chore: remove build-time model precache; enforce runtime-only downloads and ignore partial artifacts

feat: precache InsightFace buffalo_l analysis models at build; add cache listing to entrypoint

chore: prune LivePortrait artifacts from Dockerfile, remove deploy.sh

webrtc: verbose diagnostics (stats, event snapshots, /debug/webrtc_config, env vars)

hardening: unified model provisioning (sentinel, audit logs, removed force flags, /debug/models endpoint)

feat(models): entrypoint with deterministic model download + generator mandatory, debug hashing endpoint

CUDA fix: install cuda-nvrtc-11-8, set LD_LIBRARY_PATH; add ORT CUDA diagnostics

GPU-only: enforce CUDA EP when required; switch container to CUDA 11.8 + Torch cu118 to match ORT; add MIRAGE_REQUIRE_GPU=1

Upgrade to CUDA 12.1 + Torch cu121; bump ORT GPU to 1.18.1; use myn0908 generator_fix_grid (opset20); drop TensorRT plugin

Point generator to FasterLivePortrait warping_spade-fix.onnx; add GridSample plugin URL for compatibility

Switch generator to FasterLivePortrait opset<=19 to avoid opset 20 GridSample adapter issues

LivePortrait: auto-downgrade opset>19 to 19 for appearance/motion; force re-download flags; register ort-extensions; retry loaders

Use FasterLivePortrait appearance/motion (opset<=19) + myn0908 generator_fix_grid; avoid opset 20 validation error

Point LivePortrait model URLs to myn0908/Live-Portrait-ONNX (use generator_fix_grid.onnx)

Generator: switch to standard warping_spade.onnx; disable shape inference by default; force generator re-download; avoid TensorRT-dependent plugin by default

LivePortrait: use warping_spade-fix.onnx; add optional GridSample plugin; register custom ops and shape-inference toggle

Align Space image to CUDA 11.8 (Torch cu118, ORT GPU 1.17.3); add rotated-image fallback for set_reference

Improve set_reference reliability (multi-scale SCRFD, resize); add CUDA 12.1 runtime libs for ORT GPU provider

Fix generator URL to valid warping_spade.onnx; accept alt generator filenames in engine

Fix permissions for InsightFace and Matplotlib caches; set INSIGHTFACE_HOME/MPLCONFIGDIR; update default generator URL path

Neural-only pipeline: enforce SCRFD + LivePortrait ONNX (generator required); remove fallbacks/flags; fix LP engine is_ready; clean debug flags and Docker env

Models: default FasterLivePortrait ONNX URLs; add stitching.onnx support; expose MIRAGE_REQUIRE_NEURAL flag in /debug/models

Neural-only mode: require generator.onnx; disable fallbacks when MIRAGE_REQUIRE_NEURAL=1; Dockerfile env for generator URL

Fix MediaPipe dependency and enhance WebRTC debugging

Docker: set writable caches (/app/.cache) and model dirs; ENV HOME/HF_HOME/TRANSFORMERS_CACHE; Downloader: use HF_HOME/hub and avoid '/.cache' permission issues on Spaces

Default Docker ENV: enable MIRAGE_ENABLE_LIVEPORTRAIT=1 and MIRAGE_DOWNLOAD_MODELS=1 with default ONNX URLs

Ensure WebRTC router always mounts: lazy pipeline with pass-through fallback; add /webrtc/ping diagnostics in client; better 503 logging

Fix Space runtime error - remove MediaPipe dependency and build optimizations

Add landmark reenactor fallback and LivePortrait ONNX URLs

WebRTC: add /webrtc/connections; Metrics: integrate enhanced; Docker: enable SCRFD via env; Safe model loader wiring

Docker: add aiortc/librosa system libs; bind uvicorn+healthcheck to PORT; Space build stabilization

fix(docker): install wget and ca-certificates to support HF Dev Mode openvscode-server download

feat(docker): switch to Docker Space GPU runtime; prod WebRTC (aiortc) flow; remove legacy WS; token auth; instrumentation p50/p95; requirements harden

feat(docker-gpu): switch to CUDA-enabled PyTorch base and add torch deps

feat: phase1-2 baseline echo scaffold