3D engine that AI agents drive
Docs: drop the guide sections replaced by learn pages, refresh two posters
Nguyen Xuan Anb575588839c6
diff --git a/changes/docs-guide-cleanup.md b/changes/docs-guide-cleanup.mdnew file mode 100644index 00000000..16429278--- /dev/null+++ b/changes/docs-guide-cleanup.md@@ -0,0 +1 @@+- Docs: the 22 guide sections that now have their own interactive learn pages are removed from docs/guide.md, so each fact lives in one place; the fog and shrinkwrap posters are re-rendered from their final examples.diff --git a/docs/guide.md b/docs/guide.mdindex 4eb27a7a..3492e57e 100644--- a/docs/guide.md+++ b/docs/guide.md@@ -60,346 +60,6 @@ The canvas is read in the same task as the render, so no `preserveDrawingBuffer`The `matrix` is a column-major 4x4 transform, and `box()` is a unit cube centred on the origin.The MCP server (`67js-mcp`) is the route for scene recipes, review grids and video, and `npm run render` in a source checkout exports video.-## Parts and geometry--A scene is an array of parts. Each part has a unique `name`, a shared `geometry`, a column-major 16-value `Float32Array` `matrix`, and a `material`.--Geometry uses `Float32Array` positions and normals, and `Uint32Array` triangle indices. Normals must have unit length and triangles face outward. Optional `uvs` have two values per vertex. Optional `colors` are opaque linear RGB.--Transforms must be finite and invertible. Use one geometry object for repeated parts, rather than creating identical meshes for every instance. Each geometry and texture descriptor identity pair becomes an instanced batch.--```js-import { Renderer67, box } from '67js';-const geometry = box(0.08, 4);-const parts = [{- name: 'Rounded cube', geometry,- matrix: new Float32Array([1,0,0,0, 0,1,0,0, 0,0,1,0, 0,0,0,1]),- material: { color: [0.3,0.42,0.55], metallic: 0.7, roughness: 0.3, emission: 0, tag: 'surface' },-}];-const renderer = new Renderer67(canvas, parts, {- ground: false, hdr: true, samples: 4, pixelRatio: 1.25, memoryBudgetMiB: 256,-});-renderer.render({ eye: [3,2,4], target: [0,0,0] });-```--Supply a real HTML canvas with nonzero CSS width and height. The example assumes `canvas` refers to that element. `box(fillet, samples)`, `cylinder(segments)`, `sphere(segments, rings)` and `torus(tubeRadius, segments, sides)` are core primitives. Inspect the supplied demos for dimension and transform conventions.--## Procedural modeling--Version `0.4.0` provides `lathe`, `extrude` and `tube` through `67js/addons/modeling`.-They generate static geometry from radius/height profiles, polygon contours and XYZ paths.-Extrusion supports holes and bevels. Open tube ends remain uncapped.--1. Choose a profile, contour or path in local coordinates.-2. Generate the geometry once with the matching addon function.-3. Assign the result to a part before creating `Renderer67`.-4. Share the geometry among repeated parts.--The [modeling contract](/modeling.md) defines input bounds, winding, bevel limits and resource estimates.-The [scene graph](https://github.com/codepawl/67.js/blob/main/docs/scene-graph.md) supports hierarchy and pivots without changing mesh topology.-The [modeling demo](https://67js.codepawl.com/modeling.html) combines these tools in a lantern scene.-These generators do not perform solid Boolean operations or provide a mesh editor.--## Sculpt brushes--Version `0.4.0` provides `sculptGeometry` through `67js/addons/sculpt`.-It applies `inflate`, directional `displace` and `smooth` strokes to a copy of static indexed geometry.-Stroke centers, radii and strengths use geometry-local units.-The result preserves topology, UVs and colors, and rebuilds normals.--1. Define bounded strokes on the undeformed mesh.-2. Call `sculptGeometry` before creating the renderer.-3. Inspect the resulting surface for folds or intersections.-4. Rebuild surface patches and query indices after changing geometry.--The [sculpt contract](https://github.com/codepawl/67.js/blob/main/docs/sculpt.md) defines brush strength, seam handling and work limits.-Smoothing can shrink volume. Sculpting rejects skin and morph channels and does not remesh or prevent intersections.-Compatible deformations can become [relative morph targets](https://github.com/codepawl/67.js/blob/main/docs/morph-targets.md) for animation.--## Contact detection--Version `0.4.0` provides static BVH queries through `67js/addons/mesh-query`.-`createMeshQueryIndex` copies positions and indices. `createMeshPairQuery` tests two indices under explicit affine matrices.-The `surfaces` mode finds triangle contact. The `closed-solids` mode also tests containment between two validated solid indices.