fix(spaces): trim preload_from_hub to 10 entries (HF cap) + future-work doc

HF caps preload_from_hub at 10 entries; we had 16 → Configuration error.
Dropped the 6 lightest entries (4 Dolly camera LoRAs at ~0.30 GB each,
Motion-Track IC-LoRA at 0.30 GB, Pose-Control IC-LoRA at 0.61 GB).
Those still lazy-download on first use of their respective modes.

Captured deeper optimizations (drop unused 84 GB Lightricks-side
transformer preload, drop GGUF, auto-sync README<->registry, etc.) in
docs/future_improvements.md for follow-up rather than bundling here.

README.md

@@ -13,16 +13,10 @@ preload_from_hub:
   - Comfy-Org/ltx-2 split_files/text_encoders/gemma_3_12B_it.safetensors
   - Kijai/LTX2.3_comfy diffusion_models/ltx-2.3-22b-dev_transformer_only_bf16.safetensors,loras/ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors,text_encoders/ltx-2.3_text_projection_bf16.safetensors,vae/LTX23_audio_vae_bf16.safetensors,vae/LTX23_video_vae_bf16.safetensors,vae/taeltx2_3.safetensors
   - Lightricks/LTX-2-19b-IC-LoRA-Detailer ltx-2-19b-ic-lora-detailer.safetensors
-  - Lightricks/LTX-2-19b-IC-LoRA-Pose-Control ltx-2-19b-ic-lora-pose-control.safetensors
-  - Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-In ltx-2-19b-lora-camera-control-dolly-in.safetensors
-  - Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-Left ltx-2-19b-lora-camera-control-dolly-left.safetensors
-  - Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-Out ltx-2-19b-lora-camera-control-dolly-out.safetensors
-  - Lightricks/LTX-2-19b-LoRA-Camera-Control-Dolly-Right ltx-2-19b-lora-camera-control-dolly-right.safetensors
   - Lightricks/LTX-2-19b-LoRA-Camera-Control-Jib-Down ltx-2-19b-lora-camera-control-jib-down.safetensors
   - Lightricks/LTX-2-19b-LoRA-Camera-Control-Jib-Up ltx-2-19b-lora-camera-control-jib-up.safetensors
   - Lightricks/LTX-2-19b-LoRA-Camera-Control-Static ltx-2-19b-lora-camera-control-static.safetensors
   - Lightricks/LTX-2.3 ltx-2.3-22b-dev.safetensors,ltx-2.3-22b-distilled-lora-384.safetensors,ltx-2.3-22b-distilled.safetensors,ltx-2.3-spatial-upscaler-x2-1.0.safetensors
-  - Lightricks/LTX-2.3-22b-IC-LoRA-Motion-Track-Control ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors
   - Lightricks/LTX-2.3-22b-IC-LoRA-Union-Control ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors
   - google/gemma-3-12b-it-qat-q4_0-unquantized gemma-3-12b-it/model-00001-of-00005.safetensors,gemma-3-12b-it/model-00002-of-00005.safetensors,gemma-3-12b-it/model-00003-of-00005.safetensors,gemma-3-12b-it/model-00004-of-00005.safetensors,gemma-3-12b-it/model-00005-of-00005.safetensors,gemma-3-12b-it/model.safetensors.index.json,gemma-3-12b-it/preprocessor_config.json,gemma-3-12b-it/tokenizer.model
   - unsloth/LTX-2.3-GGUF ltx-2.3-22b-dev-BF16.gguf

