feat(tagger-ui): display smart analysis tags in tagger panel

feat: add anatomy, expression, and background analyzers

fix: pose tag naming + swap DWPose detector to YOLOv8m-pose for anime

refactor(ui): merge pose toggles into single checkbox + method radio (YOLO / Whole-body DWPose)

feat(tagger): DWPose whole-body pose (133 keypoints) with YOLOX detector, hand gestures, facial expressions; pose_mode toggle in UI

feat(variator): smart tag ordering, per-category token budget, semantic min-tag topup, familiarity + scene constraints, intent-aware negatives, suggestion engine

fix(pose): make standing/walking/running mutually exclusive; raise walking threshold to 1.5x shoulder-width; update tests

fix(pose): letterbox preprocessing for reliable anime detection; primary-person selection; quality gate; yolov8m-pose @ 1024 default

fix(data): migrate 41 SFW tags mis-rated as nsfw into hair/body/accessory pools; tag index first-pool-wins so cross-listed SFW tags keep their safe section

feat(tagger): NL-only caption mode (no WD14/pose passes); yolov8s-pose default with original-coord keypoints; ControlNet Preview panel (skeleton black/overlay, Canny, PNG download); caption cache in Captioner facade

fix(tagger): filter character tags from general to prevent duplicates; auto-select Qwen model/dtype by device (CPU→2B fp32, GPU→3B bf16)

feat(caption): Qwen2.5-VL-3B captioner option (Anima-style NL captions); per-call kind dispatch in Captioner facade, kind in tagger cache key, transformers>=4.53

feat(tagger): weighted 3x WD14 + JoyTag smart ensemble, richer pose analysis, NL caption tied to Smart mode

fix(caption): silence early_stopping single-beam warning for Florence

fix(caption): Florence-2 with transformers<5, use_cache=False for 4.54+ cache format, dict post-process

feat(tagger): smart ensemble — WD14 EVA02-Large v3 + JoyTag fusion, YOLO pose tags, optional VLM captions

fix(moondream): syntax + accelerate-free .to(cpu) load

feat(tagger): Moondream2 NL captioner with sidebar, pose toggle on UI, prompt-caption join

fix(tagger): Moondream2 NL captions, tagger_mode names parity, remove Git LFS weights from cache

feat(tagger): swap Qwen-VL for Moondream2 NL captioner

fix(i18n): qwen_desc parity keys; test_identifiers cleanup; tagger_mode_vit removed from drop-down (--mouseover duplication bug)

feat(tagger): NL captioning via Qwen3-VL, pose detection UI fixes, memory-safe for HF Spaces (16GB CPU)

feat(tagger): pose/Qwen safety tie-offs, CPU-first model loading, disable Qwen by default

fix(tagger): CPU-only safety, memory budget, pose toggle, fold single-model modes into slide-panel

fix(tagger): optional pose/Qwen, input validation, memory-safe HF Space runtime

fix(tagger): push lost qwen model, slide-panel UI, SSR off, ensemble wiring

feat(tagger): Qwen-VL NL captions, pose extraction (YOLOv8-n), slide-panel model selector

feat(tagger): pose extraction + Qwen3-VL optional NL captioner + WD14x3 ensemble weights recalibration

feat(prod): rewrite pipeline + NoobAI + ensemble tagger (3xWD14+DeepDanbooru-ready) + 1000 artists + negative templates + history ext + scoring config