Run pull request tests on Blacksmith

[webdevtodayjason]

Summary

• move the pull request and main branch test job from ubuntu-latest to blacksmith-4vcpu-ubuntu-2404
• make live cluster-secret reloads recognize the atomic file replacements used by NodeConfig.save() on Blacksmith filesystems
• preserve the three Spark image jobs because they run only for tags or workflow_dispatch
• audit the other branches identified in the 2026-10-05 hosted-runner report

Verification

• rg -n 'runs-on:' .github/workflows/*.yml
• no GitHub-hosted runs-on labels remain
• Ruff passes
• both cluster-secret rotation regression tests pass locally
• full local Pytest on macOS with Python 3.14: 3,250 passed, 2 skipped, 1 xfailed, and one unrelated cache-order failure in tests/test_reachable_urls.py; that test passed on both Blacksmith attempts
• the Tests gate must pass on Blacksmith before merge

Preserved self-hosted Spark jobs

These jobs use [self-hosted, dgx-spark, aarch64] and are not triggered by pull requests:

• .github/workflows/publish-image.yml: v* tags and workflow_dispatch
• .github/workflows/publish-train-image.yml: train-v* tags and workflow_dispatch
• .github/workflows/publish-whisper-image.yml: whisper-v* tags and workflow_dispatch

The lead can decide separately whether to migrate these release jobs.

Branch audit

To delete

Merged PR branches:

• codex/gui-truth-053 via fix(web,api): truthful instances everywhere — node names, stacked in Server view, live status, fleet-wide update banner #59
• fable/ainode-bench via Add ainode-bench and the report that renders bench/results #70
• fable/chat-benchchat via Make the built-in chat fleet-aware: instance picker, model card, measured turns #73
• fable/claude-effort-and-active-params via Harness bench: Claude Code effort option; catalog entries carry active params and arch (#127) #131
• fable/claude-harness-messages-proxy via Fleet proxy forwards the Anthropic Messages API; Claude Code becomes a harness bench adapter #123
• fable/engine-liveness via Bind wait: engine activity is the liveness signal, log silence is secondary (#112) #138
• fable/harness-bench via Harness bench: measure a model driving a coding agent to passing tests #110
• fable/interface-autodetect via Autodetect the cluster interface; fail cleanly on a host with no vLLM (#34, #61) #82
• fable/local-harness-placement via LOCKED CONTRACT: The harness bench records the node that serves the #140
• fable/local-harness-table-ornith via Render a Coding harness runs table in README.md from the harness bench records #132
• fable/model-card-overflow via Chat model card: stay inside the panel at any width #108
• fable/mp-distributed-launch via Distributed launch without Ray: vLLM mp multi-node shape, recipe-aware sharding, DeepSeek V4 Flash catalog entry (#84) #85
• fable/mp-local-models-flash-next via Distributed launch serves a model downloaded through AINode; Qwen3.8-Flash-Next catalog entry #120
• fable/opencode-adapter via Harness bench: make the OpenCode adapter reliable (#118) #137
• fable/replay-serialized via Startup replay: sweep first, then launch one engine at a time per node (#96) #99
• fable/routing-caps-parallel via Fleet proxy: route multimodal requests to an instance that accepts images; announce real parallelism (#83, #92) #98
• fable/tested-models via README: Models tested on AINode, generated from bench/results #69

Abandoned branch with no PR and no commit in 30 days:

• tools/benchchat, last commit 2026-09-04

No branches were deleted as part of this PR.

Open PR branches, rebase later

• fable/issue-221, Keep replay tests off the Docker boundary #290
• fable/issue-228, Keep unit tests out of the operator AINode home #291
• fable/issue-250, Count proxied inference tokens in node metrics #293
• fable/issue-262, Keep locked API access controls usable #292
• fable/issue-288, Fix System One score legend labels #294
• fable/nemotron-nano-4b-bench, Nemotron 3 Nano 4B bench records #297

Related repository audit

getainode/ainode-desktop has one flagged branch: main.