fix(docker): use a CPU-only PyTorch image

[rksharma-owg]

Summary

• install the CPU-only PyTorch wheel instead of CUDA runtime dependencies
• rely on the separate Ollama service rather than starting a nonexistent embedded daemon
• stabilize Ruff linting across releases
• add packaging regression tests

Fixes #6.

Validation

• native arm64 container build passes
• image size falls from the published 6.11 GB manifest to 776.6 MB locally (87.3% smaller)
• 204 tests pass and 65 are skipped
• installed PyTorch reports no CUDA support
• the exact CPython 3.11 x86_64 CPU wheel resolves successfully; local x86 execution was unavailable because this Docker installation lacks amd64 emulation
• Ruff, Compose validation, and actionlint pass

AI assistance

OpenAI Codex assisted with the audit, implementation, tests, and PR description. I reviewed the complete diff and validation results.

[vijayleo31]

I pulled PR #17's branch and independently validated the Dockerfile changes locally (macOS, Apple Silicon, Docker Desktop, Python 3.11.15).

Build result:

docker build -t pwnzzai-local:dev .

Build completed successfully (12/12 steps, no errors). Resulting image content size: 954MB, down from the previously published 6.11GB manifest (~84% smaller) — consistent with the PR's reported 87.3% reduction.

Regression tests:

tests/unit/test_dockerfile.py::test_docker_image_installs_cpu_only_torch_before_app_requirements PASSED
tests/unit/test_dockerfile.py::test_app_image_does_not_start_a_second_ollama_server PASSED
2 passed in 0.01s

What the fix does:

• Installs the CPU-only PyTorch wheel (torch==2.7.1 from https://download.pytorch.org/whl/cpu) before requirements.txt, avoiding the CUDA runtime dependencies that were bloating the image.
• CMD now runs only flask run; the redundant ollama serve process previously started inside the app container is removed.

Caveat: I validated the build and image size locally but haven't reproduced the original repro steps (pulling a published image on Windows) to directly confirm the EOF-on-pull error is resolved. Given the ~84% size reduction, this is very likely to fix it, but a Windows pull test against a registry-published build from this branch would fully close the loop.
<img width="1256" height="170" alt="Screenshot 2026-07-28 at 7 01 00 PM" src="https://github.com/user-attachments/assets/42f

cba1d-2b6b-4c53-a8e7-1607c1e02297" />