[88plug]

Moves the repo's tooling pins to their latest releases, checked on 2026-10-29. It also adds crescent amendment E4, which moves the crescent workspace's kubectl, helm and uv before the pending rows run.

Rebased on main e5748cd1 (2026-10-29, after batch #241, #242, #232, #244, #245 and #246). The last rebase, from f53eac15, applied cleanly: #244, #245 and #246 touch none of this PR's files except EXPERIMENTS.md and README.md, in other sections. E4 still reads true after them: H60's amendment V1 (#244) changes its llama.cpp image, not the workspace it runs from, and wildct amendment R3 (#245) is not a crescent row.

The first rebase, onto f53eac15: #220 is squash-merged, so the branch now carries only its own commits. The rebase dropped #220's commits; .github/workflows/ci.yml, docs-build/Dockerfile and tests/agent_examples_test.sh then applied cleanly. The only conflict was EXPERIMENTS.md, where E3 and the H50, H50o and H51 sections now follow E2; keep-both puts E4 after them, before the H53 pre-registration.

Every pin was read again at the rebase (2026-10-29, about 19:45Z). One had moved: langchain-core 1.7.2 was released at 19:04Z, so the langgraph example moved to it and all five constraints files were refrozen. They were read again at the last rebase (about 21:20Z), with the use-latest-version checks, PyPI for every docs lock pin, Docker Hub, MCR, Chrome for Testing, dl.k8s.io and the nixos-26.05 head. One had moved again: langchain-core 1.7.3 was released at 20:55Z, so the langgraph example moves to it. All five examples were refrozen in python:3.13.18-slim-trixie, and only that line changed. Everything else, including each image digest, the nixos-26.05 head, Chrome Stable and the docs lock regenerated from scratch, is unchanged.

How versions were checked

• The use-latest-version registry checks, plus the registry itself for each pin:
  • gh api releases, and commits/<tag> for every action SHA.
  • PyPI JSON, including requires_python.
  • scripts/images-lock digest and docker buildx imagetools inspect for every image digest.
  • dl.k8s.io stable.txt and .sha256, get.helm.sh sums, and Chrome for Testing's last-known-good Stable.
  • The MCR and Docker Hub tag lists, python.org, and the nixos-26.05 branch head.
• Docker Hub rate-limited anonymous pulls here, so local runs pulled the same digests through mirror.gcr.io. A digest names the same content on either host.

Changes (class "ours")

Pin	From	To
actions/checkout (all workflows)	v6.1.0	v7.0.1 3d3c42e5
docker/setup-buildx-action	v3.12.0	v4.4.1 f87e5991
docker/login-action	v3.7.0	v4.7.0 81c6f93d
docker/build-push-action	v6.19.2	v7.4.0 c3c9e263
softprops/action-gh-release	v2.6.2	v3.0.3 efb35369
hadolint/hadolint-action	v3.3.0 (hadolint 2.14.0)	v3.5.0 06be81ba (hadolint 2.15.1, as the devcontainer)
KSXGitHub/github-actions-deploy-aur	v4.1.3	v4.2.0 084b0d9b
promtool (CI)	3.15.0	3.15.1, sha256 6a726885...
Prometheus server (share/telemetry/prometheus.yml, images.lock)	v3.15.0	v3.15.1 fa330227
Chrome for Testing (CI)	155.0.8059.84	156.0.8078.73
packaging/install.sh kubectl	v1.37.1	v1.37.2, both arch sha256
flake.lock nixpkgs (nixos-26.05)	26f4107e	08d3a975 (branch head 2026-10-28)
nix check image	nixos/nix 2.35.3	2.35.4 f159a58f
Homebrew check image	brew 7.0.9	7.0.11 e50d8c46
devcontainer base	2.2.3-trixie	2.2.5-trixie 71b6fe7c
docs interpreter	python:3.14.8-slim	python:3.15.2-slim c4158543
docs lock	mkdocs-material 9.7.7	9.8.0, lock regenerated from scratch
agent examples image	python:3.13.16-slim-trixie	python:3.13.18-slim-trixie 76baad60
crewai / langgraph / langchain-openai / langchain-core / openai	1.15.28 / 1.2.15 / 1.7.2 / 1.6.11 / 3.31.0	1.15.33 / 1.2.18 / 1.7.3 / 1.7.3 / 3.41.0
console/convert-pins.json torch / torchvision	2.14.1 / 0.29.1	2.15.0 / 0.29.2 (hack/bump-convert-pins)
amx-autoquant probe image	python:3.12-slim (3.12.15)	python:3.14.10-slim 3cabfd7d
amx-autoquant probe torch / torchvision	2.14.1+cpu / 0.29.1+cpu	2.15.0+cpu / 0.29.2+cpu
kserve example cert-manager	v1.21.2	v1.21.3

