[88plug]

H62 asks H50's question again, this time against Intel's new default for the agent toolkit. That default is EAT 2.0 installed through EAS, as EAT release-2.0.0's README and docs/single-node-deployment.md describe it.

This PR adds the pre-registration and its runner. Nothing has run on crescent: it was only read, over read-only ssh, to check the plan is feasible. No guest exists there yet.

What is pre-registered (EXPERIMENTS.md, row H62 and its section)

• Arms.
  • A1 is H50's A1, unchanged: agent manifest, OVMS 2026.4.1 int4 on the snapped QAT IR, CPU, no draft.
  • B5 is Intel's documented single-node install: EAS v1.2.0-rc.1 (5c27fb3) with Kubespray onto localhost. It installs EAT release-2.0.0 (036b662) and enterprise-inference release-1.1 (170eb3a), and serves the catalog entry gemma-4-26b (vLLM CPU v0.28.0 on KServe, NRI balloons) on the QAT weights.
• Why both arms run in KVM guests.
  • Kubespray v2.30.0 on crescent's host would run its "Uninstall docker" reset. That removes every container, the serving pairs included.
  • EAS's configure refuses Ubuntu 26.04.
  • A1 gets a guest of the same shape, so neither arm gets bare metal. One shape does not show the guest costs vLLM CPU and OVMS alike, so the VM tax is described on both sides, never gated: A1 in its guest against H50's A1 on the host, and B5 in its guest against H50's B1 on the host (same vLLM image and weights, another deployment).
• Named deviations from stock.
  • D1: the layer manifest's two branch revs are pinned to the commits that Intel's own comments on those lines name. These are the rc.1 release notes' component table.
  • D2: gemma-4-26b's model_id is the QAT checkpoint, the same deviation as amendment M1.
  • D3: the guest itself.
  • D4: the harness reaches B5's workload Service, not EAS's LiteLLM gateway with the content-safety guardrail and the hybrid router that Intel's agents use. It leans toward B5.
• Design, statistic, gates, scorer and verdict rule. All are H50's: rule L3, seed H62-crescent-2026-10-29, and agentbench pairs-report --hyp H62 --arm-a A1 --arm-b B5.
• Before hold 1.
  • Setup gates: memory, disk, AMX on every vCPU in each guest, the NFD label, and B5's weights equal to the QAT checkpoint.
  • Install ceilings: 45 min for VM a, 120 min for VM b.
  • A wiring smoke: EAS's own validate agent-toolkit, plus eat-check.sh in EAS mode through EAS's gateway.
• Holds.
  • Holds run until 24 pairs or 8 holds, at most 40 min each.
  • Each hold restarts both serving pods first.
  • Two holds in a row that are not deliverable stop the row.
• Predictions. REFUTED 0.6, INCONCLUSIVE 0.35, SUPPORTED 0.05.
• Re-run trigger. It fires when an EAS release at v1.2.0 or later, or an EAT release at 2.0.0 or later, is tagged (bench/h62/upstream.sh); release candidates do not fire it.
  • Before hold 1, a dated amendment re-pins B5 to the tagged release.
  • After hold 1, H62 keeps its verdict and a new row, H62r, is pre-registered.
• A-EAS (k3d-gpu eas up --agent-toolkit --eas-ref v1.2.0-rc.1, from eas: k3d-gpu eas up, down, status and check (EAS with EAT 2.0 on a k3d-gpu cluster) #260) is described as a later F row and is not part of this one.

Runner

• bench/h62/:
  • h62.py: the pins, the guests' QEMU command lines and cloud-init, the gates, D1, the documented edit and D2, the pod picks, the weight-check script, the re-run trigger and the VM tax.
  • qemu.Dockerfile: QEMU from Ubuntu 26.04, pinned, so the host gets no package.
  • vm.sh, guest-a.sh, guest-b.sh, setup.sh, smoke.sh, arm.sh (with arm-A1 and arm-B5), pairs-hold.sh, campaign.sh, teardown.sh and upstream.sh.
• bench/eatflow/live.sh: ARM=H62 in pairs mode drives the two remote arms. The local arms run unchanged in the other branch, and from the LiteLLM gateways on, the hold is H50's code.
• bench/eatflow/pairs.py: DESIGNS["H62"] and its seed. A pod cgroup can now be an exec: command, so the idle-arm check reads each pod inside its guest.
• bench/eatflow/crescent/serving.py: SERVING_VMS records the guests' CPU as vm_cpu, outside the serving covariate.
• docs/intel/technical-report.md now counts 82 rows, 20 of them PENDING.

Depends on

The wiring smoke uses integrations/eat/eat-check.sh's EAS mode from #260. The deploy commit must include #260. If #260 has not merged when crescent is ready, a run branch merges #260's head, as amendment E3 did.

Checks run

• bench/h62/test_h62.py: 53 tests, against stubs, with no VM.
• bench/eatflow/test_pairs.py: 66 tests, including H62Test.
• bench/eatflow/crescent/test_crescent.py: 48 tests, including ServingVmTest.
• bench/eatflow/test_agentbench.py, test_solo.py, test_h4345.py and crescent/test_holds.py.
• tests/ledger_drift_check.py (82 rows), tests/claims_check.sh, tests/claims_check_test.sh, tests/bench_test.sh, tests/ci_wiring_test.sh, tests/image_pins_test.sh, scripts/images-lock pins and tests/docs_test.sh.
• shellcheck v0.11.0 at warning severity, on every new script and on live.sh.

Nothing here has run on crescent or in a guest. AMX passthrough, the Kubespray install behind user-mode networking, and the VM tax are named in the pre-registration as risks the setup gates check.