Publish GB300 EKS evidence from a v1.0.0-rc1 run

[mchmarny]

Summary

Publish GB300 EKS evidence from a v1.0.0-rc1 run, so the v1 tag ships with
evidence against the recipes v1 actually contains.

Split out of #2806. That issue asked for evidence at a released version and
is satisfied by the 0.21.1 run published in #2812 — "any release counts" was
the intended reading. This issue is the separate thing #2806's sequencing was
carrying implicitly: a run on the release candidate itself.

Why a released-version run is not enough

#2772 (GB300 node tuning, nodewright package bumps) merged 2026-09-18, nine
days after v0.21.1 shipped. The committed 0.21.1 pointers therefore certify
the untuned GB300 recipe. Every GB300 recipe change since is uncovered.

ROADMAP §2
requires GB300 training/Kubeflow and inference/Dynamo to "remain green against
the shipped recipes". At the v1.0.0 tag the shipped recipe is the tuned one,
and nothing attests it.

This was called out on #2806 at the time: "Note v0.21.1 predates #2772 (GB300
node tuning), so these pointers certify the untuned recipe. Per the sequencing
in #2807 the definitive runs are on v1.0.0-rc1." Closing #2806 on the
released-version reading is correct and leaves that observation untracked. This
is its tracker.

Scope

• Run both coordinates against v1.0.0-rc1 and publish signed evidence
  • eks/gb300-ubuntu/training-kubeflow
  • eks/gb300-ubuntu/inference-dynamo
• Confirm both Evidence links resolve against live board data
• aicr evidence digest matches each pointer's predicate.recipe.digest
• Publish without AICR_VALIDATOR_IMAGE_TAG set, so the attestation
  records an immutable validator tag (Validator image tag override silently replaces immutable provenance in evidence #2873)

Hardware

aws-us-east-2-nhensley-gb300, in the Sep 28-30 window. Per #2806: both GPU
nodes needed package repair (unattended-upgrades installed a 7.0.0 kernel EFA
3.0.0's DKMS cannot build against); the repair purged the linux-aws
metapackages, so the nodes no longer auto-advance kernels. Stable for a re-run;
the kernel-update policy remains an open question for the cluster owner.

If the window is missed

v1 ships with GB300 evidence at 0.21.1 against the untuned recipe. That is a
ROADMAP §2 carve-out, not a silent gap — amend the §2 acceptance text to say
so rather than leaving the criterion asserting something untrue.

Related

• Publish GB300 evidence at a released version and register both coordinates #2806 — the released-version half, closed by fix(testgrid): register GB300 and GB200 EKS coordinates for evidence linking #2872
• v1.0.0 release execution: correct the v0.23 binding, freeze the baselines, cut the tag #2807 — release execution; its sequencing references this run
• feat(gb300): enable node tuning and bump nodewright packages #2772 — the GB300 tuning change that made 0.21.1 evidence stale