docs/future_improvements.md

@@ -0,0 +1,98 @@
+# Future improvements
+
+A backlog of optimizations that aren't blocking but would tighten the deploy.
+None of these are required for current functionality. Order is rough priority,
+not commitment.
+
+## Spaces / preload
+
+### 1. Stop preloading models that aren't referenced by any workflow
+
+Audit on 2026-05-02 (`tools/audit-models` style script) showed two `Lightricks/LTX-2.3`
+files in `preload_from_hub` that aren't actually referenced by any workflow JSON
+we ship:
+
+- `ltx-2.3-22b-dev.safetensors` (~42 GB)
+- `ltx-2.3-22b-distilled.safetensors` (~42 GB)
+
+The active path uses `Kijai/LTX2.3_comfy diffusion_models/ltx-2.3-22b-dev_transformer_only_bf16.safetensors`.
+Removing both saves ~84 GB of preload bandwidth/storage. Risk: if a future
+workflow update reintroduces the Lightricks-side filenames, lazy download
+takes over (slow first inference) — acceptable for the tradeoff.
+
+### 2. Drop `unsloth/LTX-2.3-GGUF` from preload (~39 GB)
+
+The GGUF transformer is the low-VRAM alternative. ZeroGPU H200 has 70 GB so
+the BF16 transformer always fits. Lazy-load when a future "Low VRAM" preset
+actually wires the GGUF path.
+
+### 3. Drop the `Lightricks/LTX-2-19b-LoRA-Camera-Control-Static/Jib-Up/Jib-Down` preload
+
+Each is ~2 GB. The Power Lora Loader has them all listed but defaults all to
+`on: false`, so they only load when the user picks one. Lazy-load is
+appropriate. Currently kept in preload because of the 10-entry cap +
+"easier to keep what we had".
+
+### 4. Auto-generate `preload_from_hub` from `MODEL_REGISTRY`
+
+Today the README list and `MODEL_REGISTRY` in `models.py` can drift. Build a
+small `tools/sync_preload.py` that:
+
+1. Reads `MODEL_REGISTRY`
+2. Walks the workflow JSONs to find which entries are actually referenced
+3. Sorts referenced entries by size (using `huggingface_hub` `repo_info`)
+4. Picks the top N entries that fit in the 10-cap
+5. Writes them back into the README YAML
+
+Run as a pre-commit or CI step.
+
+### 5. Bake custom-node clones into the build via `requirements.txt` git installs
+
+We currently `git clone` 10 custom-node repos in `_bootstrap()` at runtime.
+That's ~30 s of cold start. Some custom nodes ship as pip-installable; for
+the others, we could write a small `tools/install_custom_nodes.py` that
+runs at build time (via `pip install --no-deps` against git URLs) so the
+repos land in the image instead of being fetched at boot.
+
+Tradeoff: Spaces' build pipeline runs the gradio SDK Dockerfile which we
+don't control directly. The custom-node clone has to happen at runtime
+unless we can move it into the standard `requirements.txt` build step.
+
+### 6. Persistent storage add-on as the "$25/mo button"
+
+If iteration speed becomes the binding constraint, the persistent storage
+add-on (Spaces > Settings) at $25/mo for 150 GB makes everything just work
+— `/data` is writable, models live there forever, no preload dance.
+Sketched approach: `HF_HOME=/data/hf-cache` env var + `_bootstrap()` mkdir
+fallback. One-line code change.
+
+## Workflow / runtime
+
+### 7. Move ComfyUI custom-node `requirements.txt` install to build time
+
+Bootstrap currently `pip install`s each custom node's requirements at
+runtime. Most are no-ops (deps already in our top-level `requirements.txt`)
+but the `pip install --quiet` calls still take a few seconds each. Could
+audit and just merge them into the top-level `requirements.txt`.
+
+### 8. Clean up `nodes_replacements.py` warning
+
+ComfyUI core at our pinned commit (`eb0686bb`) emits
+`'function' object has no attribute 'register'` because the node-replacement
+API surface is incomplete at that SHA. Bumping `COMFYUI_COMMIT` to a newer
+tag should silence it. Pure cosmetic — no functional impact.
+
+### 9. Auto-close drawer when user navigates away from header
+
+Currently relies on document-level click listener. Works but has a
+microsecond race when the click target is between elements. Could use
+`pointerleave` on the drawer instead.
+
+## Cost-of-running
+
+### 10. Trim ZeroGPU duration cap
+
+Currently `@spaces.GPU(duration=300)` reserves 5 min per call. For Fast preset
+(distilled 8 steps) actual usage is ~30 s. Could shorten to 120 s — improves
+queue priority for the user (per HF docs). Use dynamic duration based on
+preset.