Pinned dependencies

The .in files list direct dependencies with their minimum supported versions. The .lock files pin every package to an exact version with hashes (the versions listed in the thesis software table, Appendix B, Table B.1), per role and platform, and are what gets installed:

Lock file	Role	Platform
edge_server.lock	edge server (NNPM, SSH dispatch, aggregation, optimiser, model profiling)	x86_64, Python 3.11
edge_device.lock	edge device, DNN experiments (TensorFlow)	aarch64, Python 3.11
edge_device_llm.lock	edge device, LLM experiments (PyTorch, Transformers)	aarch64, Python 3.11
dev.lock	development and continuous integration	x86_64, Python 3.11
docs.lock	the documentation site (MkDocs) and the screenshot tooling (Playwright)	x86_64, Python 3.11

Regenerate with scripts/lock_requirements.sh (uses uv pip compile --generate-hashes). Install with uv pip install --require-hashes -r requirements/<role>.lock or pip install --require-hashes -r .... The edge-server lock pins the x86_64 CPU build of PyTorch (torch==2.1.2+cpu), which is on the PyTorch index: add --extra-index-url https://download.pytorch.org/whl/cpu (and, with uv, --index-strategy unsafe-best-match), or install from the local mirror with --no-index --find-links.

Transformers is pinned to 4.46.3 because the Qwen2 (4.37), Llama 3.2 (4.43) and HeadSplit (4.40) adapters need it; the Chapter 6 experiments of the thesis used 4.36, and the testbed's code accepts both the 4.36 and the 4.38+ interfaces.

The optional INA260 power-monitor packages (power.in) depend on the Raspberry Pi platform (adafruit-blinka) and are installed on the devices only; they are listed without a lock file because their platform wheels cannot be resolved off the device.