Notes on what the bumps needed:

• Homebrew 7.0.11 loads a non-official tap's formulae implicitly only for install, reinstall and upgrade (Limit implicit item trust to install commands Homebrew/brew#24352). The first run of packaging/check.sh brew failed at brew style with "Refusing to load formula ... from untrusted tap". homebrew-check.sh now runs brew trust --tap 88plug/k3d-gpu after tapping, and the full check passes.
• Python 3.15.2 is the newest Python. The docs toolchain installs from the hashed lock and builds --strict on it.
• The agent examples stay on 3.13, because crewai 1.15.33 and litellm 1.83.14 both require Python < 3.14.
• The probe moves to 3.14, the newest Python openvino 2026.4.1 ships wheels for (cp310-cp314).
• torchvision 0.29.2+cpu exists on the PyTorch CPU index (py3-none), so the probe keeps the +cpu pair.
• The docs lock was regenerated from scratch, so every transitive pin is the newest allowed: certifi 2026.10.23, markdown 3.11.2, platformdirs 4.14.0, python-dateutil 3.0.0, plus tzdata 2026.6. A plain lock.sh run keeps the old transitive pins.
• actions/checkout v7 blocks fork-PR checkouts under pull_request_target and workflow_run. No workflow here uses either trigger.
• The docker action majors drop deprecated inputs and the DOCKER_BUILD_* envs. None are used here.

Crescent amendment E4 (EXPERIMENTS.md)

• What moves. Shared rule 5 pins kubectl v1.35.2, helm v4.1.1 and uv 0.8.0. The rows' node image (0.1.61) runs k3s v1.37.1, two minors past kubectl v1.35, outside kubectl's one-minor skew.
• Why an amendment. E2 item 3 says these moves take a further amendment.
• When. E4 was recorded before any H52b, H53 or H58 hold, and covers H53, H58 and later crescent rows that run from ~/k3dgpu-bench/bin.
• Revised at the rebase (19:50Z), after H52b's hold and before any H53 or H58 hold. No pin, checksum, check or render changed.
  • Item 2: H52b came first. Its hold (18:32-18:43Z, ledger(H52b): REFUTED on the Xeon 6741P after host-setup --apply; amendment D3 #232) ran before setup.sh, on kubectl v1.35.2 and helm v4.1.1, and its records name them.
  • Item 4: H59 and H60 have merged. H60 runs from ~/k3dgpu-xhost/bin (its Reproduce block), which cross-host amendment K1 covers, so E4 does not cover it. H59 has no runner yet.
  • Item 5: H54q's runner amendment Q2 was pushed at 19:34Z (ledger(H54q): INCONCLUSIVE on the Xeon 6741P; runner amendment Q2 and the run #243). It keeps H54a's tools in ~/k3dgpu-gemma and renders no arm with helm.
• Scope. Mechanics only: no arm, prediction, statistic, gate or cap changes.
• setup.sh installs kubectl v1.37.2, helm v4.3.1 and uv 0.13.6, each still checked against its upstream checksum.
  • uv 0.13 refuses to overwrite an existing venv (uv venv exits 2 without --clear), so setup.sh keeps an existing mafvenv.
  • Under a scratch HOME, a fresh run and a re-run both pass, and the venv equals banking-venv.txt.
  • A uv 0.13.6 build of the venv freezes identical to a uv 0.8.0 build. The interpreter is a later python-build-standalone build of the same 3.12.11.
• Render check, helm v4.1.1 vs v4.3.1. It covers H53's B2, H54's eat-vllm and the 0d3438d ovms chart.
  • Each pair renders the same objects. The text differs only in blank lines and trailing whitespace.
  • After stage.py, the staged B2 and B1g are byte-identical.
• H54q keeps its own ~/k3dgpu-gemma/bin, as its runner amendment Q2 (ledger(H54q): INCONCLUSIVE on the Xeon 6741P; runner amendment Q2 and the run #243) says. E4 records the render check for its EAT renders.
• EAT_REV stays 0d3438d, against the audit's suggestion.
  • Its checkout serves only H50o's B3 chart, a row that has run.
  • The shared LiteLLM gateway is not rendered from EAT's genai-gateway chart. live.sh runs EAT's image as a container with its own config.
  • EAT release-1.1.0, still EAT's latest release, pins the same image tag, and its digest 73cbb215... is unchanged. Moving EAT_REV would change no pending row's objects.

