Without persisted models a rescheduled swarm task re-downloads the argos
packages on a fresh node, which exceeds the reschedule reachability
probe window and fails the chaos test. Build a custom image that bakes
the configured language pairs (services.libretranslate.load_only, joined
into LT_LOAD_ONLY and the install_models build arg from a single SPOT)
so every node starts warm. Proven live: the baked image serves HTTP 200
within ~20s with no runtime download. Pin the base to v1.9.6 instead of
latest.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>