fix: keep classified items when a run is cancelled

feat: hybrid ML allowlist rescue on WD needs-review

Enhance media preview functionality in backend API. Introduce caching for JPEG stills from GIFs and videos to improve performance under load. Update preview endpoint to support raw media serving and adjust response handling for inline display. Modify frontend to reflect changes in media preview handling and improve UI for video indicators. Add tests to ensure proper functionality for GIF and MP4 previews.

Fix GIF/video preview and classify with length-scaled multi-frame sampling.

Speed up WD inference with real batching and harden reclassify cancel/retry.

Enhance run management in backend API to support cancellation and directory validation. Implement checks for root and categories directories, ensuring proper error handling for missing or invalid paths. Update run status handling to allow immediate cancellation during execution. Add tests to validate new behavior for run initiation and directory checks.

Add multi-model tagging and SHUCK3R-style classifier UX.

Enhance README and backend API for GPU inference support. Updated README to include GPU visibility verification, runtime controls, and examples for inference settings. Modified backend API to add new endpoints for provider health checks and improved inference handling with batch processing capabilities. Updated database schema to include scan_recursive setting and enhanced telemetry tracking for inference runs. Improved tests to validate new API responses and functionality.

Enhance backend API and frontend components to support full scores for classified items. Updated the database schema to include a new column for full scores, modified the item retrieval logic to optionally include these scores, and added a new endpoint for fetching item scores. Updated the frontend to display score debug information and improved state management for score loading. Enhanced tests to cover new functionality.

Enhance backend API with run status management and logging. Added functionality to track and update the status of image processing runs, including cancellation and progress reporting. Improved error handling and logging throughout the application. Updated database schema to support new run lifecycle states and added corresponding response models. Enhanced frontend API interactions to include run status polling and management.