avbd-babylon 0.2.0 → 0.3.0

package/docs/README.md

@@ -1,459 +1,465 @@
-# Integrating AVBD into a Babylon.js project
-
-Use `AvbdPhysics` for GPU collision detection and rigid-body simulation, with
-Babylon meshes for the interactive view. The aggregate/body methods follow
-familiar Babylon conventions: mass, friction, velocities, impulses and disposal.
-AVBD computes body motion and contacts on the GPU; Babylon renders the results.
-This adapter has its own world; it does not implement Babylon's full
-Physics V2 plugin interface or install through `scene.enablePhysics()`.
-
-**Jump to:**
-
-- [Install the library](#install-the-library)
-- [A complete small scene](#a-complete-small-scene)
-- [Move, launch and reuse bodies](#move-launch-and-reuse-bodies)
-- [Choose a rendering path](#choose-a-rendering-path)
-- [Reproduce the collision and solver benchmarks](#reproduce-the-collision-and-solver-benchmarks)
-- [Current shape and API support](#current-shape-and-api-support)
-- [Choosing collision detection](#choosing-collision-detection)
-
-## Install the library
-
-The package is named `avbd-babylon`. For a local build, run `npm install` and
-`npm run build` in this repository. From your
-Babylon project, run `npm install /absolute/path/to/avbd` or install the archive
-created by `npm pack`. Import from `avbd-babylon` using your normal bundler.
-The ESM output already contains the solver and shader strings. No TypeScript
-source compilation or WGSL loader is needed in your project. The package has
-no Babylon runtime imports; it uses the mesh objects your project provides.
-
-`npm run build` builds only the library. `npm run build:web` builds the lab and
-browser tests. `npm start` builds the library and serves the lab together.
-`npm run build:cdn` creates `dist/avbd.global.js`, which exposes `globalThis.AVBD`.
-Host that file yourself and load it with a normal script tag; use
-`const { AvbdPhysics, AvbdPhysicsAggregate, AvbdShapeType } = AVBD`.
-There is no public CDN release URL yet.
-
-WebGPU
-needs a supported browser and localhost or HTTPS. The default world requests
-9 storage bindings, a 512 MiB storage binding limit and a 1 GiB buffer limit;
-unsupported adapters fail during creation. Babylon drawing can use either its
-WebGL or WebGPU engine when using the small-scene mesh bridge.
-
-## A complete small scene
-
-This example expects `<canvas id="canvas"></canvas>` styled to fill the page.
-Use your project's existing Babylon dependency.
-
-```js
-// src/main.js
-import {
-  Engine,
-  Scene,
-  ArcRotateCamera,
-  HemisphericLight,
-  MeshBuilder,
-  Vector3,
-} from "@babylonjs/core";
-import { AvbdPhysics, AvbdPhysicsAggregate, AvbdShapeType } from "avbd-babylon";
-
-const canvas = document.getElementById("canvas");
-const engine = new Engine(canvas, true);
-const scene = new Scene(engine);
-const camera = new ArcRotateCamera(
-  "camera",
-  -Math.PI / 2,
-  Math.PI / 3,
-  12,
-  new Vector3(0, 2, 0),
-  scene,
-);
-camera.attachControl(canvas, true);
-new HemisphericLight("light", new Vector3(0, 1, 0), scene);
-
-const floor = MeshBuilder.CreateBox(
-  "floor",
-  { width: 20, height: 1, depth: 20 },
-  scene,
-);
-floor.position.y = -0.5;
-const box = MeshBuilder.CreateBox("box", { size: 1 }, scene);
-box.position.y = 4;
-
-const physics = await AvbdPhysics.create({
-  scene,
-  gravity: new Vector3(0, -9.81, 0),
-  timeStep: 1 / 60,
-  iterations: 10,
-  capacity: 64,
-});
-new AvbdPhysicsAggregate(floor, AvbdShapeType.BOX, { mass: 0 }, scene);
-const fallingBox = new AvbdPhysicsAggregate(
-  box,
-  AvbdShapeType.BOX,
-  { mass: 2, friction: 0.6 },
-  scene,
-);
-
-// The world attaches to the scene and allocates on its first simulation step.
-// Create aggregates before the render loop starts.
-fallingBox.body.setAngularVelocity(new Vector3(0, 1, 0.5));
-
-engine.runRenderLoop(() => scene.render());
-const resize = () => engine.resize();
-window.addEventListener("resize", resize);
-window.addEventListener(
-  "pagehide",
-  () => {
-    window.removeEventListener("resize", resize);
-    physics.dispose();
-    scene.dispose();
-    engine.dispose();
-  },
-  { once: true },
-);
-```
-
-The fixed-step observer allows up to three catch-up steps per rendered frame.
-Mesh poses arrive asynchronously; do not await them in the render loop. `create({ scene })` attaches automatically and disposes with the scene. For a
-custom simulation loop, pass `autoAttach: false` and call `step()` yourself.
-`attachToScene()` is also available and does not attach a second observer when
-called again for the same scene.
-
-## Move, launch and reuse bodies
-
-```js
-fallingBox.body.setLinearVelocity(new Vector3(0, 0, 5));
-fallingBox.body.applyImpulse(new Vector3(3, 0, 0), box.position);
-fallingBox.body.applyAngularImpulse(new Vector3(0, 2, 0));
-
-// Teleport retains existing velocities; reset them explicitly for a fresh launch.
