[88plug]

Moves the engine pins we own to the newest stable releases, read on 2026-10-29. No EXPERIMENTS.md row, prediction or arm changes.

Stacked on #182. The base is claude/llamacpp-backend, because #182 changes the same vver line in agent_optimize and adds the b11960 llama.cpp images beside the Vulkan pin. Merge #182 first, then retarget this PR to main (or rebase it onto main if #182 is squashed). Only the four commits after #182's head d2e2dc91 belong to this PR.

What moves

Every digest below was resolved with docker buildx imagetools inspect on 2026-10-29. Every version was checked as the newest stable release with the use-latest-version registry lookup and the registry itself: PyPI and GitHub releases for vLLM, GitHub releases for llama.cpp and Ollama, and the Docker Hub and ghcr tag lists.

Pin	Was	Now
vllm agent backend (Intel GPU), scripts/k3d-gpu + images.lock	vllm/vllm-openai-xpu v0.31.0 @95ac8150	v0.32.0 @20509fab (index; linux/amd64 9dd913e9)
model optimizer vLLM version: agent_optimize vver, console/plan.py AGENT_VLLM_VERSION["GPU"] and its fallback, console/model_opt.py default	0.31.0	0.32.0
LLAMA_VULKAN_IMAGE (the llama.cpp pin in facts/v1 that plan's Vulkan cells compare against), images.lock	server-vulkan-v0.5.0 @be586c64	server-vulkan-b11960 @c9aba1a6 (release v0.8.0 = tag b11960, commit 8642541b, the image's revision label; amd64 and arm64)
console/moe_predict.py llama.cpp engine images (plan --predict)	server-b11749 / server-intel-b11749	server-b11960@8052c13e / server-intel-b11960@068cd41c, the same digests as #182's LLAMA_CPU_IMAGE and LLAMA_SYCL_IMAGE
bench/competitors.lock llamacpp stack (k3d-gpu bench --vs)	server-b11706@f41d47ee	server-b11960@8052c13e; docs link to v0.8.0 common/arg.cpp
bench/competitors.lock ollama stack	ollama/ollama:0.40.5@5d3a2f25	0.40.13@b2f0122c (GitHub latest, 2026-10-28, not a prerelease; amd64 and arm64); docs link to v0.40.13 openai/openai.go

llama.cpp's GitHub "latest" release is v0.8.0 (2026-10-28). The b119xx tags after it are marked pre-release, so b11960 is the newest stable build. ghcr has published no semver image tags since v0.5.0, so the v0.8.0 images are the *-b11960 tags.

Compatibility, read from source at the new versions

• vLLM v0.32.0, from its release notes, vllm/engine/arg_utils.py, vllm/entrypoints/launchers/cli_args.py, envs.py, the tool and reasoning parser registries and docker/Dockerfile.xpu:
  • Every flag the agent manifest and the model optimizer emit still exists: --revision, --served-model-name, --port, --dtype, --block-size, --max-model-len, --max-num-batched-tokens, --gpu-memory-utilization, --enforce-eager, --no-enable-prefix-caching, --max-num-seqs, --kv-cache-dtype, --tensor-parallel-size, --enable-auto-tool-choice, --tool-call-parser, --reasoning-parser, --default-chat-template-kwargs, --language-model-only, --override-generation-config, --speculative-config and --trust-remote-code.
  • The parsers the optimizer names (hermes, qwen3_coder, gemma4; reasoning qwen3, gemma4) are still registered.
  • The env vars in VLLM_XPU_ENV are still read.
  • The image entrypoint is still exec-form vllm serve, and there is no /opt/intel/oneapi.
  • The image still ships Level Zero 1.32 and compute-runtime 26.27, so the launcher comment stays true.
  • The optimizer always sets num_speculative_tokens, so v0.32.0's check of that field before it applies a default does not apply here.
  • The breaking changes do not reach this path. CORS now defaults to no origins, but nothing calls vLLM from a browser: Open WebUI and the gateway call it server-side. qwen3_xml changed, but the optimizer names qwen3_coder. The other breaking changes are cascade attention, mamba_cache_mode, and transformers inside the image.
  • The model optimizer's vLLM source links move to v0.32.0 at the lines re-read there: --language-model-only is now arg_utils.py L1460, and the hardware table (L67) and the Gemma 4 MTP section (L12-36) are unchanged.
• llama.cpp v0.8.0: the competitor args -mu, -m, -a/--alias, --host and --port are in common/arg.cpp (L3066, L3057, L3033, L3343, L3359), and /health is still served (tools/server/server.cpp L171). b11749 to b11960 and b11706 to b11960 are ahead-only. The Vulkan pin is not deployed by any launcher manifest. It is the facts/v1 release.pins["llama.cpp"] that plan compares against a matrix cell's image.
• Ollama v0.40.13: v0.40.5 to v0.40.13 is 55 commits ahead and 0 behind. openai/openai.go still streams thinking in delta.reasoning, which bench/llm_load.py counts. The model blob pin is a registry digest and does not move.
• README: the vllm row no longer says "validated on Arc Pro B-series". It names what vLLM's own XPU guide lists (Data Center GPU and Arc) and says v0.32.0 has not been run here on an Arc B-series card. No B-series card was available for this change.

