# 🗜️ SHRINKGOD — local GLB optimizer

Drop in a heavy GLB, pick a target, compare original vs optimized side-by-side (synced camera, animation playback, wireframe), export. **Everything runs client-side in the browser** — gltf-transform + meshoptimizer WASM. No uploads, files never leave the machine.

## Run

```bash
npm install
npm run dev        # http://localhost:5173
npm run build      # static build in dist/ — host anywhere, still fully local
```

Dev test hook: `?test=<file-in-public>.glb&run=1` auto-loads and optimizes.

## CLI (batch / farm use)

Same pipeline, headless — sharp for textures (faster than the browser canvas path):

```bash
node bin/cli.js model.glb                          # -> model.opt.glb (balanced)
node bin/cli.js --preset crunch *.glb --outdir out/
node bin/cli.js model.glb --ratio 0.15 --max-tex 512 -o small.glb
node bin/cli.js --help                             # all flags
```

`npm link` (or install from the repo) puts a `shrinkgod` command on PATH. Rig protection applies identically.

## Pipeline

dedup → flatten+join (static models only) → weld → per-primitive meshopt simplify → animation resample → prune → texture resize/re-encode → quantize → (optional) meshopt compression.

### Rig protection (automatic)
- Skinned meshes never go below **50% kept geometry** and error is capped at 0.5% — vertex collapse near joints wrecks skin weights in motion.
- Morph-target (blendshape) primitives are left untouched.
- Mesh joining is disabled for rigged/animated models (would break node animation).
- Best results: **optimize before rigging** — static geometry can be crunched far harder.

### Textures
- Longest side clamped to the chosen max (canvas re-encode).
- Color/emissive maps optionally converted to WebP (quality slider); normal/roughness/occlusion maps always stay lossless PNG.
- Re-encode is skipped if it wouldn't shrink the file.

### Presets
| | geometry kept | error | max texture | webp |
|---|---|---|---|---|
| Light | 50% | 0.5% | 2048 | no |
| Balanced | 25% | 1% | 1024 | yes |
| Crunch | 10% | 5% | 512 | yes |

## Draco input
Draco-compressed GLBs are decoded on load (decoder WASM is vendored in `public/draco/` — still fully local, no CDN) and written back out as standard uncompressed GLB. Draco *output* is intentionally not supported; use meshopt compression for web targets instead.

## Known limits
- Meshopt compression output (`EXT_meshopt_compression`) reads fine in three.js but **not in Blender** — leave it off for assets going back into a DCC.
- KTX2/basis textures pass through untouched.

## Bench (M5 MacBook Pro, Balanced preset)
- 15.8 MB photogrammetry room scan (151k tris, 13 MB textures) → **1.5 MB** (−90%) in 5 s.
- 3.5 MB rigged character (18.6k tris, 20 clips) → **2.7 MB** (−24%), clips + skinning intact (rig clamp limits geometry reduction by design).