Run every media model in its own runtime

[data-angel]

Summary

Public main installs all 14 ComfyUI media recipes into one shared lloom-comfyui-media container. That hides the individual models from LLooM's admission, eviction and memory accounting. This PR gives every media model its own runtime.

• Per-model recipes (from codex/isolated-media-runtimes-20260924). Each of the 14 ComfyUI recipes gets its own runtime, container, data volume and port. Only that model's checkpoint, encoder, VAE and adapter files are bind-mounted, read-only. The bridge serves exactly the LLOOM_MEDIA_MODEL it was started with, and refuses an empty or unknown selection. Memory estimates are now per model; FLUX.2 klein, for example, drops from 95 GB to 20 GB.
• The shared runtime is removed.
  • build/launch.py refuses to start without LLOOM_MEDIA_MODEL and no longer defaults to the aggregate /opt/lloom-models root.
  • Re-applying a recipe on a host that still routes a model through a legacy shared comfyui-media runtime moves that model to its own runtime. The shared runtime and backend are dropped once no model references them. Before this, setup silently kept the shared route.
  • Recipes pin the rebuilt image (source-381d99b…).
• OpenRouter Lyria audio generation adapter (from codex/lyria-music-adapter, which was stacked on the isolation commit).

Test plan

• npm run test:unit (full chain, pass)
• node test/comfyui-media.test.mjs, including a new legacy shared → per-model migration case
• scripts/test-comfyui-media.py (209 pass, including the new launcher refusal test)
• node --test test/audio-providers.test.mjs
• node scripts/check-package.mjs
• Hosts: rebuild the image with lloom setup --recipe <id> --additive --apply, then remove the orphaned shared container

🤖 Generated with Claude Code