (is_workflow_fix_session(2331) is False — its candidate block is an HTML comment, not a workflow_fix_target: Provenance line), so filing this companion here is not a recursion-guard bypass.