AINode 0.5.4

Added

• Models launch with their own engine image and flags (#62) — per-instance
  extra_vllm_args and engine_image on the node config, threaded through the
  per-load override set so they persist and survive restart-replay. A model whose
  published recipe needs flags AINode doesn't model (speculative decoding,
  MoE/mamba backends, reasoning + tool-call parsers) now launches through the
  normal load path instead of a hand-rolled container. A flag supplied by the
  caller suppresses the matching built-in rather than duplicating it.
• Catalog recipes for NVIDIA Nemotron 3.5 Lightning 30B-A3B (NVFP4) and
  Qwen3.8-27B (NVFP4, native vision), each carrying its proven engine image,
  flag set, and recommended memory fraction — applied as defaults so a bare
  {"model": ...} load (what the dashboard sends) launches correctly.

Fixed

• The 0.17-era GB10 workarounds (--enforce-eager, the NVFP4 MARLIN env) now
  apply only to the pinned default engine image. They are bugs in that build;
  on newer engines they only disable CUDA graphs and cost throughput.
• vllm serve argv is normalized across engine images. vllm/vllm-openai bakes
  ENTRYPOINT ["vllm","serve"] while the default image uses NVIDIA's passthrough
  shim, so emitting our own prefix produced vllm serve vllm serve <model> and
  the engine exited with "unrecognized arguments". The entrypoint is deliberately
  not overridden — that would bypass CUDA setup.
• Engine containers no longer run with --rm, so an engine that dies during
  startup leaves a readable corpse instead of erasing itself. start_solo() now
  confirms the container reached Running and logs the engine's last output on
  failure, rather than reporting success as soon as the docker CLI forked.
• Eject survives reboot. It was memory-only, so startup replay resurrected
  ejected models; it now rewrites the instance manifest and clears the node's
  model claim when the ejected instance was the primary.
• The boot path no longer launches the legacy host-venv engine when vLLM isn't
  importable — it uses the configured container backend instead of starting a
  guaranteed "No module named 'vllm'" failure behind an "Engine starting" banner.
• Vision models served from the HF cache now get --kv-cache-dtype auto
  explicitly; the automatic fp8 downgrade only fires for models on local disk,
  and fp8 KV corrupts VLM generation on GB10.

Changed

• Working branches use the fable/* prefix (was codex/*).
• The ruff rule set is pinned explicitly so lint is deterministic across ruff
  upgrades; an unpinned ruff>=0.1.0 had started failing every PR on rules the
  repo never opted into.

---