[PWJCSqiushan]

What changed

• Integrated Fursee subject intelligence into FurColor Studio 4.0 with isolated optional dependencies and verified model assets.
• Added fursuit detection, DINOv3 embeddings, anonymous per-project subject clustering, subject-level keep/reject controls, and conservative fusion with YuNet and face-memory evidence.
• Added local subject exposure correction with highlight protection, subject preview crops, new API/service/UI flows, and supporting documentation.
• Hardened project-path isolation, cleanup behavior, demo isolation, BOM-compatible selection loading, dependency setup, licensing, and third-party notices.
• Preserved the Windows CI definition as docs/ci/tests.yml.example. It is intentionally inactive because the current HTTPS publishing credential cannot create files under .github/workflows; copy it back after a credential with GitHub''s workflow scope is available.

Why

FurColor Studio previously lacked reliable fursuit-aware subject understanding. This version uses the released Fursee detector and embedding model to improve subject-aware privacy review, selection, grouping, and exposure adjustment while keeping model files, photos, feature vectors, and private project data outside Git.

Compatibility notes

• The Fursee DINOv3 checkpoint uses development-era key names. Loading now validates exact key sets and tensor shapes before a deterministic remap and strict load.
• The matching DINOv3ViTImageProcessorFast path is used for Transformers 4.56.0 compatibility.
• YOLO may use FP16 on CUDA; the DINOv3 embedding path remains FP32 to match the upstream feature projection.

Validation

• 20 passed with the base project virtual environment.
• Publication security audit passed: no photos, credentials, private paths, model weights, or instance identifiers were detected.
• Full model SHA-256 verification passed.
• RTX 5060 Laptop GPU model smoke test passed.
• Three-image end-to-end subject job completed with detections, 512-dimensional embeddings, clustering, and preview crops.
• Same-crop embedding comparison against Fursee''s packaged feature database reached cosine similarity 0.9999997616.
• Python syntax, JavaScript syntax, JSON parsing, PowerShell parsing, dependency checks, API smoke tests, and git diff --check passed.