Remove PyTorch dependencies

[adamtheturtle]

Summary

Replace PyTorch and torchmetrics with lighter-weight alternatives:

• Replaced torchmetrics.image.StructuralSimilarityIndexMeasure with skimage.metrics.structural_similarity
• Replaced piq BRISQUE implementation with brisque package wrapper
• Removed torch, torchvision, torchmetrics, piq from dependencies (~3.5GB of binaries removed)
• Added brisque, scikit-image, scipy as replacements

Changes

• image_matchers.py: Use scikit-image for SSIM calculation instead of torch
• target_raters.py: Use brisque package wrapper instead of piq/torch
• pyproject.toml: Update dependencies and remove PyTorch index configuration
• _brisque.py: New wrapper module around brisque package

🤖 Generated with Claude Code

---

Note

Medium Risk
Replaces SSIM/BRISQUE implementations and swaps major dependencies, which may slightly change image matching/target rating behavior and introduces native build requirements in Docker.

Overview
Removes the PyTorch-based image-quality stack by dropping torch, torchvision, torchmetrics, and piq, and adding brisque, scikit-image, and scipy (plus deleting the custom uv PyTorch CPU index config).

Updates StructuralSimilarityMatcher to compute SSIM via skimage.metrics.structural_similarity on normalized NumPy arrays, and rewires target tracking ratings to use a new mock_vws._brisque.brisque_score wrapper around the brisque package.

Adjusts the Flask server Docker image to install g++ so native extensions needed by the new dependency chain can compile, and updates lint/dep tooling config and spelling dictionary accordingly.

^{Written by Cursor Bugbot for commit 259c63b. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

SSIM fails on grayscale images missing channel dimension

High Severity

The structural_similarity call unconditionally passes channel_axis=2, but grayscale ("L" mode) images produce 2D numpy arrays with shape (H, W) — no axis 2 exists. The image validator explicitly allows grayscale images, so they can reach this code path. The old torch-based code handled this correctly via getbands() and .view(). The _brisque.py wrapper avoids this by calling .convert(mode="RGB"), but image_matchers.py does not convert images before creating numpy arrays.