-fallingBox.body.teleport(new Vector3(0, 3, -5));
-fallingBox.body.setLinearVelocity(new Vector3(0, 0, 12));
-fallingBox.body.setAngularVelocity(Vector3.Zero());
-```
-
-Commands are combined and applied against the current GPU state. They do not
-upload a stale CPU pose. Create projectile slots before initialization and reuse
-them for spawning. `aggregate.dispose()` disables its collision participation;
-the caller still owns the mesh. Dispose the world before disposing its scene.
-
-## Choose a rendering path
-
-| Scene size and needs                | Rendering path                                       | Cost to watch                             |
-| ----------------------------------- | ---------------------------------------------------- | ----------------------------------------- |
-| Small interactive scenes            | Default Babylon mesh bridge                          | Pose readback and individual mesh updates |
-| Repeated geometry, moderate counts  | Instances or thin instances with a batched pose copy | CPU matrix updates and uploading matrices |
-| Tens of thousands of moving objects | Draw directly from the live GPU body buffer          | GPU rendering and collision/solver work   |
-
-Babylon [instances](https://doc.babylonjs.com/features/featuresDeepDive/mesh/copies/instances)
-share source geometry and material. [Thin instances](https://doc.babylonjs.com/features/featuresDeepDive/mesh/copies/thinInstances)
-store repeated transforms in buffers and avoid a separate scene object for every
-visible copy. Group objects by geometry and material. Instancing reduces drawing
-overhead; it does not reduce the number of physics contacts.
-
-For a moderate number of identical, origin-centred unit boxes, this bridge copies
-all poses once and updates a single thin-instance matrix buffer. Here `physics`
-was created with `syncMeshes: false`, and `boxes` contains box aggregates created
-with `sync: false`. Their original visual meshes should be hidden; `source` is
-an independent unit-box mesh at the origin with identity rotation and scale.
-
-```js
-import { Matrix, Quaternion, Vector3 } from "@babylonjs/core";
-
-physics.initialize();
-const indices = boxes.map((aggregate) => aggregate.body.gpuIndex);
-const matrices = new Float32Array(boxes.length * 16);
-const matrix = Matrix.Identity();
-const position = Vector3.Zero();
-const rotation = Quaternion.Identity();
-const scale = Vector3.One();
-let pending = false;
-
-source.thinInstanceSetBuffer("matrix", matrices, 16, false);
-scene.onBeforeRenderObservable.add(() => {
-  if (pending) return;
-  pending = true;
-  physics
-    .readBodies()
-    .then((poses) => {
-      for (let i = 0; i < indices.length; i++) {
-        const offset = indices[i] * 40;
-        position.set(poses[offset], poses[offset + 1], poses[offset + 2]);
-        rotation.set(...poses.subarray(offset + 4, offset + 8));
-        Matrix.ComposeToRef(scale, rotation, position, matrix);
-        matrix.copyToArray(matrices, i * 16);
-      }
-      source.thinInstanceBufferUpdated("matrix");
-      source.thinInstanceRefreshBoundingInfo();
-    })
-    .catch(console.error)
-    .finally(() => {
-      pending = false;
-    });
-});
-```
-
-This is a CPU bridge for moderate counts. It still copies the full body buffer,
-calculates matrices and uploads them. Bounding updates also cost CPU time. For
-known contained motion, a deliberately conservative bounding volume can avoid
-rebuilding bounds each update; ensure it covers all moving instances.
-For hulls or meshes with offset origins, reconstruct the mesh transform using
-the aggregate's local centre and principal rotation as the adapter does; the
-unit-box example does not handle those offsets.
-
-For a large scene:
-
-```js
-const physics = await AvbdPhysics.create({
-  capacity: 100_001,
-  iterations: 4,
-  syncMeshes: false,
-});
-// Create all bodies before initialize(), with { sync: false }.
-physics.initialize();
-const bodyBuffer = physics.bodyBuffer;
-const gpuIndex = aggregate.body.gpuIndex;
-physics.step(); // One fixed step; submission does not await completion.
-```
-
-A custom WebGPU renderer can read `bodyBuffer` on the **same GPUDevice** and draw
-instances from live poses. Buffers cannot be shared across devices. The adapter
-does not automatically connect this storage buffer to Babylon thin instances,
-and it does not access Babylon private engine fields. If your renderer exposes
-a device you own, pass it to `AvbdPhysics.create({ device, ... })`. A caller-owned
-device stays alive when the world is disposed.
-
-Thin instances do not automatically create colliders. The aggregate API takes
-one mesh per physics body, which has a startup cost for very large scenes.
-For 100K bodies, the laboratory's native body/scene builders avoid creating
-100K Babylon mesh objects; see [canonicalScenes.js](../demo/canonicalScenes.js).
-Import `AvbdScene2D` or `AvbdScene3D` from `avbd-babylon/native` for the same
-renderer-independent integration. Both expose live body handles, bulk commands,
-packed velocity uploads, selected pose reads, triggers, masks, restitution,
-sleeping and contact events. See the package README's
-[native examples](../README.md#native-physics-with-any-renderer).
-The direct-buffer renderer and small-scene aggregate bridge are separate
-integration paths.
