Pinned candidate reconstruction

This is an experimental recipe, not a promoted configuration. The image is a local immutable build, not a published image. Build Dockerfile.preparation against NVIDIA/flare-local-inference commit 47581ccac82660953a5b59cb04b2750d64edc97d through the managed image build surface. The exact resulting image ID is retained in serve-recipe.toml; a future rebuild must record its own identity rather than assume byte equality.

Install the reviewed seccomp.json at the operator-selected absolute path and update only that path in the recipe. Its hash is pinned. The profile adds only io_uring_setup, io_uring_enter and io_uring_register to the pinned Moby default at 4b8214fc4971e06172938a1f1554ccff6647ee80. Keep it immutable through launch. The published path is synthetic. Stage flare-config.toml in the private configuration volume. The explicit recipe --compile-workers 8 overrides its configuration-file value of 2.

Use managed models recipes load, status, logs, and unload with this registry. The tested source adds pinned inspection-image and seccomp support; these changes are not yet released. GPU ownership and host admission must pass. The private operator packet retains exact selected GPU identities and commands.

Named model, kernel/package and private configuration volumes survive managed unload. The private volume contains a local package-signing key; never publish or copy that key into evidence. Completed compiler caches are reusable with this pinned runtime/hardware. Package export succeeded. The retained package index hash is in package identity; it excludes external weight blobs. A cached native restart took21.555seconds and passed7/7 preflight. 32,768 is the tested workload budget, not a Flare runtime context limit. TP2/C4/MTP3 and 2 GiB reserve are configured values. No supported vision replacement or MTP speedup is established by this recipe.