[Housekeeping] Security hardening + repo hygiene cleanup

[inii-sayaa]

Labels: security, chore, documentation

Items

1. qml_app/models.py (save/load, lines ~118, ~124): hyperparameters are stored
   with raw pickle.dumps/pickle.loads. If this .pkl file is ever shared or
   downloaded from an untrusted source, pickle.loads can execute arbitrary code.
   Suggest switching to json for this simple hyperparameter dict.

2. qml_app/main.py (line 132): torch.load(...) doesn't pass weights_only=True
   explicitly. This is safe by default on torch>=2.6 (this repo pins torch==2.9.0),
   but being explicit protects against future default changes.

3. README.md lists Qiskit as part of the tech stack (badge + description), but
   there is no import qiskit anywhere in the codebase, and qiskit is not in
   requirements.txt. Either add the dependency/usage or update the README.

4. 5 .DS_Store files are committed to the repo (.DS_Store, artifacts/.DS_Store,
   artifacts/kernel/.DS_Store, artifacts/vqc/.DS_Store, qml_app/.DS_Store).
   Suggest git rm --cached them and add .DS_Store to .gitignore.

5. .gitignore excludes artifacts/, but artifacts/convert.py, artifacts/script.py,
   artifacts/vqc/convert_pt_to_json.py, and the .DS_Store files under artifacts/
   are already tracked (added before the ignore rule existed). This is confusing
   for contributors — worth clarifying intent (untrack them, or move the helper
   scripts outside artifacts/).

6. No automated tests exist in the repo. Even small unit tests around
   _align_with_qubits (data.py) and the kernel circuit output shape would have
   caught issue [Bug] Best-checkpoint saved during VQC training is not actually the best epoch's weights #2 above much earlier.

Severity

Low-to-medium - none of these block correctness, but #1 is worth prioritizing
if model files are ever distributed publicly.