-
-See [canonicalRenderer.js](../demo/canonicalRenderer.js) and
-[canonicalRender.wgsl](../demo/canonicalRender.wgsl) for the direct rendering path,
-and [mixedColliderDemo.js](../demo/mixedColliderDemo.js) for the Babylon mesh bridge.
-The laboratory scene builders use z up; the Babylon adapter uses y up by
-default. Match your renderer's coordinates to the world you create.
-
-The 3D body stride is 40 floats / 160 bytes: position at 0, quaternion xyzw at 4,
-size at 16, mass at 19, principal inertia at 20, linear velocity at 32 and angular
-velocity at 36. Spatial sorting changes GPU order. Use `body.gpuIndex` after
-initialization rather than assuming creation order equals buffer order.
-
-Use shared geometry/materials and a small number of draw calls. For independent
-solver measurements, omit drawing entirely. For whole-app comparisons, keep
-resolution, antialiasing, shadows, materials and camera fixed. Enhanced colors,
-lighting and 4× MSAA are interactive presentation options in this app;
-`quality=benchmark` disables them. Keeping physics settings fixed makes a
-rendering optimization easier to identify.
-
-## Reproduce the collision and solver benchmarks
-
-Start this repository with `npm start`, then open
-[/tests/performance.html](http://127.0.0.1:8080/tests/performance.html).
-Select a scene and play it before measuring. The preview closes during a run
-and restarts afterward. “Run all scenes” measures the complete scene catalog, including
-10K, 50K and 100K dense 2D piles.
-
-For a repeatable solver-only run, use this JavaScript in the benchmark page's
-developer console after it is ready. `__PERFORMANCE__` is the repository's test
-harness, not part of the Babylon world API.
-
-```js
-const measurements = [];
-await __BENCHMARK_VIEW__.suspendPreview();
-try {
-  for (const scene of [
-    "2d-stress-pile-10k",
-    "2d-stress-pile-50k",
-    "2d-stress-pile-100k",
-    "showcase-box-columns-100k",
-    "showcase-brick-ring-110k",
-  ]) {
-    for (let repeat = 0; repeat < 3; repeat++) {
-      const result = await __PERFORMANCE__.run(scene, "production", 60, 60);
-      measurements.push({ ...result, repeat });
-    }
-  }
-  console.table(
-    measurements.map((r) => ({
-      scene: r.scene,
-      objects: r.bodyCount,
-      contacts: r.peakContacts,
-      iterations: r.params.iterations,
-      collisionMs: r.summary.collision.mean,
-      movementMs: r.summary.solve.mean,
-      physicsMs: r.summary.total.mean,
-      slowerStepsMs: r.summary.total.p95,
-    })),
-  );
-  const data = { adapter: __PERFORMANCE__.info, measurements };
-  const url = URL.createObjectURL(
-    new Blob([JSON.stringify(data, null, 2)], {
-      type: "application/json",
-    }),
-  );
-  const download = document.createElement("a");
-  download.href = url;
-  download.download = "avbd-replication.json";
-  download.click();
-  setTimeout(() => URL.revokeObjectURL(url), 1000);
-} finally {
-  await __BENCHMARK_VIEW__.preview.load(document.getElementById("scene").value);
-}
-```
-
-Each run starts fresh, warms up for 60 steps and records 60 distinct GPU
-timestamp samples. Every step checks contact capacity and scheduling conflicts.
-The 2D stress scenes use touching boxes, gravity, friction and ten iterations,
-without sleeping. The 3D showcase uses each scene's recorded settings. Compare
-collision time with contact count; equal body counts across dimensions do not
-guarantee equal collision work.
-
-For controlled engineering profiles, use `scripts/profile-solver.mjs` or
-`scripts/profile-paired.mjs`. Set `AVBD_PROFILE_SCENES`,
-`AVBD_PROFILE_VARIANTS`, `AVBD_PROFILE_WARMUP`, `AVBD_PROFILE_SAMPLES` and
-`AVBD_PROFILE_REPEATS` explicitly. Public benchmark controls use the production
-automatic choice.
-
-GPU work layout selection is automatic in `AppGpuSolver3D` and the Babylon
-adapter. Body count and active joints select the contact calculation; contact
-load selects thread sharing for very large scenes. The 100K columns preset
-shares face-contact points across threads. Dense rigid scenes above 250K bodies
-with at least 1.75 manifolds per body use eight threads per body. The same large
-scenes cache sorting keys for contact lists of up to sixteen entries; longer
-lists retain their general sorting path. Deformable shaders remain intact.
-
-```js
-const world = await AvbdPhysics.create({ scene }); // automatic by default
-// Create aggregates before the render loop.
-world.initialize();
-console.log(world.solverDecision); // implementation and reason
-```
-
-Body and joint changes are reconsidered before the next step. Contact-load
-adaptation reuses existing capacity-counter readbacks rather than adding a
-per-frame pose copy. Positions, velocities, friction, rotation and iteration
-counts are retained. The policy follows measurements on the test laptop;
-results on other GPUs may differ. Advanced `setSolverMode()` overrides remain
-available for controlled engineering measurements and compatibility. The
-browser benchmark runs the production automatic choice.
-
-Web Workers can keep scene construction and command submission from blocking
-the page. They do not divide the already-parallel GPU kernels among CPU threads.
