feat(comfyui): migrate 325G model tree worktank -> /storetank/arbo

ComfyUI's ~325G model tree moved off the near-full worktank NVMe (97%->26%,
342G free) to /storetank/arbo (roomy SATA SSD on irv-ml1), overlay-mounted
back at /basedir/models so ComfyUI behaviour is unchanged. rsync byte-verified
(src==dst), one comfyui restart, worktank original removed. Inventory of the
set in docs/arbo-comfyui-model-catalog.md for the retain decision. The older
919G /storetank/image-models/comfy archive is untouched (separate reclaim).

docs/arbo-comfyui-model-catalog.md

@@ -0,0 +1,60 @@
+# Arbo ComfyUI model catalog (worktank set)
+
+Snapshot **2026-06-13** of the ComfyUI model tree that arbo (hero/asset
+generation) consumes, taken at migration time. **~325 G**, migrated from
+`/worktank/comfyui/basedir/models` (irv-ml1 NVMe, was 97% full) →
+**`/storetank/arbo/models`** (irv-ml1 SATA-SSD) and overlay-mounted back into
+the ComfyUI container at `/basedir/models`.
+
+Purpose: decide what to **retain for future arbo use** vs prune. This is the
+*live* set ComfyUI serves today — distinct from the older 919 G CivitAI-managed
+archive at `/storetank/image-models/comfy` (untouched, separate question).
+
+## Per-category sizes
+
+| Category | Size | Notes |
+|---|---|---|
+| `diffusion_models/` | **203 G** | the bulk — flux2 / wan2.2 / qwen-image / z-image / ideogram (GGUF + fp8) |
+| `text_encoders/` | **64 G** | qwen3-VL, qwen2.5-VL, gemma4, umt5/t5-xxl, qwen3 (encoders for the above) |
+| `checkpoints/` | **31 G** | SDXL/Pony bases + SUPIR upscalers |
+| `loras/` | **14 G** | style/character LoRAs |
+| `controlnet/` | 5.7 G | flux upscaler + sdxl union-promax |
+| `clip_vision/` | 3.6 G | CLIP-ViT-H, clip_vision_h |
+| `vae/` | 3.0 G | wan2.2, flux2, flux1, z-image VAEs |
+| `ipadapter/` | 1.5 G | ip-adapter sdxl vit-h |
+| `RMBG/` `sams/` `upscale_models/` `ultralytics/` | <1.5 G ea | bg-removal, SAM, upscalers, yolo |
+| (empty placeholders) | — | unet, embeddings, hypernetworks, gligen, … (`put_*_here` only) |
+
+## Notable models (size · file)
+
+**diffusion_models/ (203 G):**
+- 15.7 G · Qwen-Image-Edit-2509-Q6_K.gguf
+- 14.3 G ×4 · wan2.2 i2v/t2v a14b low+high-noise Q8_0
+- 13.6 G · Wan2.2-Animate-14B-Q6_K · 13.5 G · Wan2.1_14B_VACE-Q6_K
+- 11.8 G · flux1-dev-Q8_0
+- 9.3 G ×2 · flux-2-klein 9b + base-9b Q8_0 · 8.9/8.8 G · same in fp8
+- 8.6 G ×2 · ideogram4 (+ unconditional) fp8
+- 6.7 G · z_image_turbo-Q8_0 · 5.0 G ×2 · Wan2.2-Fun-5B-Control, wan2.2-ti2v-5b
+- 4.0 G ×2 · flux-2-klein 4b + base-4b Q8_0 · 3.9 G · anima-base-v1.0 · 3.8 G ×2 · klein-4b fp8
+
+**text_encoders/ (64 G):** qwen3vl_8b_fp8 (9.9 G), qwen2.5_vl_7b_fp8 (8.7 G),
+gemma4_e4b_it_fp8 (8.4 G), flux2-klein-9b-uncensored-q8 (8.1 G), qwen_3_8b_fp8mixed
+(8.1 G), umt5_xxl_fp8 (6.3 G), t5xxl_fp8 (4.6 G), qwen3-4b-zimage-abliterated-q8
+(4.0 G), flux2-klein-4b-uncensored-q8 (4.0 G), qwen_3_06b_base (1.1 G).
+
+**checkpoints/ (31 G):** cyberrealisticPony_v180Coreshift (12.9 G),
+novaAnimeXL_illustrious (6.5 G), lustifySDXLNSFW_apexV8 (6.5 G),
+SUPIR-v0Q/v0F_fp16 (2.5 G ea).
+
+**loras/ (14 G):** gameart-game-icon sdxl (2.6 G) + diablo (0.9 G), RetroAnimeFluxV1
+(1.0 G), flux1_turbo_alpha, zit_* (nsfw/cinematic/ultrareal), LH_Pixar3D, oil-painting,
+hentai_hottie2_klein9b, … (~30 files).
+
+## Retain guidance (for operator)
+- **Likely keep (active gen stack):** the flux2-klein family + their text_encoders,
+  wan2.2 video set, qwen-image-edit, z-image, the SDXL/Pony checkpoints + their LoRAs,
+  VAEs, controlnet, clip_vision, ipadapter. This is a coherent current toolkit.
+- **Prune candidates:** redundant precisions (e.g., both Q8 *and* fp8 of the same
+  klein model where one suffices), superseded older variants, one-off LoRAs.
+- The 919 G `/storetank/image-models/comfy` archive (CivitAI-managed, older) is a
+  separate, larger reclaim target — abandoned (nothing mounts it). Operator's call.

stacks/comfyui/compose.yaml

@@ -49,6 +49,12 @@ services:
       - COMFY_CMDLINE_EXTRA=--disable-cuda-malloc --fp8_e4m3fn-text-enc
     volumes:
       - ${COMFYUI_BASEDIR}:/basedir
+      # models/ overlaid from storetank. The ~325 GB model tree was migrated
+      # off the near-full worktank NVMe (2026-06-13) to /storetank/arbo (roomy
+      # SATA SSD). This nested mount shadows the models subdir of /basedir;
+      # everything else (custom_nodes, output, input, user, workflows) stays on
+      # worktank. Inventory: docs/arbo-comfyui-model-catalog.md.
+      - ${COMFYUI_MODELS_DIR:-/storetank/arbo/models}:/basedir/models
       - ${COMFYUI_RUNDIR}:/comfy/mnt
     healthcheck:
       test: ["CMD-SHELL", "curl -fsS http://localhost:8188/ >/dev/null || exit 1"]