fix(model-artifacts): gate inferred artifact materialization by backend

[localai-bot]

Problem

Follow-up to #10909 (single-file snapshot resolution). That PR fixed the case where an explicitly-resolved single-file snapshot was handed to a backend as a directory. This PR closes the remaining gap: the loader still infers a managed artifact from any HuggingFace-shaped parameters.model regardless of backend.

PrimaryArtifactSpec materializes a bare HF reference (huggingface://owner/repo/model.gguf, hf://…, or a bare owner/repo) into a snapshot directory .artifacts/huggingface/<key>/snapshot/. That directory is the correct load target for directory-consuming backends (transformers, vLLM, diffusers, …), but a single-file backend such as llama.cpp or whisper is then handed the directory instead of the weight file:

gguf_init_from_reader: failed to read magic
llama_model_load: error loading model: llama_model_loader: failed to load model from
  /models/.artifacts/huggingface/<key>/snapshot

The /import-model importer already guards this with a backend allow-list (managedArtifactBackends), but the loader-side inference in core/config never applied the same gate. So a hand-written YAML or local-ai run huggingface://…/x.gguf with a single-file backend and no files: would break — the exact class of failure that motivated #10909, just via the inference path rather than an explicit artifacts: block.

Fix

• Move the backend allow-list from core/gallery/importers/helpers.go into core/config as an exported predicate IsManagedArtifactBackend(backend).
• PrimaryArtifactSpec now gates the inferred cases: only directory-consuming backends may have an artifact inferred from a bare reference. Every other backend stays on the legacy download-to-file path, so the backend receives the weight file itself.
• An explicit artifacts: block still bypasses the gate — single-file snapshot resolution (fix(model-artifacts): load single-file HF snapshots from the file, not the directory #10909) handles the load path there.
• The importer now consumes the shared config.IsManagedArtifactBackend, so both paths agree on which backends auto-materialize (no duplicate list).

Tests

• New core/config/model_artifact_inference_test.go (Ginkgo): a bare HF file/repo reference on llama-cpp stays legacy (managed=false); a reference on transformers still infers a managed artifact; an explicit artifacts: block on llama-cpp stays managed.
• Written test-first (2 of 4 cases failed before the change, then green).

Verified locally: go test ./core/config/ ./core/gallery/importers/ ./core/services/galleryop/ ./core/application/ pass; make lint (new-from-merge-base) reports 0 issues.

Assisted-by: Claude:opus-4.8 [Claude Code]

🤖 Generated with Claude Code