-The current API runs on the caller's thread. A complete worker integration would
-own the GPU solver and an OffscreenCanvas renderer together, send inputs and
-small status messages, and draw directly from the GPU body buffer. Copying every
-body's pose between threads each frame would defeat that design. Worker-based
-Babylon drawing also requires a compatible Babylon engine setup and replacing
-DOM-dependent controls with forwarded input messages. The root [README worker example](../README.md#worker-integration) shows the
-OffscreenCanvas setup and message forwarding. Use the benchmark's `setup` and
-`summary.submission` fields to
-judge the CPU cost separately from GPU physics.
-
-Both GPU solvers hold poses and velocities if a step runs
-out of contact storage or has conflicting solve groups. The host can then grow
-capacity without advancing an incomplete movement solve. Counters still report
-the failure: a held step does not pass correctness checks. This safeguard does
-not change valid-step equations or remove the need to reserve enough capacity.
-
-`npm run test:web` also checks screenshots, controls, benchmark preview lifecycle,
-navigation, and six-second dense 2D motion. Keep other GPU workloads closed.
-
-Lower GPU milliseconds are better; higher app FPS is better. Physics timings
-separate collision checks, contact setup, scheduling and movement calculations.
-They exclude drawing, sleeping, CPU submission and copying results back.
-Instrumented timing passes add overhead. The “CPU + wait” column includes GPU
-work and readback, so it is not an isolated CPU cost. Report GPU model, driver,
-scene, timestep, iterations, sample counts and contact counts with comparisons.
-
-The premade reports were recorded on an **NVIDIA GeForce RTX 4070 Laptop GPU**.
-They are saved examples; running the browser benchmark measures your own GPU.
-
-The [solver report](http://127.0.0.1:8080/test-results/solver-performance.html)
-includes the [AVBD paper](https://graphics.cs.utah.edu/research/projects/avbd/Augmented_VBD-SIGGRAPH25.pdf)
-comparison. Its RTX 4090 / DirectX / LBVH results and this browser's
-RTX 4070 Laptop / WebGPU / spatial-grid results have different hardware and
-timing boundaries. They are workload context, not a same-hardware speed ratio.
-
-## Current shape and API support
-
-Boxes, spheres and capsules have native collision geometry. Capsules use exact
-rounded segments along local Y. Cylinders and custom hulls use polygonal convex
-hulls, limited to 32 vertices. Meshes must be root
-meshes with positive scale; bake parent or negative transforms first. Colliders
-capture geometry, mass properties and transforms at creation. Sphere dimensions
-must be uniform.
-
-The adapter supports restitution, friction, velocity and impulse commands,
-teleports, collision masks, contact/trigger events, GPU raycasts, per-body sleep
-eligibility, ball joints, welds, elastic or breakable springs, 3D hinges with
-angle stops and torque-limited motors. Both dimensions expose GPU raycasts;
-3D sphere casts and 2D circle casts include rounded corner tests. It has no
-complete Physics V2 plugin or concave mesh collider.
-Static bodies with velocity move kinematically. Portable builders include 3D
-ropes, ragdolls and fabric. Automatic mesh synchronization is limited to 2048
-bodies by default; large scenes should draw GPU poses directly.
-For full method signatures and ownership rules, see [BABYLON_API.md](BABYLON_API.md).
-
-## Choosing collision detection
-
-`AvbdPhysics.create` accepts `broadphase: "auto"` (default), `"grid"` or
-`"hploc"`. Automatic selection uses a GPU tree for large scenes with varied
-collider sizes or many linked thin plates, such as the ragdoll net.
-Small scenes and uniform block scenes retain the grid. Measure your own workload;
-body count alone does not predict which method is faster.
-
-```js
-const physics = await AvbdPhysics.create({
-  scene,
-  capacity: 50000,
-  syncMeshes: false,
-  broadphase: "auto",
-  bvh: { rebuildInterval: 64 },
-});
-```
-
-The tree refits every collider each step, including rotations and teleports.
-Rebuilds improve its quality; refitting preserves coverage between rebuilds.
-The same narrowphase, friction and rotational AVBD equations handle its pairs.
-See [GPU tree design and limits](HPLOC.md) for implementation details.
-
-For a controlled comparison, open `/tests/performance.html`. Its collision
-selector compares the grid and tree without changing solving rounds. Both
-methods are also available in the playable laboratory.
-
-```js
-try {
-  const pair = await __PERFORMANCE__.runPaired(
-    "3d-mixed-sizes-50k",
-    ["grid", "hploc"],
-    60,
-    60,
-  );
-  console.table(
-    pair.results.map((r) => ({
-      collisions: r.broadphase,
-      collisionMs: r.summary.collision.mean,
-      physicsMs: r.summary.total.mean,
-      firstStepMs: r.startup.total,
-    })),
-  );
-} finally {
-  await __BENCHMARK_VIEW__.preview.load(document.querySelector("#scene").value);
-}
-```
-
-Use `grid-detail` / `hploc-detail` to split collision preparation, nearby-pair
-search, exact contact checks, body solving and contact-force updates. These
-extra timers add pass boundaries, so compare matching modes. Paired runs keep
-two independent simulations in memory; their absolute times are separate from
-the single-world paper comparison. Raw JSON retains startup, rebuild spikes,
-all samples, buffer bytes, selection decisions and noisy timing flags.
+# Integrating AVBD into a Babylon.js project
+
+Use `AvbdPhysics` for GPU collision detection and rigid-body simulation, with
+Babylon meshes for the interactive view. The aggregate/body methods follow
+familiar Babylon conventions: mass, friction, velocities, impulses and disposal.