What stays, and why

• Already the newest stable:
  • Gaudi vllm-0.30.0-ptupstream-2.13.0 1.24.2-481. vllm-gaudi v0.30.0 is the newest release, and vault has no 1.24.3 or 1.25.x repository.
  • intel/llm-scaler-vllm:1.5. The 0.31.0-b1-experimental tag has no release.
  • OVMS 2026.4.1 / 2026.4.1-gpu.
  • vllm/vllm-openai-cpu:v0.32.0.
  • lmsysorg/sglang:v0.5.22-xeon.
  • intel/dlstreamer:2026.2.0. The 2026.3.0 tags are weeklies.
• EAT-stock arms of the pending crescent rows are already EAT release-1.1.0 (0c726f57, still EAT's latest release): B1 on vllm-cpu v0.28.0 per amendment E1, B2 on vllm-xpu v0.28.0 per E2 item 5, and B3 on OVMS 2025.4. So no mechanics amendment is needed on main.
• Kept on purpose for pending rows:
  • H54q's Lg stays at b11749 in bench/h54/h54.py: amendment Q1 runs items 50-199 at the configs of H54a's items 0-49, and H57d and H57w import h54.IMG.
  • The wildct H57wa/H57wm pull list stays at b11749: it replicates H57d's arms.
  • bench/eatflow/crescent/h52.sh's enumeration image is H52b's probe tool. H52b has run (ledger(H52b): REFUTED on the Xeon 6741P after host-setup --apply; amendment D3 #232).
• Historical records are untouched: bench/apb/*, bench/eatflow/agent.lock, live.sh C1/C2/A3, the B1 default, and every results tree.

Not in this PR (each needs the owning branch)

• H60 (ledger(H60): pre-register Qwen3.8-Flash-Next from host DRAM on the Xeon 6741P #199): its runner and Engine paragraph pin server-b11783. The latest is server-b11960@8052c13e, which is ahead-only and has the Flash-Next MTP fix. Amend it before H60 runs.
• H7x (ledger(H7x): pre-register H7 to its full protocol on the Xeon 6741P #196) and integrations/eat: H7x's B and C arms read integrations/eat/vllm-values.yaml. That file and eat-up.sh stay on EAT 0d3438d with vLLM CPU v0.24.0 until ledger(H7x): pre-register H7 to its full protocol on the Xeon 6741P #196 moves B/C to release-1.1.0 with v0.28.0 @197354e6. Moving eat-up.sh also needs the guardrail switch docs(eat): refresh the EAT upstream status, checked 2026-10-28 #211 names and a fresh H1b-style deploy. That deploy was not run here: this host is the desktop bench box.
• H57dp/H57de (ledger(H57dp, H57de): pre-register the unpinned llama.cpp arm on the i9-13900T, with its runner (amendment R1) #207, ledger(H57dp): amendment R2, the bench lock, the hybrid vendor, the quality records and the Reproduce header #215): these rows use server-b11749 through h54.IMG. A pin scoped to these rows at server-b11960@8052c13e (with the drafter reconverted with full-b11960@7afc81bf, or the b11749 GGUF shown to load) belongs in those PRs before the rows run.
• H58: it has no pins of its own and takes main's agent defaults at its runner amendment. Landing this PR (after agent: llama.cpp backend the fit picks, and busy Intel cards left out of the device plugin #182) first gives it the newest engines.

Checks run locally (on top of #182's head)

• scripts/images-lock check and scripts/images-lock pins pass.
• Suites:
  • tests/image_pins_test.sh 39/0
  • tests/console_test.sh 61/0
  • tests/report_html_test.sh 48/0
  • tests/sim_matrix_test.sh 256/0
  • tests/model_opt_test.sh 37/0
  • tests/console/test_model_opt.py 45 OK
  • tests/audit_test.sh 49/0
  • tests/fit_test.sh 34/0
  • tests/llamacpp_test.sh 47/0
  • tests/eat_test.sh 64/0
  • tests/platforms_doc_test.sh 195/0
  • tests/docs_test.sh 61/0 (mkdocs from docs-build/requirements.txt)
  • tests/bench_vs_test.sh 23/0
  • tests/bench_vs_test.py 27 OK
  • tests/bench_test.sh 58/0
  • tests/agent_kubeconform.sh 100 valid
  • tests/detect_test.sh 1185/0
  • tests/console_mcp_test.sh
  • tests/image_test.sh 26/0
  • tests/prefill_guard_test.sh 77/0
  • tests/claims_check.sh and tests/claims_check_test.sh
  • tests/ledger_drift_check.py and its self-test
• The goldens that change were regenerated with each suite's --update mode. The diffs are the Vulkan pin line in the 32 fixture facts.json files, the console fallback goldens and report.html, plus each report's facts sha256 line, nothing else. tests/console/matrix-fallback.json (synthetic cells) and tests/console/model-opt/xeon-b70.facts.json move to the new pin so the fallback cells stay green.
• tests/bench_vs_test.sh expects the new competitor pins.

[88plug]

Landed in #241 (merge commit 76bbfbd).