Kept on purpose

• EAT-stock pins at EAT's latest release (release-1.1.0, 0c726f57): the bastion's helm v3.14.0, the LiteLLM example's litellm 1.83.14, and the banking venv (EAB v1.0.0).
• Records of runs already done: the BFCL and coding constraints, the crescent ov-export-venv.txt and export_qat_ov.sh, and docs/security-audit.md's dated table.
• Builds that pending rows reuse: bench/h54/convert.sh and ov-export-venv.txt (H54q continues H54a's builds), and bench/h57d/convert.sh (H57dp, H57de and H57w* need H57d's drafter bytes).
• hack/tools.env stays on helm v3.22.1 (the newest v3), because helm v4 renders differently.
• Bench harness images that pending rows run: bench/eatflow/agent.lock proxy_python and live.sh RAPL_IMAGE (python:3.14.8-slim), and the python:3.13-slim images of the wildct, zero-copy and model-share paths. Moving them needs those rows' own amendments.
• Already current: fedora:44 (45 is still beta), debian:trixie-slim, nfpm, shellcheck, actionlint, kubeconform, jsonschema, packaging, pip-tools and k3d.

Not in this PR

• The cross-host and wildct kubectl are already v1.37.2 on main: bench/xhost/tools.sh under cross-host amendment K1, and bench/wildct/install.sh under wildct amendment P1 (both landed with batch batch: land #182, #236, #195, #224, #237, #226, #239, #230, #210, #233, #196, #235, #207, #213, #228, #227 #241).
• Nothing was pushed to other workflows' branches.

Tests

At the last rebase (main e5748cd1, with langchain-core 1.7.3):

• tests/claims_check.sh 26 files clean; claims_check_test.sh 81 passed; tests/ledger_drift_check.py 41 files agree with 81 rows; ledger_drift_check_test.py OK; pe_premise_check 16 passed and ok; model_cards_license_test.py OK.
• scripts/images-lock check and pins ok; image_pins_test 43, ci_wiring_test 69, gate_merge_test 31, packaging_test 126, kserve_test 29, platforms_doc_test 195, detect_test 1185, console_test 61, quickstart_test 174, wildct_hold_test 41, xhost_tools_test 18, hold_own_test 63, bench_test 71, autoquant_test 27 passed. crescent/test_holds.py, wildct/test_wildct.py, test_pairs.py, h54/test_h54.py and h59/test_h59.py OK.
• tests/agent_examples_test.sh: all five examples pass in python:3.13.18-slim-trixie.
• shellcheck 0.11.0 at warning severity clean on every changed script.

At the first rebase:

Run at the rebase. After #232 landed, the ledger, pin, packaging, kserve, bench, console, docs and crescent checks were run again at the final head (main f53eac15):

• tests/claims_check.sh 26 files clean; tests/claims_check_test.sh 81 passed; tests/ledger_drift_check.py 41 files agree with 81 rows; ledger_drift_check_test.py OK; pe_premise_check 16 passed and ok; model_cards_license_test.py OK.
• scripts/images-lock check and pins both ok. scripts/images-lock repin over every file whose image this PR moves changes nothing.
• tests/image_pins_test.sh 43 passed; tests/ci_wiring_test.sh 69 passed.
• tests/agent_examples_test.sh: all five examples pass in python:3.13.18-slim-trixie, with the refrozen constraints.
• Docs in python:3.15.2-slim: hashed install, tests/docs_test.sh (63 passed) and mkdocs build --strict. docs-build/lock.sh from scratch reproduces the committed lock byte for byte.
• tests/telemetry_test.sh with promtool 3.15.1, sha256-checked as CI installs it: 27 passed.
• tests/report_html_test.sh with Chrome for Testing 156.0.8078.73, headless required: 48 passed.
• packaging/check.sh nix in nixos/nix 2.35.4: nix flake OK. packaging/check.sh brew on brew 7.0.11: homebrew formula OK.
• tests/packaging_test.sh 126, kserve_test 29, quickstart_test 174, quickstart_adversarial 8, bench_test 71, console_test 61, model_opt_test 37, autoquant_test 27, audit_test 49, xhost_tools_test 18, model_share_test 52, image_test 26, eat_test 64, host_setup_test 269, gate_merge_test 31, auto_release_cas_test 17, register_model_test 23, chat_ui_test 25, wildct_hold_test 38, hold_own_test 63. crescent/test_holds.py, crescent/test_crescent.py, test_pairs.py, h54/test_h54.py and h59/test_h59.py OK.
• Lint: actionlint 1.7.12 clean; hadolint 2.15.1 clean at the error threshold on the three Dockerfiles; shellcheck 0.11.0 at warning severity clean on every changed script.

