[88plug]

This PR moves the bench arms to the newest releases where the ledger allows it. Where a pin has to stay, the reason is now written next to it. Every version below was re-read on 2026-10-29 from the registry or the project's releases:

Project	Newest release on 2026-10-29	Where it was read
vLLM	v0.32.0	PyPI and GitHub; v0.32.1rc0 and v0.33.0rc0/rc1 are prereleases
EAT	release-1.1.0 (0c726f57)	GitHub; main is one diagram commit ahead
llama.cpp	v0.8.0, commit 8642541b, build b11960	GitHub; b11962 to b11981 are prereleases
OVMS	2026.4.1	registry
SGLang	v0.5.22-xeon	registry
stress-ng	V0.22.01	GitHub

Every new or kept digest in these changes resolves with docker buildx imagetools inspect.

Bumps, each with a dated amendment, stacked on the unmerged pre-registrations

These two rows have not run, and their pre-registrations are not on main yet, so the amendments target those PRs' branches:

• ledger(H60): amendment V1, both arms on llama.cpp v0.8.0 (server-b11960) #234 into ledger(H60): pre-register Qwen3.8-Flash-Next from host DRAM on the Xeon 6741P #199 (H60), amendment V1. Both arms move from llama.cpp server-b11783 to server-b11960 (v0.8.0, sha256:8052c13e…). Checked before the bump:
  • b11783 is an ancestor of b11960.
  • The two images' llama-server --help outputs differ only by a router-mode flag.
  • The qwen4exp loader reads the same GGUF keys. It falls back to the separate tensors, and the pinned Flash-Next GGUF carries only those (its two shard headers were read).
  • The IQ4_NL repack lines are unchanged.
• ledger(H7x): amendment V1, arms B and C on EAT release-1.1.0 (vLLM CPU v0.28.0) #235 into ledger(H7x): pre-register H7 to its full protocol on the Xeon 6741P #196 (H7x), amendment V1. Arms B and C move to EAT release-1.1.0's stock: vLLM CPU v0.28.0@sha256:197354e6… and the chart at 0c726f57.
  • h7x.sh now pins the tag itself and refuses any chart that is not release-1.1.0's (checked by sha256).
  • Rendered through h7x.sh, the two releases differ only in the image and the chart's fsGroup: 1001.
  • The amendment records that v0.28.0 no longer reads VLLM_CPU_SGL_KERNEL, so B's dense bf16 projections run on oneDNN.
  • It asks for one more setup smoke before the hold.

Merge each into its base before that pre-registration lands.

This PR (main)

• bench/apb/h7.sh and bench/apb/arm.sh (eat-vllm-cpu) pin v0.24.0, the image H7 and H8 ran. Both took the tag from integrations/eat/vllm-values.yaml, which is due to move to EAT's v0.28.0. After this PR, that move no longer changes what H7's and H8's scripts reproduce.
  • Renders from EAT 0d3438d's chart are byte-identical before and after the change.
  • They stay on v0.24.0 when the overlay's tag is moved to v0.28.0.
  • A new bench_test.sh check covers the pin. It failed before the change.
• Reasons recorded next to the held pins (comments only):
  • h54.py's llama.cpp b11749: H54q (amendment Q1), H57dp/H57de and H57wa/wm/wb pair with or replicate b11749 records. A newer build needs a dated amendment for each of those rows first.
  • The vLLM CUDA v0.31.0 arms in live.sh (A3/A3ref) and apb/arm.sh (vllm-cuda): only done rows ran them, and the pin is kept so those rows reproduce.

Kept as they are

• Already at the newest release:
  • h54.py's vLLM CPU v0.32.0, OVMS 2026.4.1 and SGLang v0.5.22-xeon.
  • stress-ng V0.22.01.
  • The EAT-stock pins H50, H50o, H51, H52b and H53 use: live.sh B2 and LiteLLM, B3 2025.4, and the pairs/apb hold wrappers' v0.28.0. These match EAT release-1.1.0.
• Records of runs that are done (class c):
  • agent.lock's H40-H48 entries.
  • live.sh's B1 default (v0.24.0).
  • h54.py's vllm024 (B1g).
  • The H54b/H54c GPU images.
  • The C1/C2 desktop arms.
• Pins that pending rows replicate:
  • full-b11749, the drafter converter that the committed weights' sha256 depends on.
  • The k3d-gpu 0.1.61 and 0.1.63 node images, under E2 item 1 and H54 R1.
  • probe.sh's python:3.13-slim, the image of H28's measured pods, which H28x and H28w replicate.

Coordination

• nightly: repin the moved digests and keep the digest PR out of .github/workflows/ #220 (nightly repin).
  • It rewrites bench/eatflow/agent.lock (proxy_python, redis and the redis read date) and bench/zero-copy-models/probe.sh's PYIMG. The second one is the measured pod image of the pending H28x and H28w, so either nightly: repin the moved digests and keep the digest PR out of .github/workflows/ #220 drops those files or the two rows get a dated amendment before they run.
  • scripts/images-lock repin scans bench/ and will re-pin any bench pin whose tag the registry re-points, pending rows included.
  • This PR does not touch the lines nightly: repin the moved digests and keep the digest PR out of .github/workflows/ #220 edits. The RAPL reader (python:3.14.8-slim → 3.14.10-slim@sha256:3cabfd7d…) and the H53 card placeholder (busybox:1.37 → 1.38.0@sha256:a46257e0…) in live.sh move in a follow-up on top of nightly: repin the moved digests and keep the digest PR out of .github/workflows/ #220. That follow-up carries an H53 mechanics amendment. BusyBox labels 1.38.0 "unstable", as it labelled 1.37.0; its stable line is 1.36.1.
• agent: llama.cpp backend the fit picks, and busy Intel cards left out of the device plugin #182 already pins llama.cpp b11960 for the agent backend, so H59's arms are current once it lands.
• integrations/eat/vllm-values.yaml (integrations area) can move to v0.28.0 once this PR and ledger(H7x): amendment V1, arms B and C on EAT release-1.1.0 (vLLM CPU v0.28.0) #235 are in. After that, no bench row reads its tag.

Tests run

• tests/bench_test.sh: 59 passed.
• scripts/images-lock check.
• tests/image_pins_test.sh: 39 passed.
• tests/claims_check.sh.
• tests/ledger_drift_check.py.
• The bench unit suites in bench/eatflow (pairs, solo, agentbench, crescent holds), bench/h54, bench/h57d and bench/apb (apb, stage).
• shellcheck at warning severity on every changed script.