+AVBD computes body motion and contacts on the GPU; Babylon renders the results.
+This adapter has its own world; it does not implement Babylon's full
+Physics V2 plugin interface or install through `scene.enablePhysics()`.
+
+**Jump to:**
+
+- [Install the library](#install-the-library)
+- [A complete small scene](#a-complete-small-scene)
+- [Move, launch and reuse bodies](#move-launch-and-reuse-bodies)
+- [Choose a rendering path](#choose-a-rendering-path)
+- [Reproduce the collision and solver benchmarks](#reproduce-the-collision-and-solver-benchmarks)
+- [Current shape and API support](#current-shape-and-api-support)
+- [Choosing collision detection](#choosing-collision-detection)
+
+## Install the library
+
+The package is named `avbd-babylon`. For a local build, run `npm install` and
+`npm run build` in this repository. From your
+Babylon project, run `npm install /absolute/path/to/avbd` or install the archive
+created by `npm pack`. Import from `avbd-babylon` using your normal bundler.
+The ESM output already contains the solver and shader strings. No TypeScript
+source compilation or WGSL loader is needed in your project. The package has
+no Babylon runtime imports; it uses the mesh objects your project provides.
+
+`npm run build` builds only the library. `npm run build:web` builds the lab and
+browser tests. `npm start` builds the library and serves the lab together.
+`npm run build:cdn` creates `dist/avbd.global.js`, which exposes `globalThis.AVBD`.
+Host that file yourself and load it with a normal script tag; use
+`const { AvbdPhysics, AvbdPhysicsAggregate, AvbdShapeType } = AVBD`.
+
+WebGPU
+needs a supported browser and localhost or HTTPS. The default world requests
+9 storage bindings, a 512 MiB storage binding limit and a 1 GiB buffer limit;
+unsupported adapters fail during creation. Babylon drawing can use either its
+WebGL or WebGPU engine when using the small-scene mesh bridge.
+
+For CPU physics, use the [native pipeline examples](../README.md#choose-a-solver-backend-once).
+Select the backend once; 2D and 3D scene builders, body commands, joints, triggers,
+restitution, collision masks and state reads keep the same API. The package
+supports WebGPU and CPU. The repository demos retain an experimental WebGL2
+solver for comparisons; its measured performance was too poor to include in
+the package. See the [backend benchmark](../README.md#backend-differences-and-benchmarks).
+
+## A complete small scene
+
+This example expects `<canvas id="canvas"></canvas>` styled to fill the page.
+Use your project's existing Babylon dependency.
+
+```js
+// src/main.js
+import {
+  Engine,
+  Scene,
+  ArcRotateCamera,
+  HemisphericLight,
+  MeshBuilder,
+  Vector3,
+} from "@babylonjs/core";
+import { AvbdPhysics, AvbdPhysicsAggregate, AvbdShapeType } from "avbd-babylon";
+
+const canvas = document.getElementById("canvas");
+const engine = new Engine(canvas, true);
+const scene = new Scene(engine);
+const camera = new ArcRotateCamera(
+  "camera",
+  -Math.PI / 2,
+  Math.PI / 3,
+  12,
+  new Vector3(0, 2, 0),
+  scene,
+);
+camera.attachControl(canvas, true);
+new HemisphericLight("light", new Vector3(0, 1, 0), scene);
+
+const floor = MeshBuilder.CreateBox(
+  "floor",
+  { width: 20, height: 1, depth: 20 },
+  scene,
+);
+floor.position.y = -0.5;
+const box = MeshBuilder.CreateBox("box", { size: 1 }, scene);
+box.position.y = 4;
+
+const physics = await AvbdPhysics.create({
+  scene,
+  gravity: new Vector3(0, -9.81, 0),
+  timeStep: 1 / 60,
+  iterations: 10,
+  capacity: 64,
+});
+new AvbdPhysicsAggregate(floor, AvbdShapeType.BOX, { mass: 0 }, scene);
+const fallingBox = new AvbdPhysicsAggregate(
+  box,
+  AvbdShapeType.BOX,
+  { mass: 2, friction: 0.6 },
+  scene,
+);
+
+// The world attaches to the scene and allocates on its first simulation step.
+// Create aggregates before the render loop starts.
+fallingBox.body.setAngularVelocity(new Vector3(0, 1, 0.5));
+
+engine.runRenderLoop(() => scene.render());
+const resize = () => engine.resize();
+window.addEventListener("resize", resize);
+window.addEventListener(
+  "pagehide",
+  () => {
+    window.removeEventListener("resize", resize);
+    physics.dispose();
+    scene.dispose();
+    engine.dispose();
+  },
+  { once: true },
+);
+```
+
+The fixed-step observer allows up to three catch-up steps per rendered frame.
+Mesh poses arrive asynchronously; do not await them in the render loop. `create({ scene })` attaches automatically and disposes with the scene. For a
+custom simulation loop, pass `autoAttach: false` and call `step()` yourself.
+`attachToScene()` is also available and does not attach a second observer when
+called again for the same scene.
