fix: handle PrettyMIDI object return from app.transcribe() by writing directly

fix: pin keras==2.12.0 to match TF version and support SavedModel checkpoints

fix: use protobuf==4.25.9 which has runtime_version and is TF 2.12 compatible

fix: clear PIP_CONSTRAINT for pyharp install to avoid descript-audiotools conflict

fix: use PIP_CONSTRAINT to globally lock protobuf==3.20.3

debug: pip show protobuf to see what version is actually installed

fix: force-reinstall protobuf==3.20.3 after omnizart to fix runtime_version error

fix: use omnizart Python API directly instead of CLI subprocess to avoid SIGSEGV

fix: pin tensorflow-cpu==2.12.0 to avoid SIGSEGV on CPU without AVX512

trigger restart

debug: don't capture subprocess output to see real omnizart error

fix: check output files instead of return code (TF warnings cause false exit 1)

fix: add file_types to gr.File for pyharp compatibility

fix: add portaudio19-dev for pyaudio build

fix: pre-install Cython+numpy with PIP_NO_BUILD_ISOLATION for omnizart

fix: valid colorTo value

initial Omnizart HARP deployment