feat(isr): create async assessor loop with verdict merging

feat(isr): create ISRAssessor with GPT-4o-mini vision integration

fix: weight download race condition + rewrite CLAUDE.md

feat: replace drone_yolo with YOLOv8-VisDrone detector

feat: add YOLOv8-VisDrone detector for aerial object detection

Eliminate redundant JPEG frame loading via shared frame store

feat(ttfs): instrument segmentation time-to-first-stream

fix: download yolo11m.pt via huggingface_hub to writable /tmp path

fix: use YOLO("yolo11m.pt") with writable config dir

refactor: rename hf_yolov8 → yolo11 across codebase

feat: replace custom YOLOv8m_defence with YOLO11m COCO-pretrained

feat(segmentation): add YSAM2 variants and mission-aware detector mapping

chore: remove dead code and unused frontend modules

refactor(gsam2): make SAM2 detector-agnostic

fix: SAM2 _prep_prompts() signature

perf: defer GPU->CPU mask transfer to render thread via LazyFrameObjects

fix: correct benchmark metrics — remove double-counting and add missing timers

perf: pipeline GSAM2 tracking + rendering with startup buffer

feat: add num_maskmem parameter to /benchmark endpoint

revert: remove torch.compile — runtime failures on HF Space

fix: drop fullgraph=True from torch.compile for broader compatibility

fix: add runtime fallback for torch.compile inductor/triton failures

fix: manual torch.compile with graceful fallback for SAM2 components

perf: enable torch.compile for SAM2 via vos_optimized flag

perf: GPU-resident tensor pipeline for SAM2 video propagation

perf: eliminate CUDA sync points in SAM2 video propagation hot-path

feat: add detailed timing metrics to GroundedSAM2 segmenter

Fix segmentation mission routing and first-frame GPT alignment

Fix cross-GPU mask device mismatch in GSAM2 reconciliation

Harden GSAM2 parallel pipeline and tracking reconciliation

Refine Grounded-SAM2 tracking behavior and video frame handling

Remove SAM3 and standardize segmentation on Grounded-SAM2

Handle set-typed missing_keys in Grounding DINO loader

Fix Grounding DINO load reliability and writer error handling

fix: replaced batch with sequential

Fix: Grounding DINO batch inference

Add diagnostic logging to pinpoint tracker dropping detections

Fix Grounding DINO detections dropped by ByteTracker threshold mismatch

Fix GroundingDino loading: manually materialize meta tensors to support tied weights

Refine GPT enrichment logic, add relevance badges to UI, and fix minor model/logging issues

feat: Radar historic traces and fix DETR crash

Use DETR no_timm revision to avoid timm dependency

Fix BGR to RGB conversion for DETR and GroundingDino inference

fix: Resolve IndentationError in DroneYoloDetector due to displaced import

feat: Implement SAHI Tiling for 4K video detection

optimize: Force high-res (1280px) inference for YOLO models to improve small object detection

perf: Tune batch sizes and queues for A10 GPUs

Simplify drone YOLO loading with ultralytics native HF Hub support

Update Lite model to use spencercdz/YOLOv8m_defence

Remove InternVL2 logic as requested