+
+## Move, launch and reuse bodies
+
+```js
+fallingBox.body.setLinearVelocity(new Vector3(0, 0, 5));
+fallingBox.body.applyImpulse(new Vector3(3, 0, 0), box.position);
+fallingBox.body.applyAngularImpulse(new Vector3(0, 2, 0));
+
+// Teleport retains existing velocities; reset them explicitly for a fresh launch.
+fallingBox.body.teleport(new Vector3(0, 3, -5));
+fallingBox.body.setLinearVelocity(new Vector3(0, 0, 12));
+fallingBox.body.setAngularVelocity(Vector3.Zero());
+```
+
+Commands are combined and applied against the current GPU state. They do not
+upload a stale CPU pose. Create projectile slots before initialization and reuse
+them for spawning. `aggregate.dispose()` disables its collision participation;
+the caller still owns the mesh. Dispose the world before disposing its scene.
+
+## Choose a rendering path
+
+| Scene size and needs                | Rendering path                                       | Cost to watch                             |
+| ----------------------------------- | ---------------------------------------------------- | ----------------------------------------- |
+| Small interactive scenes            | Default Babylon mesh bridge                          | Pose readback and individual mesh updates |
+| Repeated geometry, moderate counts  | Instances or thin instances with a batched pose copy | CPU matrix updates and uploading matrices |
+| Tens of thousands of moving objects | Draw directly from the live GPU body buffer          | GPU rendering and collision/solver work   |
+
+Babylon [instances](https://doc.babylonjs.com/features/featuresDeepDive/mesh/copies/instances)
+share source geometry and material. [Thin instances](https://doc.babylonjs.com/features/featuresDeepDive/mesh/copies/thinInstances)
+store repeated transforms in buffers and avoid a separate scene object for every
+visible copy. Group objects by geometry and material. Instancing reduces drawing
+overhead; it does not reduce the number of physics contacts.
+
+For a moderate number of identical, origin-centred unit boxes, this bridge copies
+all poses once and updates a single thin-instance matrix buffer. Here `physics`
+was created with `syncMeshes: false`, and `boxes` contains box aggregates created
+with `sync: false`. Their original visual meshes should be hidden; `source` is
+an independent unit-box mesh at the origin with identity rotation and scale.
+
+```js
+import { Matrix, Quaternion, Vector3 } from "@babylonjs/core";
+
+physics.initialize();
+const indices = boxes.map((aggregate) => aggregate.body.gpuIndex);
+const matrices = new Float32Array(boxes.length * 16);
+const matrix = Matrix.Identity();
+const position = Vector3.Zero();
+const rotation = Quaternion.Identity();
+const scale = Vector3.One();
+let pending = false;
+
+source.thinInstanceSetBuffer("matrix", matrices, 16, false);
+scene.onBeforeRenderObservable.add(() => {
+  if (pending) return;
+  pending = true;
+  physics
+    .readBodies()
+    .then((poses) => {
+      for (let i = 0; i < indices.length; i++) {
+        const offset = indices[i] * 40;
+        position.set(poses[offset], poses[offset + 1], poses[offset + 2]);
+        rotation.set(...poses.subarray(offset + 4, offset + 8));
+        Matrix.ComposeToRef(scale, rotation, position, matrix);
+        matrix.copyToArray(matrices, i * 16);
+      }
+      source.thinInstanceBufferUpdated("matrix");
+      source.thinInstanceRefreshBoundingInfo();
+    })
+    .catch(console.error)
+    .finally(() => {
+      pending = false;
+    });
+});
+```
+
+This is a CPU bridge for moderate counts. It still copies the full body buffer,
+calculates matrices and uploads them. Bounding updates also cost CPU time. For
+known contained motion, a deliberately conservative bounding volume can avoid
+rebuilding bounds each update; ensure it covers all moving instances.
+For hulls or meshes with offset origins, reconstruct the mesh transform using
+the aggregate's local centre and principal rotation as the adapter does; the
+unit-box example does not handle those offsets.
+
+For a large scene:
+
+```js
+const physics = await AvbdPhysics.create({
+  capacity: 100_001,
+  iterations: 4,
+  syncMeshes: false,
+});
+// Create all bodies before initialize(), with { sync: false }.
+physics.initialize();
+const bodyBuffer = physics.bodyBuffer;
+const gpuIndex = aggregate.body.gpuIndex;
+physics.step(); // One fixed step; submission does not await completion.
+```
+
+A custom WebGPU renderer can read `bodyBuffer` on the **same GPUDevice** and draw
+instances from live poses. Buffers cannot be shared across devices. The adapter
+does not automatically connect this storage buffer to Babylon thin instances,
+and it does not access Babylon private engine fields. If your renderer exposes
+a device you own, pass it to `AvbdPhysics.create({ device, ... })`. A caller-owned
+device stays alive when the world is disposed.
+
+Thin instances do not automatically create colliders. The aggregate API takes
+one mesh per physics body, which has a startup cost for very large scenes.
+For 100K bodies, the laboratory's native body/scene builders avoid creating
+100K Babylon mesh objects; see [canonicalScenes.js](../demo/canonicalScenes.js).
+Import `AvbdScene2D` or `AvbdScene3D` from `avbd-babylon/native` for the same
+renderer-independent integration. Both expose live body handles, bulk commands,
+packed velocity uploads, selected pose reads, triggers, masks, restitution,
+sleeping and contact events. See the package README's
+[native examples](../README.md#native-physics-with-any-renderer).
+The direct-buffer renderer and small-scene aggregate bridge are separate
+integration paths.
