Summary

Adds support for MOOZY, a slide- and patient-level foundation model built on the Lunit ViT-S/8 patch encoder.

• lunit — new tile encoder (ViT-S/8 DINO, 384-dim, 0.5 µm)
• moozy-slide — standalone slide encoder from the MOOZY stage-2 checkpoint (768-dim)
• moozy — patient-level encoder: Lunit → MOOZY slide encoder → CaseAggregator transformer (768-dim)

New concepts

Patient-level encoders (PatientEncoder base class) aggregate slide embeddings across multiple slides per patient. They require a patient_id column in the input manifest CSV.

New Python API: Model.embed_patient(slides, patient_id=...) and Model.embed_patients(slides) return EmbeddedPatient with patient_embedding and per-slide slide_embeddings.

New config option: save_slide_embeddings: true saves intermediate per-slide MOOZY embeddings to slide_embeddings/ when running the patient pipeline. Patient embeddings are written to patient_embeddings/.

Files changed

• encoders/base.py — PatientEncoder base class
• encoders/registry.py — extended level validation to "patient"
• encoders/models/lunit.py, moozy.py — new encoder registrations
• artifacts.py — PatientEmbeddingArtifact, write_patient_embeddings()
• api.py — EmbeddedPatient, Model.embed_patient/embed_patients, ExecutionOptions.save_slide_embeddings, RunResult.patient_artifacts
• inference.py — embed_patients(), _run_patient_pipeline(), _encode_slide_from_tiles() helper
• utils/tiling_io.py — load_patient_id_mapping()