k8s: NFD v0.20.1, agent-sandbox v1.0.8, EAS patch 0003 on nri-plugins v0.15.0, kubectl v1.37.2 for the wildct and cross-host rows

.github/workflows/ci.yml

@@ -252,6 +252,7 @@ jobs:
       - run: python3 bench/eatflow/test_agentbench.py
       - run: python3 bench/eatflow/test_pairs.py
       - run: python3 bench/eatflow/test_solo.py
+      - run: tests/xhost_tools_test.sh
       - run: python3 bench/eatflow/test_h4345.py
       - run: python3 bench/models/test_export_qat_ov.py
       - run: python3 bench/models/test_qat_grid.py

EXPERIMENTS.md

@@ -1735,7 +1735,7 @@ entries that ask for another, Xeon/AMX, XMX, non-throttling or quiet host, and t
    below showed the cause: without C1 the agent pod stayed Pending. C1 changes when the kubelet
    evicts, nothing else, and applies to both arms alike. X1's 97% guard stays.
 5. **Pins.** k3d v5.9.0 and kubectl v1.35.2 in `~/k3dgpu-xhost/bin`, each checked against its
-   upstream checksum; node image `cryptoandcoffee/k3d-gpu:0.1.61` (the launcher's default at this pre-registration's base, set explicitly with `K3D_GPU_IMAGE` or `IMAGE` in every command,
+   upstream checksum (kubectl v1.37.2 for the rows [amendment K1](#cross-host-amendment-k1-kubectl-v1372-in-the-cross-host-workspace-2026-10-29) moves); node image `cryptoandcoffee/k3d-gpu:0.1.61` (the launcher's default at this pre-registration's base, set explicitly with `K3D_GPU_IMAGE` or `IMAGE` in every command,
    k3s v1.37.1; a change: the originals ran 0.1.28 (H25 run 4), 0.1.55 (H34) and 0.1.59 (H28),
    with the agent-manifest code unchanged since H25 run 4); OVMS `openvino/model_server@sha256:a928903f...76fc58` (2026.4.1, as every
    original row); `load.py`'s stress-ng pin. Models are downloaded on the host at the original
@@ -1801,6 +1801,36 @@ failure C1 fixes. Smoke 2 (22:14-22:20Z, with C1): `up` 48 s, the agent pod Read
 C1 as `--k3s-arg ...@server:*`) had no taint after 90 s. No inference request was sent in either.
 Smoke records: `bench/xhost/results/2026-10-20/smoke{1,2}/`.
 