+
+See [canonicalRenderer.js](../demo/canonicalRenderer.js) and
+[canonicalRender.wgsl](../demo/canonicalRender.wgsl) for the direct rendering path,
+and [mixedColliderDemo.js](../demo/mixedColliderDemo.js) for the Babylon mesh bridge.
+The laboratory scene builders use z up; the Babylon adapter uses y up by
+default. Match your renderer's coordinates to the world you create.
+
+The 3D body stride is 40 floats / 160 bytes: position at 0, quaternion xyzw at 4,
+size at 16, mass at 19, principal inertia at 20, linear velocity at 32 and angular
+velocity at 36. Spatial sorting changes GPU order. Use `body.gpuIndex` after
+initialization rather than assuming creation order equals buffer order.
+
+Use shared geometry/materials and a small number of draw calls. For independent
+solver measurements, omit drawing entirely. For whole-app comparisons, keep
+resolution, antialiasing, shadows, materials and camera fixed. Enhanced colors,
+lighting and 4× MSAA are interactive presentation options in this app;
+`quality=benchmark` disables them. Keeping physics settings fixed makes a
+rendering optimization easier to identify.
+
+## Reproduce the collision and solver benchmarks
+
+Start this repository with `npm start`, then open
+[/tests/performance.html](http://127.0.0.1:8080/tests/performance.html).
+Select a scene and play it before measuring. The preview closes during a run
+and restarts afterward. “Run all scenes” measures the complete scene catalog, including
+10K, 50K and 100K dense 2D piles.
+
+For a repeatable solver-only run, use this JavaScript in the benchmark page's
+developer console after it is ready. `__PERFORMANCE__` is the repository's test
+harness, not part of the Babylon world API.
+
+```js
+const measurements = [];
+await __BENCHMARK_VIEW__.suspendPreview();
+try {
+  for (const scene of [
+    "2d-stress-pile-10k",
+    "2d-stress-pile-50k",
+    "2d-stress-pile-100k",
+    "showcase-box-columns-100k",
+    "showcase-brick-ring-110k",
+  ]) {
+    for (let repeat = 0; repeat < 3; repeat++) {
+      const result = await __PERFORMANCE__.run(scene, "production", 60, 60);
+      measurements.push({ ...result, repeat });
+    }
+  }
+  console.table(
+    measurements.map((r) => ({
+      scene: r.scene,
+      objects: r.bodyCount,
+      contacts: r.peakContacts,
+      iterations: r.params.iterations,
+      collisionMs: r.summary.collision.mean,
+      movementMs: r.summary.solve.mean,
+      physicsMs: r.summary.total.mean,
+      slowerStepsMs: r.summary.total.p95,
+    })),
+  );
+  const data = { adapter: __PERFORMANCE__.info, measurements };
+  const url = URL.createObjectURL(
+    new Blob([JSON.stringify(data, null, 2)], {
+      type: "application/json",
+    }),
+  );
+  const download = document.createElement("a");
+  download.href = url;
+  download.download = "avbd-replication.json";
+  download.click();
+  setTimeout(() => URL.revokeObjectURL(url), 1000);
+} finally {
+  await __BENCHMARK_VIEW__.preview.load(document.getElementById("scene").value);
+}
+```
+
+Each run starts fresh, warms up for 60 steps and records 60 distinct GPU
+timestamp samples. Every step checks contact capacity and scheduling conflicts.
+The 2D stress scenes use touching boxes, gravity, friction and ten iterations,
+without sleeping. The 3D showcase uses each scene's recorded settings. Compare
+collision time with contact count; equal body counts across dimensions do not
+guarantee equal collision work.
+
+For controlled engineering profiles, use `scripts/profile-solver.mjs` or
+`scripts/profile-paired.mjs`. Set `AVBD_PROFILE_SCENES`,
+`AVBD_PROFILE_VARIANTS`, `AVBD_PROFILE_WARMUP`, `AVBD_PROFILE_SAMPLES` and
+`AVBD_PROFILE_REPEATS` explicitly. Public benchmark controls use the production
+automatic choice.
+
+GPU work layout selection is automatic in `AppGpuSolver3D` and the Babylon
+adapter. Body count and active joints select the contact calculation; contact
+load selects thread sharing for very large scenes. The 100K columns preset
+shares face-contact points across threads. Dense rigid scenes above 250K bodies
+with at least 1.75 manifolds per body use eight threads per body. The same large
+scenes cache sorting keys for contact lists of up to sixteen entries; longer
+lists retain their general sorting path. Deformable shaders remain intact.
+
+```js
+const world = await AvbdPhysics.create({ scene }); // automatic by default
+// Create aggregates before the render loop.
+world.initialize();
+console.log(world.solverDecision); // implementation and reason
+```
+
+Body and joint changes are reconsidered before the next step. Contact-load
+adaptation reuses existing capacity-counter readbacks rather than adding a
+per-frame pose copy. Positions, velocities, friction, rotation and iteration
+counts are retained. The policy follows measurements on the test laptop;
+results on other GPUs may differ. Advanced `setSolverMode()` overrides remain
+available for controlled engineering measurements and compatibility. The
+browser benchmark runs the production automatic choice.
+
+Web Workers can keep scene construction and command submission from blocking
+the page. They do not divide the already-parallel GPU kernels among CPU threads.
