- add _output_dir_for(): output/<relpath_without_ext>\n- use it in sequential, staged, LRC and embed pipelines\n- keep track folders unique for equal stems in different albums\n- log the layout and report folder collisions\n- add a safe one-shot flat-to-path migration script
- detect unrecoverable CUDA errors before OOM retries and clear model caches\n- keep normal OOM retries to two attempts (configured batch, then 1)\n- do not create text.txt when ASR returns only whitespace\n- warn per file and report empty_result / lrc+empty in the summary\n- keep at least 12 log lines visible in the GUI
- keep WhisperX, align and GigaAM caches process-wide and thread-safe\n- preload the ASR model at the start of the staged stage 3\n- log that the model is loaded once and then reused\n- never clear caches between files, batches or OOM retries
- retry only lowers batch_size: configured, 1, 1\n- keep vad_filter parameters for future model-load-level control\n- do not log success in staged finalize when no text was produced
- add _final_artifact_present and use it for resume decisions\n- text.txt no longer marks a track done when separation is the last stage\n- apply the same rule to pre-scan, staged seeding and sequential skipping\n- keep identification-only runs resolving through text.txt\n- preserve existing metadata tags when a stage-only rerun has less info
- pre-scan all inputs before chunking so ready files stay out of batches\n- start overall progress at the already-finished count\n- count skipped/resumed files correctly in sequential and staged runs\n- keep batches full and log ready/pending counts before starting\n- avoid importing Torch at startup or in cleanup for separation-only runs\n- move TF32 setup into lazy WhisperX model loading\n- detect the GPU through nvidia-smi for the GUI status line
- replace permanent _CUDA_BROKEN flag with CudaOOMError and retry helper
- retry WhisperX with configured settings, then batch=1 without VAD, then once more
- keep the loaded model cached between retries instead of reloading it
- actually pass vad_filter through transcribe() and use it in aligned ASR
- skip only the OOM file in sequential, staged and LRC pipelines
- mark OOM files with oom_skipped / lrc+oom for the summary
- fix staged vocal separation writing input_dir/vocals.wav when convert is off
- route original inputs and resumed files through safe per-file separation
- validate ffmpeg before starting the LRC task
- make the main page scrollable and keep the console reachable on small screens
- add fullscreen toggle (button + F11, Escape to exit)
- add enable switches for LRC generation/embedding and skip ready text.lrc
- sanitize multi-line ASR text before writing LRC timestamps
- emit LRC/embed progress once per file instead of twice
- increase progress-bar contrast
- store source track path in metadata.json and validate it when skipping ready files
- add skip_existing/resume_partial options to config and GUI
- identify lyrics via embedded text, Shazam, then one LRCLIB author search
- keep post-ASR LRCLIB text matching and allow cancelling safely
- recover from CUDA OOM and free VRAM between batch stages
- add overall completed/total file progress bar to GUI and pipelines