+### Cross-host amendment K1: kubectl v1.37.2 in the cross-host workspace (2026-10-29)
+
+Recorded 2026-10-29, before any H7x hold and before H28x's run 2. Mechanics only. No hypothesis,
+arm, model, node image, prediction, outcome table, scorer, threshold, cap or order changes. It
+is pre-registered by its push time.
+
+1. **Why.** Rule 5 pins kubectl v1.35.2 in `~/k3dgpu-xhost/bin`, and `bench/xhost/hold.sh` puts
+   that directory first on `PATH`. The node image these rows run (0.1.61) serves Kubernetes
+   v1.37.1 (k3s v1.37.1+k3s1). Kubernetes' version skew policy supports kubectl within one minor
+   version of kube-apiserver, so v1.35 against v1.37 is outside it. No hold has failed from this.
+   Read on 2026-10-29: kubectl v1.37.2 (2026-10-28, `dl.k8s.io/release/stable.txt`) is the newest
+   release; k3d v5.9.0 is still its newest and stays. Crescent amendment E4 (PR #238) makes the
+   same move for shared rule 5's workspace, `~/k3dgpu-bench/bin`; this amendment covers the
+   cross-host workspace, which E4 does not touch.
+2. **Pins.** k3d v5.9.0 (unchanged) and kubectl v1.37.2 (linux/amd64 sha256 `c0448752...`).
+   `bench/xhost/tools.sh`, new, installs exactly these two into `~/k3dgpu-xhost/bin`, each checked
+   against its upstream checksum file (k3d's `checksums.txt`, kubectl's `.sha256`), and prints
+   both versions. It touches nothing else. Rule 5's tools had no script in this repo until now.
+3. **When.** `tools.sh` runs once on crescent from a clean checkout at or after this amendment's
+   merge, with the bench lock free and no cross-host row between holds. Its output goes into the
+   next cross-host row's records.
+4. **Rows.** A row runs all its holds on the kubectl in place at its first measuring hold. Rows
+   whose first measuring hold starts after `tools.sh` has run use v1.37.2: H28x's run 2, if PR
+   #208's amendment R1 allows it, and H7x (PRs #196 and #235), if its pre-registration merges
+   under these rules. H25x, H27x and H34x have run and keep what they ran; H25x and H34x record
+   kubectl v1.35.2. A smoke, or a hold that measured nothing (H28x run 1), does not fix a row's
+   tools.
+5. **Not changed.** The node image, the models, OVMS, `load.py`'s stress-ng pin and every other
+   item of rule 5.
+
 ### H25x run (Xeon 6741P, 2026-10-21)
 
 ```text
@@ -4181,7 +4211,8 @@ re-graded here.
    whose need never fits before the campaign ends is INCONCLUSIVE (not run), with the
    readings recorded. Everything this campaign creates on wildct is deleted at the end
    except the run records, which are copied back into the repo first.
-7. **Pins.** k3d v5.9.0, helm v4.3.1, kubectl v1.37.1 (each the newest release on 2026-10-20)
+7. **Pins.** k3d v5.9.0, helm v4.3.1, kubectl v1.37.1 (each the newest release on 2026-10-20;
+   kubectl v1.37.2 from [amendment P1](#wildct-amendment-p1-kubectl-v1372-2026-10-29))
    in `~/k3dgpu-bench-wildct/bin`, checked against their upstream checksums
    (`bench/wildct/install.sh`); k3d-gpu: the newest release at the first hold (the first one
    carrying PR #180: 0.1.65 or later), its `.deb` checked against the release's `SHA256SUMS` and unpacked in
@@ -4282,6 +4313,24 @@ request, discarded and disclosed). Then H52w (1 hold), H18w (1), H25w (1), H34w
 repeat a failed hold once. At most 12 holds in all. A protocol change is a dated amendment
 merged before the data it governs.
 
+### Wildct amendment P1: kubectl v1.37.2 (2026-10-29)
+
+Recorded 2026-10-29, before any wildct hold (none has run; `smoke1` was discarded). Mechanics only.
+No row's hypothesis, prediction, arm, model, statistic, threshold, cap, disk need, hold count or
+order changes.
+
+Rule W7 pins kubectl v1.37.1, the newest release on 2026-10-20. Kubernetes v1.37.2 came out on
+2026-10-28 and is the newest now (`dl.k8s.io/release/stable.txt`, read 2026-10-29); k3d v5.9.0
+and helm v4.3.1 are still the newest of theirs. W7 now reads kubectl v1.37.2.
+`bench/wildct/install.sh` fetches it and checks it against its `.sha256` file (linux/amd64
+`c0448752...`), as before. The node images serve Kubernetes v1.37.1 (k3s v1.37.1+k3s1), so
+client and server share a minor version, as before. `install.sh` runs again before
+the first hold anyway, to install W7's newest k3d-gpu release (0.1.75 or later, not the 0.1.71
+installed on 2026-10-21), and that run installs kubectl v1.37.2. So every wildct row (H52w,
+H18w, H25w, H28w, H34w, H57wa, H57wm and H57wb, and H59w if it is pre-registered under W7) runs on
+it. If wildct amendment R1 (PR #226) merges, its sentence that keeps kubectl v1.37.1 is superseded
+by this amendment. Each row's Reproduce header names the kubectl that ran.
+
 ### H5 amendment A1: stage-1 harness and definitions (2026-10-19)
 
 Written 2026-10-19T22:30-07:00, before any H5 measurement. It fixes what the

README.md

@@ -573,7 +573,7 @@ name in place of `gpu`):
 ```bash
 docker exec k3d-gpu-server-0 touch /var/lib/rancher/k3s/server/manifests/nri-balloons.yaml.skip
 helm upgrade --install nri-resource-policy-balloons nri-resource-policy-balloons \
-  --repo https://containers.github.io/nri-plugins --version v0.14.0 \
+  --repo https://containers.github.io/nri-plugins --version v0.15.0 \
   --namespace kube-system -f values.yaml --force   # Helm 4: --force-conflicts
 ```
 
@@ -713,7 +713,7 @@ documents GPU support for NVIDIA only.
 ### Node Feature Discovery
 
 Every cluster also runs [Node Feature Discovery](https://github.com/kubernetes-sigs/node-feature-discovery)
-v0.19.0 with Intel's
+v0.20.1 with Intel's
 [NodeFeatureRules](https://github.com/intel/intel-device-plugins-for-kubernetes/tree/v0.37.1/deployments/nfd/overlays/node-feature-rules)
 v0.37.1, baked into the image (namespace `node-feature-discovery`). About 30 s
 after `up`, nodes carry the host's CPU, kernel and PCI features. On an
@@ -1485,7 +1485,7 @@ not switch the Intel GPU path.
 ### Agent sandboxes on E-cores
 
 `k3d-gpu sandbox up` installs [kubernetes-sigs/agent-sandbox](https://github.com/kubernetes-sigs/agent-sandbox)
-v1.0.7 (controller image pinned by digest) and a `python-ecores` SandboxTemplate
+v1.0.8 (controller image pinned by digest) and a `python-ecores` SandboxTemplate
 plus a one-replica warm pool in namespace `agent-sandbox`. Sandbox pods carry
 `k3d-gpu.io/cores: efficient` and request one CPU, so on a hybrid host untrusted
 agent tool execution runs on E-cores, apart from the P-cores that serve inference.
@@ -1506,7 +1506,7 @@ EOF
 ```
 
