Production overhaul: pre-trained AdaptFormer model + detection quality improvements

Add Siamese U-Net training notebook and model inference integration

Clean visualization: subtle tint overlay, outline-only boxes, 30-region cap, high-contrast labels

Redesign detection pipeline: CVA, vegetation/shadow suppression, SNR fusion, multi-scale SSIM

Thinner, more precise detection lines and compact labels

AI/Hybrid overhaul: bidirectional AI fusion, stronger hybrid sensitivity, keep large unclassified regions

AI rework: percentile+blur threshold + less aggressive cleanup; Hybrid lower threshold

Fix: fallback to Image Difference when AI/Hybrid mask is too sparse

Detection fix: ECC registration fallback + less aggressive large-region rejection

Detection tuning controls + threshold diagnostics in API response

Full codebase audit: fix critical perf bug, security, error handling, dead code

Improve detection accuracy: border exclusion, threshold floors, fill-ratio filter, NMS, tighter bboxes

History table view, zoom, severity markers, row hover highlight, before/after thumbnails

Add vegetation and structural sub-classification for detected changes

Add 3D building analysis: shadow-based height/stories estimation and construction stage classification

Code review fixes: SSIM stability, symmetric CLAHE, confidence scoring, KMeans perf, auth hardening, remove debug endpoint

Initial commit: SatDetect satellite change detection web app