[88plug]

Independent review of c2a68d6: approve. No defects found; nothing pushed to the branch.

Versions, re-read 2026-10-29 (~20:40Z)

• Every action SHA equals gh api repos/<repo>/commits/<tag> for its latest release. The kept actions (setup-chrome v2.2.0, setup-helm v5.0.1, action-shellcheck 2.0.0, create-pull-request v8.1.1, the pages actions) are also latest.
• hadolint-action v3.5.0's Dockerfile is hadolint:v2.15.1-debian.
• docker buildx imagetools inspect matches every moved digest: brew 7.0.11, devcontainers/base 2.2.5-trixie, prometheus v3.15.1, python 3.15.2-slim, 3.13.18-slim-trixie and 3.14.10-slim. Each is the newest tag in its line. The kept debian:trixie-slim, fedora:44 and nfpm v2.48.0 digests are current.
• The sha256 values match upstream: kubectl v1.37.2 (amd64 and arm64), promtool 3.15.1, helm v4.3.1 and uv 0.13.6.
• The nixos-26.05 head is 08d3a975, Chrome for Testing Stable is 156.0.8078.73, nix 2.35.4 is the newest tag, and k3s v1.37.1+k3s1 is still the newest k3s.
• The PyPI pins are all latest. The deliberate exceptions are correct:
  • litellm 1.83.14 is EAT release-1.1.0's gateway tag, and the index digest 73cbb215 is unchanged.
  • transformers and huggingface-hub sit at optimum-intel 2.2.0's <5.6 and <1.22 bounds.
  • crewai 1.15.33 and litellm 1.83.14 both require Python <3.14.
• torchvision 0.29.2 requires torch>=2.14.0 (py3-none wheels), so the torch 2.15.0 pair resolves. hack/bump-convert-pins --dry-run moves nothing. With uv, the probe's PIP_PINS resolve for cp314 to 57 packages.

Amendment E4

• It adds lines only, and no prediction is edited.
• PR #238 was opened at 11:14Z with E4, before H52b's 18:32Z hold. H53 and H58 are still PENDING.
• All five anchors resolve. Q2's commit time (19:34:29Z, #243) matches the text.
• On current main, H60 still runs from ~/k3dgpu-xhost/bin after V1 (#244), so item 4 holds.
• Item 3, reproduced with uv 0.13.6:
  • uv venv over an existing venv exits 2 without --clear.
  • A fresh build gets the 2025-10-07 build of 3.12.11, and its freeze equals banking-venv.txt (71 packages).
  • A uv 0.8.0 build gets the 2025-07-11 build, and uv 0.13.6 over it reports "Checked 71 packages" with nothing installed.
• The render check, reproduced with helm v4.1.1 against v4.3.1 at c86c42b and 0d3438d:
  • B2 has 5 objects, eat-vllm 6 and ovms 4.
  • Each pair parses equal, and the text differs only in blank lines.
  • After stage.py, the staged B2 (xpu) and B1g (vllm) are byte-identical.

Suites (PR head, and a local merge with main d997ea89)

• The ledger checks pass on both trees: claims_check (26 files clean), ledger_drift_check (41 files, 81 rows) and pe_premise_check.
• The PR head merges cleanly with main d997ea89.
• Test suites, all passing:
  • image_pins 43, ci_wiring 69, packaging 126, kserve 29, quickstart 174 and adversarial 8.
  • bench 71, autoquant 27, console 61, audit 49, image 26, eat 64, xhost_tools 18, model_share 52, model_opt 37.
  • host_setup 269, gate_merge 31, register_model 23, chat_ui 25, wildct_hold 38, hold_own 63.
  • test_holds, test_crescent, test_pairs (63), test_h54 and test_h59 all report OK.
• Container checks:
  • telemetry_test with sha-checked promtool 3.15.1: 27 passed.
  • report_html_test with Chrome for Testing 156.0.8078.73, headless required: 48 passed.
  • All five agent examples pass.
  • In python:3.15.2-slim, the docs install with hashes, docs_test passes 63, and mkdocs build --strict succeeds. lock.sh from scratch reproduces requirements.txt byte for byte.
  • check.sh brew (7.0.11, with brew trust --tap) and check.sh nix (2.35.4) pass.
• Lint is clean: actionlint 1.7.12, hadolint 2.15.1 at the error threshold, and shellcheck 0.11.0 at warning severity on the changed scripts.

Non-blocking: in E4's "Why", the "(released 2026-10-27)" reads as if it covers helm v4.3.1 too. helm v4.3.1 was released on 2026-10-15; only uv 0.13.6 was released on 2026-10-27.