docs(simulation): record collision publication contract

docs/20_SIMULATION_WORLD_SURFACE_BRIEF.md

@@ -57,7 +57,8 @@ Mission Core backend
 DC Gaussian Pipeline
   ├─ TUS bundle or archive admission
   ├─ secure ZIP/RAR/7z normalization
-  └─ SplatTransform build on Worker 006
+  ├─ native Vulkan SplatTransform build on Worker 006
+  └─ bounded collision simplification + mandatory Draco publication
 ```
 
 The browser never receives the Worker token. Mission Core does not embed archive-format behavior
@@ -70,10 +71,22 @@ Mission Core pins the direct `playcanvas` package. `PlayCanvasRuntime` owns one
 camera controls, GSplat asset lifecycle and later physical/debug entities. React owns no per-frame
 entity state. A new world is data and does not require a new frontend build.
 
-The first Worker 006 provider is CPU-only and therefore does not advertise collision outputs. The
-UI disables collision/combined modes and states this explicitly; it does not synthesize a proxy.
-The manifest already keeps visual and collision URLs and their independent world transforms so the
-GPU collision build can be admitted without changing the project surface.
+Worker 006 runs the portable API in its own read-only Docker composition and delegates only the
+SplatTransform GPU command through a confined filesystem spool to the pinned native Vulkan runtime
+on the RTX 4090. It does not install into or share the Python, CUDA, Triton or computer-vision
+environments on the host.
+
+Collision publication is deliberately two-stage. SplatTransform generates smooth collision and
+SVO query data at a scene-type voxel size (`0.2 m` outdoors, `0.1 m` indoors/objects). The isolated
+pipeline container then bounds the debug/physics mesh to a default 250,000-triangle target and
+always writes `KHR_draco_mesh_compression` before the artifact digest is sealed. Simplification
+controls decoded geometry cost; Draco controls transfer and stored bytes and is never treated as a
+replacement for a physics mesh budget. The browser loads the result through a pinned local Draco
+WASM decoder with no external decoder request.
+
+Visual and collision layers retain independent X/Y/Z correction settings, while PlayCanvas world,
+camera, navigation and future physics stay in the canonical Y-up coordinate system. Quality, both
+layer transforms and camera-axis inversion are persisted atomically in each project's metadata.
 
 Primary implementation references:
 
@@ -92,8 +105,9 @@ Primary implementation references:
 - encrypted, linked, traversing, duplicate and over-limit archive entries fail closed;
 - project status is durable and reflects provider state without fabricated percentages;
 - ready artifacts are imported digest-bound and served from Mission Core same-origin URLs;
+- every published collision GLB is simplified to the configured budget and requires Draco;
 - edit changes project metadata; delete removes both the Mission Core project and terminal provider
   job;
 - a ready project mounts direct PlayCanvas Engine and loads Streamed SOG with preview fallback;
-- visual, collision and combined remain distinct runtime modes even when collision is not yet
-  available on the CPU provider.
+- visual, collision and combined remain distinct runtime modes, and collision is loaded lazily;
+- viewer quality, layer-axis correction and camera inversion survive navigation and reload.