Add custom user-LoRA (Quality mode) for the MPS route

feat(ui): add ZeroGPU 2x-quota caution banner (Spaces only); drop dev banner from README

feat(zerogpu): get all modes working on ZeroGPU

fix(ui): replace track_tqdm with a manual step callback for clear progress labels

feat(mps): run locally on Apple Silicon MPS with a pre-inference OOM preflight

fix(fp8): bump diffusers (unpin) + torchao (latest) for a compatible fp8 stack

fix(fp8): use optimum-quanto (qfloat8) instead of torchao

fix(fp8): pin torchao==0.12.0 to match the pinned diffusers commit

feat(fp8): torchao Float8WeightOnly quantize the transformer at load

fix(zerogpu): raise duration to 145s + QIE_GPU_DURATION override

fix(zerogpu): lower duration to ~54s (108s req) — per-call ceiling is ~120s req

fix(zerogpu): scale duration with steps; materialization is ~15-40s, not 130s

fix(zerogpu): cap @spaces.GPU duration to 145s (xlarge 2x -> 290s, under the cap)

chore(ui): add active-development "do not run inference" banner (app + README)

fix(zerogpu): set duration to 180 and lower Quality default steps to 28

fix(zerogpu): cap queue concurrency and raise @spaces.GPU duration (Fast 120 / Quality 180)

chore: untrack .github/workflows/ci.yml (push token lacks workflow scope; retained locally)

fix: pass local gate and harden ZeroGPU path

feat: build Qwen Image Editor app (Edit/Compose tabs, Fast/Quality, ZeroGPU + local CUDA) with tests

chore: scaffold project — design spec, implementation plan, reference workflows