Build reproducibility: pin pytorch_wavelets, recheck the PyTorch 2.1.2 pin, lock dependencies, fix Docker user handling

[joeljose]

Severity: Medium.

Problems

1. An unpinned git dependency. Dockerfile.gpu runs pip install git+https://github.com/fbcotter/pytorch_wavelets.git, which installs whatever is on the default branch at build time. That is a reproducibility risk and a supply-chain risk.
2. The PyTorch pin may no longer be needed. The Dockerfile comment pins pytorch/pytorch:2.1.2-cuda12.1 because pytorch_wavelets "uses old-style autograd.Function". In this audit, current pytorch_wavelets HEAD imported and ran extract_audio_gpu correctly on torch 2.14 (CPU). So the pin (a 2023 base image) may be removable. Please confirm on CUDA.
3. Inconsistent requirements. requirements-gpu.txt pins numpy<2, while requirements.txt allows <3. PyWavelets has no upper bound. There's no lock file. The CPU image uses opencv-python (needs libgl1/libglib2.0-0), while the GPU image uses the headless build.
4. The builder's UID is baked into the image. Both Dockerfiles need UID/GID/UNAME build args: a plain docker build . fails, building as root fails, and the images can't be published for other users.
5. Base image and workflow. python:3.11-slim isn't pinned to a patch release. The workflow has no permissions: block, and actions aren't SHA-pinned.

Suggested fix

• Install pytorch_wavelets @ git+…@<commit-sha>, or vendor the DTCWT forward pass: only DTCWTForward is used, about 300 lines under MIT licence.
• Test a current PyTorch base (2.4+), and drop the 2.1.2 pin and numpy<2 if they pass.
• Keep version ranges in pyproject.toml, generate lock files with uv pip compile, and install from them in Docker and CI.
• Use opencv-python-headless everywhere.
• Use a fixed non-root user and document docker run --user "$(id -u):$(id -g)".
• Add permissions: contents: read, pin actions by SHA, and enable Dependabot.

Acceptance criteria

• No unpinned VCS dependencies remain.
• docker build . works with no build args.
• Builds are reproducible from the lock files.