v0.6.1

Added

• -DPROFILING=ON: frame pointers and debug info on every target, for perf/Valgrind call graphs.
• CPU decode-path benchmarks (tests/benchmark/bench_cpu_pipeline.cpp).
• GPU gate on Google Colab (scripts/colab/gpu_gate.ipynb) and headless --display check (scripts/check_display.sh).
• DALI_SOURCE=pip ./scripts/fetch_dali.sh stages DALI without Docker.
• run_gate.sh: CUDA_ARCH defaults to the card's compute capability; SKIP_BUILD_MATRIX=1.

Changed

• clang-tidy is enforced (WarningsAsErrors: '*'); all findings fixed or suppressed per line.
• CLI parsing moved to rfdetr::cli::parse_cli (src/cli_options.*) with 47 unit tests; main() split up.
• std::endl replaced with '\n': stdout is no longer flushed per line when piped.

Fixed

• run_gate.sh failed engine builds on a fresh machine (TensorRT libs not yet downloaded).
• run_gate.sh reported a missing DALI as FAIL instead of UNRUN.
• Documented clang-tidy command now excludes tensorrt_backend.cpp, as CI does.

Known issues

• ONNX Runtime runs single-threaded (SetIntraOpNumThreads(1) is hardcoded).
• GPU-decoded JPEGs (nvJPEG) do not bit-match the CPU tensor (stb); within GPU-preprocess tolerances.
• --display unverified on the DALI build and on a real (non-Xvfb) screen.