+The current API runs on the caller's thread. A complete worker integration would
+own the GPU solver and an OffscreenCanvas renderer together, send inputs and
+small status messages, and draw directly from the GPU body buffer. Copying every
+body's pose between threads each frame would defeat that design. Worker-based
+Babylon drawing also requires a compatible Babylon engine setup and replacing
+DOM-dependent controls with forwarded input messages. The root [README worker example](../README.md#worker-integration) shows the
+OffscreenCanvas setup and message forwarding. Use the benchmark's `setup` and
+`summary.submission` fields to
+judge the CPU cost separately from GPU physics.
+
+Both GPU solvers hold poses and velocities if a step runs
+out of contact storage or has conflicting solve groups. The host can then grow
+capacity without advancing an incomplete movement solve. Counters still report
+the failure: a held step does not pass correctness checks. This safeguard does
+not change valid-step equations or remove the need to reserve enough capacity.
+
+`npm run test:web` also checks screenshots, controls, benchmark preview lifecycle,
+navigation, and six-second dense 2D motion. Keep other GPU workloads closed.
+
+Lower GPU milliseconds are better; higher app FPS is better. Physics timings
+separate collision checks, contact setup, scheduling and movement calculations.
+They exclude drawing, sleeping, CPU submission and copying results back.
+Instrumented timing passes add overhead. The “CPU + wait” column includes GPU
+work and readback, so it is not an isolated CPU cost. Report GPU model, driver,
+scene, timestep, iterations, sample counts and contact counts with comparisons.
+
+The premade reports were recorded on an **NVIDIA GeForce RTX 4070 Laptop GPU**.
+They are saved examples; running the browser benchmark measures your own GPU.
+
+The [solver report](http://127.0.0.1:8080/test-results/solver-performance.html)
+includes the [AVBD paper](https://graphics.cs.utah.edu/research/projects/avbd/Augmented_VBD-SIGGRAPH25.pdf)
+comparison. Its RTX 4090 / DirectX / LBVH results and this browser's
+RTX 4070 Laptop / WebGPU / spatial-grid results have different hardware and
+timing boundaries. They are workload context, not a same-hardware speed ratio.
+
+## Current shape and API support
+
+Boxes, spheres and capsules have native collision geometry. Capsules use exact
+rounded segments along local Y. Cylinders and custom hulls use polygonal convex
+hulls, limited to 32 vertices. Meshes must be root
+meshes with positive scale; bake parent or negative transforms first. Colliders
+capture geometry, mass properties and transforms at creation. Sphere dimensions
+must be uniform.
+
+The adapter supports restitution, friction, velocity and impulse commands,
+teleports, collision masks, contact/trigger events, GPU raycasts, per-body sleep
+eligibility, ball joints, welds, elastic or breakable springs, 3D hinges with
+angle stops and torque-limited motors. Both dimensions expose GPU raycasts;
+3D sphere casts and 2D circle casts include rounded corner tests. It has no
+complete Physics V2 plugin or concave mesh collider.
+Static bodies with velocity move kinematically. Portable builders include 3D
+ropes, ragdolls and fabric. Automatic mesh synchronization is limited to 2048
+bodies by default; large scenes should draw GPU poses directly.
+For full method signatures and ownership rules, see [BABYLON_API.md](BABYLON_API.md).
+
+## Choosing collision detection
+
+`AvbdPhysics.create` accepts `broadphase: "auto"` (default), `"grid"` or
+`"hploc"`. Automatic selection uses a GPU tree for large scenes with varied
+collider sizes or many linked thin plates, such as the ragdoll net.
+Small scenes and uniform block scenes retain the grid. Measure your own workload;
+body count alone does not predict which method is faster.
+
+```js
+const physics = await AvbdPhysics.create({
+  scene,
+  capacity: 50000,
+  syncMeshes: false,
+  broadphase: "auto",
+  bvh: { rebuildInterval: 64 },
+});
+```
+
+The tree refits every collider each step, including rotations and teleports.
+Rebuilds improve its quality; refitting preserves coverage between rebuilds.
+The same narrowphase, friction and rotational AVBD equations handle its pairs.
+See [GPU tree design and limits](HPLOC.md) for implementation details.
+
+For a controlled comparison, open `/tests/performance.html`. Its collision
+selector compares the grid and tree without changing solving rounds. Both
+methods are also available in the playable laboratory.
+
+```js
+try {
+  const pair = await __PERFORMANCE__.runPaired(
+    "3d-mixed-sizes-50k",
+    ["grid", "hploc"],
+    60,
+    60,
+  );
+  console.table(
+    pair.results.map((r) => ({
+      collisions: r.broadphase,
+      collisionMs: r.summary.collision.mean,
+      physicsMs: r.summary.total.mean,
+      firstStepMs: r.startup.total,
+    })),
+  );
+} finally {
+  await __BENCHMARK_VIEW__.preview.load(document.querySelector("#scene").value);
+}
+```
+
+Use `grid-detail` / `hploc-detail` to split collision preparation, nearby-pair
+search, exact contact checks, body solving and contact-force updates. These
+extra timers add pass boundaries, so compare matching modes. Paired runs keep
+two independent simulations in memory; their absolute times are separate from
+the single-world paper comparison. Raw JSON retains startup, rebuild spikes,
+all samples, buffer bytes, selection decisions and noisy timing flags.