 The sandbox serves `POST /execute {"command": "..."}` on port 8888
-(`registry.k8s.io/agent-sandbox/python-runtime-sandbox:v1.0.7`, pinned by digest).
+(`registry.k8s.io/agent-sandbox/python-runtime-sandbox:v1.0.8`, pinned by digest).
 
 What works on k3d, and what does not:
 

bench/wildct/install.sh

@@ -1,11 +1,11 @@
 #!/usr/bin/env bash
 # install.sh K3D_GPU_VERSION: the wildct campaign's tools, pinned and checksum-checked, into
 # ~/k3dgpu-bench-wildct (EXPERIMENTS.md "Wildct pre-registration", rule W7); nothing
-# system-wide. k3d v5.9.0, helm v4.3.1 and kubectl v1.37.1 (each the newest release on
-# 2026-10-20) into bin/; the k3d-gpu release's .deb, checked against the release's
-# SHA256SUMS, unpacked into k3d-gpu/ (k3d-gpu/usr/bin/k3d-gpu with K3D_GPU_SHARE pointing at
-# k3d-gpu/usr/share/k3d-gpu); huggingface_hub 1.21.0 (crescent's pin) in hfvenv/. Prints
-# every version and checksum.
+# system-wide. k3d v5.9.0 and helm v4.3.1 (each the newest release on 2026-10-20) and kubectl
+# v1.37.2 (the newest on 2026-10-29, wildct amendment P1) into bin/; the k3d-gpu release's
+# .deb, checked against the release's SHA256SUMS, unpacked into k3d-gpu/
+# (k3d-gpu/usr/bin/k3d-gpu with K3D_GPU_SHARE pointing at k3d-gpu/usr/share/k3d-gpu);
+# huggingface_hub 1.21.0 (crescent's pin) in hfvenv/. Prints every version and checksum.
 set -euo pipefail
 V=${1:?usage: install.sh K3D_GPU_VERSION}
 W=${HOME}/k3dgpu-bench-wildct
@@ -23,8 +23,8 @@ check 0403a88ae20913d11b10a2467b50087721fea7581736da6b9d3992f5357fe54b helm.tgz
 tar -xzf helm.tgz linux-amd64/helm
 install -m 0755 linux-amd64/helm "${W}/bin/helm"
 
-get https://dl.k8s.io/release/v1.37.1/bin/linux/amd64/kubectl kubectl
-get https://dl.k8s.io/release/v1.37.1/bin/linux/amd64/kubectl.sha256 kubectl.sha256
+get https://dl.k8s.io/release/v1.37.2/bin/linux/amd64/kubectl kubectl
+get https://dl.k8s.io/release/v1.37.2/bin/linux/amd64/kubectl.sha256 kubectl.sha256
 check "$(cat kubectl.sha256)" kubectl
 install -m 0755 kubectl "${W}/bin/kubectl"
 

bench/xhost/tools.sh

@@ -0,0 +1,21 @@
+#!/usr/bin/env bash
+# tools.sh: the cross-host rows' k3d and kubectl in ~/k3dgpu-xhost/bin (EXPERIMENTS.md
+# "Cross-host pre-registration" rule 5, as cross-host amendment K1 moves it), each checked
+# against its upstream checksum file. Installs nothing else and touches no other file. Run on
+# crescent from a clean checkout of this repo, with the bench lock free. Prints both versions.
+set -euo pipefail
+X=${HOME}/k3dgpu-xhost
+K3D_V=v5.9.0 KUBECTL_V=v1.37.2
+mkdir -p "${X}/bin"
+DL=$(mktemp -d "${X}/tools-dl.XXXXXX")
+trap 'rm -rf "${DL}"' EXIT
+cd "${DL}"
+curl -fsSLo k3d "https://github.com/k3d-io/k3d/releases/download/${K3D_V}/k3d-linux-amd64"
+curl -fsSLo k3d.sums "https://github.com/k3d-io/k3d/releases/download/${K3D_V}/checksums.txt"
+curl -fsSLo kubectl "https://dl.k8s.io/release/${KUBECTL_V}/bin/linux/amd64/kubectl"
+curl -fsSLo kubectl.sha256 "https://dl.k8s.io/release/${KUBECTL_V}/bin/linux/amd64/kubectl.sha256"
+echo "$(grep 'k3d-linux-amd64$' k3d.sums | awk '{print $1}')  k3d" | sha256sum -c
+echo "$(cat kubectl.sha256)  kubectl" | sha256sum -c
+install -m755 k3d kubectl "${X}/bin/"
+"${X}/bin/k3d" version | head -1
+"${X}/bin/kubectl" version --client | head -1

