Major Changes

• Python 3.10+ minimum (previously 3.9)
• NumPy ≥ 2.2.0 required
• LiTERT-Torch exporter replaces the old TFLite export path
• TFLite backend fixes: XNNPACK dynamic reshape fallback, dimension mismatch fixes, flatten(1) in backbones for dynamic shape compatibility

New Features

• GTA (Global Track Association) postprocessing — pure-appearance recovery pass for re-identifying dead tracks by embedding similarity, with GP-smoothed gap interpolation
• ReID training pipeline improvements (feat/reid-training PR [#2300])
• SAM2MOT and OccluBoost tracker configs added
• tune_analysis module for hyperparameter tuning results
• Dataset configs, cache improvements, and CLI enhancements — optional pipeline steps, new modes, SportsMOT support
• Part-3 ViT model and resume validation metrics

Refactors

• Tracker directory reorganization into bbox/, hybrid/, mask/ subcategories
• Config directory flattening — removed benchmarks/ dir, flattened detectors/ subdirs, merged mot17-ablation into mot17 with a splits: mechanism
• Engine rework — cache, replay, tracker, tuner, CLI updates
• Native C++ replay backend updates with C++/Python parity

Performance

• OBB IoU vectorization — batch rect pre-computation (~3.5× faster)
• Frame-indexed detection lookup — O(1) instead of O(N) per frame
• Image I/O skipping after first frame during tracking replay
• Parallel TrackEval using NUM_THREADS

Fixes

• KF per-class class-ID alignment
• OBB KF tuning injection
• KF velocity noise derived from GT accelerations instead of fixed 0.01 ratio
• CI fixes (Python 3.9 → 3.10 matrices, test datasets switched to mot17-mini)
• TUI blank-space cleanup, GTA warning fix
• iou_batch_obb crash on empty inputs