- Add official GNU Affero General Public License v3.0 to uvr_cli/LICENSE
- Add SPDX-License-Identifier: AGPL-3.0-or-later and Copyright (C) 2026 RevLayDust headers to all uvr_cli modules
- Document multi-license architecture in root README.md and uvr_cli/README.md
- Add automated regression tests in tests/test_uvr_cli.py verifying license file and module SPDX headers
- Implement TensorRT-RTX backend via onnxruntime-ep-nv-tensorrt-rtx plugin in engine/tensorrt_rtx/
- Support in-memory FP16 downcast for MDX-Net ONNX models with zero disk writes
- Implement automatic fallback chain: TensorRT-RTX -> CUDA -> CPU
- Add GUI execution provider selection with dynamic hardware gating
- Add CLI provider flags and model resolver support for trt_rtx
- Add install/requirements_trt_rtx.txt and stack profile tracking
- Add test suite in tests/test_tensorrt_rtx.py and test_uvr_cli.py
- Update run_uvr.bat to execute UVR.py via native CMD
- Implement PEP 562 lazy loading in uvr_cli/__init__.py to reduce import time from 7.85s to 0.02s
- Defer heavy inference dependencies (torch, separate, onnxruntime, numpy) into CLI command handlers
- Reduce uvr_cli --help startup latency from ~8.98s to ~0.20s (~45x faster)
- Introduce InferenceResult tuple subclass in runner.py to isolate pure model execution timing
- Accurately measure and report pure inference elapsed time in uvr_cli run and benchmark engine
- Add unit tests for CLI startup backend isolation, PEP 562 exports, and InferenceResult structure
- Synchronize user-specified overlap, segment_size, and batch_size attributes onto HeadlessModelData
- Prioritize active overlap_mdx value when writing benchmark CSV rows and JSON summaries
- Support shorthand precision strings (fp32, fp16, bf16) in build_headless_model_data
- Add unit test coverage verifying custom bench parameters are captured in model data and CLI parsing