[bmachek]

Why

The nightly Golden tests (all model assets) run failed (job):

auraface_matches_insightface_reference ... FAILED
LRG_REQUIRE_GOLDENS demands the `auraface` golden assets, but they are missing:
needs …/yunet_face_detection.onnx and …/auraface_glintr100.onnx

#460 added the auraface golden family and committed its reference embedding to face_goldens.json, so LRG_REQUIRE_GOLDENS=all now requires the AuraFace model. golden-tests-full.yml never downloaded it. (YuNet is present; the message lists both files because the test checks for both.)

What

• golden-tests-full.yml: download glintr100.onnx from Hugging Face as auraface_glintr100.onnx. AuraFace isn't a release asset, and the server fetches it the same way (routes/face_recognizer.rs). The step takes repo, revision, file and sha256 from auraface_source in face_goldens.json, which records the exact file the reference came from. It checks the hash with sha256sum -c, so a moved or changed upstream file fails the job instead of producing a confusing numeric mismatch.
• Docs: added auraface to the LRG_REQUIRE_GOLDENS family names in CLAUDE.md, AGENTS.md, CONTRIBUTING.md, server-rs/README.md and docs/wiki/Dev-Server-README.md.

PR CI (server-rs-tests.yml, LRG_REQUIRE_GOLDENS=face) is unchanged. auraface is a separate family, so PRs still skip it.

Testing

• Workflow YAML parses. The jq extraction gives fal/AuraFace-v1 af6d057c… glintr100.onnx a7933ea5….
• scripts/check-docs.py passes.
• Not tested: the download itself. huggingface.co is blocked from the environment this was written in. Running the workflow by hand (workflow_dispatch) on this branch will check it end to end.

---

Generated by Claude Code