What problem does it solve? Large digital-twin scenes in Godot stutter or miss frame budgets, and naive fixes like fixed-grid chunking or blanket occlusion culling can actually regress performance. This Skill provides measured recipes and a headless optimizer tool that scale a twin viewer to high instance counts without lying about fps. ## Core Features & Use Cases - Headless scene optimization: tools/optimize_scene.gd groups MeshInstance3Ds into auto-chunked MultiMeshInstance3D fields, preserves GlobalId joins via twin_globalids metadata, and emits a JSON report. - Scoped culling recipes: Measured guidance on when --occluders and --vis-ranges win (many-unique-mesh scenes at street level) versus when they are net-negative (single buildings) or no-ops (instanced/aerial scenes). - Honest benchmarking: tools/bench_scene.gd and tools/bench_sweep.sh implement a frames-drawn-delta methodology with vsync off, warmup, interleaved repeats, and determinism assertions that survive macOS display caps. - Use Case: A 28,600-mesh city scene renders at 27 fps aerial; run the optimizer with auto chunking to reach the 120 fps cap, then bench both street and aerial vantages to confirm the win before reporting it. ## Quick Start Run the headless optimizer on my imported twin model with auto chunking and then benchmark the before and after scenes at both street and aerial vantages.