docs/intel.md

@@ -32,8 +32,8 @@ with detection, enablement and what k3d-gpu does with each, see the
   manifests into k3s's auto-deploy directory:
   [intel-device-plugins-for-kubernetes](https://github.com/intel/intel-device-plugins-for-kubernetes)
   `0.37.1` (GPU, NPU, DSA, IAA, QAT), `intel/gaudi-device-plugin:1.24.2-481` and
-  [nri-plugins](https://github.com/containers/nri-plugins/releases/tag/v0.14.0)
-  balloons `v0.14.0`. With the default image nothing is `kubectl apply`'d at
+  [nri-plugins](https://github.com/containers/nri-plugins/releases/tag/v0.15.0)
+  balloons `v0.15.0`. With the default image nothing is `kubectl apply`'d at
   run time. If a custom `K3D_GPU_IMAGE` lacks a baked DaemonSet, `up` applies
   the manifest from `K3D_GPU_SHARE`; for NVIDIA with no local manifest it falls
   back to the upstream `main` branch manifest from GitHub, which is unpinned.

docs/intel/upstream-eat.md

@@ -78,7 +78,7 @@ EAS (main e8f3ebc; one concern per PR, in this order)
 ├── roles/kubernetes/tasks/apply_node_topology.yaml 0004
 ├── docs/customize/node_topology.md                 0004  "label by hand" text replaced
 ├── docs/deploy/topologies.md                       0004
-├── roles/nri_cpu_balloons/defaults/main.yaml       0003  nri_cpu_balloons_chart_version: "v0.14.0"
+├── roles/nri_cpu_balloons/defaults/main.yaml       0003  nri_cpu_balloons_chart_version: "v0.15.0"
 ├── roles/nri_cpu_balloons/tasks/install_advanced.yaml 0003
 ├── docs/deploy/k3d.md                              0005  new page "Single-Node Dev on k3d"
 └── docs/README.md                                  0005  +1 row under Deploy

images.lock

@@ -388,28 +388,28 @@
       "tag": "v3.15.0"
     },
     {
-      "digest": "sha256:f254384c24605436ce35d7077427559e24bda8f5691cff5a54a4c2e030d81083",
+      "digest": "sha256:e1d96312487eba7d8f001e848a9d4e644b7deea8c2a4564ccc318abb729be731",
       "name": "registry.k8s.io/agent-sandbox/agent-sandbox-controller",
       "sources": [
         "share/agent-sandbox.yml"
       ],
-      "tag": "v1.0.7"
+      "tag": "v1.0.8"
     },
     {
-      "digest": "sha256:fc92654c97fe5dff2a7935a845820981026c075d0c1014e5c2c569f5e58204f1",
+      "digest": "sha256:41da84809426098cb04cdafbb41855f15e3604f7455ef53816bad666e532e787",
       "name": "registry.k8s.io/agent-sandbox/python-runtime-sandbox",
       "sources": [
         "share/agent-sandbox-k3d.yml"
       ],
-      "tag": "v1.0.7"
+      "tag": "v1.0.8"
     },
     {
-      "digest": "sha256:2fa1c99ad09bdf2c8ad97706a4ad2fd548c84d5ecd70ba32a6152c667b96c4d2",
+      "digest": "sha256:6fc5c9af9176e1acb5ce57a4a1d0c0d3eb25cd34f5bdd0efefebf0efe45a8379",
       "name": "registry.k8s.io/nfd/node-feature-discovery",
       "sources": [
         "share/node-feature-discovery.yml"
       ],
-      "tag": "v0.19.0"
+      "tag": "v0.20.1"
     },
     {
       "digest": "sha256:5d064465d751a73a729e7ecdc32c0ba5de0df1d305debb60a3d9a85b4fd3a967",