--1. Build one index per static geometry with explicit count and memory budgets.-2. For containment, create both indices with `solid: true`.-3. Create a pair query with the required mode.-4. Evaluate the query with both current world matrices.-5. Dispose the indices when their geometry is no longer needed.--An index also supports `raycast(matrix, ray)` for a triangle number, hit point and barycentric coordinates.-The [mesh query contract](https://github.com/codepawl/67.js/blob/main/docs/mesh-query.md) describes results, solid validation, numerical limits and budget errors.-Rebuild indices after geometry edits. Queries reject skin and morph channels.-Contact results do not measure penetration depth or resolve collisions. Z-fighting has its own check: `renderer.diagnostics` and `scene_inspect` report `coplanar-overlap` for two parts that draw the same plane (within 1 mm, same facing, at least 1 cm2 shared). Fix it with `decal: true` on the smaller part (layer 1; `depthLayer` 0 to 4 orders stacked decals), or `environment.resolveCoplanar: 'offset'` to bias the smaller part of every flagged pair. Both draw with a scoped polygon offset in the colour pass and render passes, and never move geometry. Use a real offset along the normal when the gap should be visible in exports or the path tracer. The coplanar code is not in the core bundle: call `installCoplanarDiagnostics(renderer)` from `67js/addons/coplanar-overlap` to add the warning to `renderer.diagnostics` (checked once per layout), and `scene_inspect` shows it once `loadAllAgentSceneModules()` (or a recipe with `resolveCoplanar`) has loaded the code. Far from the camera, depth precision is the other cause: `installReversedDepth(renderer, 'auto')` from `67js/addons/reversed-depth` switches on reversed-Z with a 32-bit float depth buffer and camera-relative drawing when a frame's far / near passes 10,000, or when `installDepthPrecision` finds a flush pair more than 40 units from the eye. It still has a float32 limit, because vertex positions are float32. On the RTX 3060 a 0.1 mm gap at 500 m still draws the far surface on 6.9% of pixels, and a 1 mm gap at 500 m on 1.4%; at 50 m both stay under 1%. Keep surfaces that must not flicker at least 1 mm apart when they sit hundreds of metres from the camera, or mark the top one `decal: true`.--## Shrinkwrap and surface patches--Version `0.4.0` provides `shrinkwrapGeometry` through `67js/addons/shrinkwrap`.-It returns new static geometry in target-local coordinates while preserving source topology, UVs and colors.-Nearest mode chooses a surface point. Project mode casts along a target-local direction.--1. Define `sourceToTarget` to map source-local positions into target-local coordinates.-2. Choose `nearest` or `project` mode and a bounded `maxDistance`.-3. Set `clearance` and an explicit `onMiss` policy.-4. Use the returned geometry with the target part's retained matrix.-5. Inspect `stats` and the resulting mesh before rendering.--The [shrinkwrap contract](https://github.com/codepawl/67.js/blob/main/docs/shrinkwrap.md) defines local distances, work budgets and miss handling.-Collapsed output triangles cause errors. Projection does not guarantee clearance across the complete accessory surface.--`createSurfacePatch` from `67js/addons/surface` clips target triangles into an overlay with its own UVs and material.-The [surface patch contract](https://github.com/codepawl/67.js/blob/main/docs/surface.md) defines projector orientation, depth and local clearance.-Patches and shrinkwrap are static operations. They do not follow skin or morph deformation.--## HDR environment lighting--Pass `environment: { map }` to light and reflect from an equirectangular HDR image. `loadHdri(url)` from `67js/addons/hdri` parses a Poly Haven `.hdr` file, and `createEnvironmentMap` accepts your own float data. Without a map, rendering is unchanged. The [lighting contract](/lighting.md#environment-maps-ibl) covers rotation, background display, memory and limits.--## Humanoid retargeting and crowds--`67js/addons/retarget` is available from version `0.10.0`. It moves a humanoid clip onto another skeleton and offsets many copies in time.--1. Build rigs with `createRigFromSkin(asset.skins[0])`. Name families (Mixamo, VRM, Unreal, Rigify, Quaternius) map through `mapHumanoidBones`; fix a miss with `targetOverrides`.-2. Call `retargetClip(sourceRig, targetRig, clip, { rootMotion })`.-3. Play it with `createSkinClipPlayer(skin, clip).apply(seconds)`, then `renderer.updateParts(parts)`.-4. For a crowd, `createCrowd({ count, clip, seed })` gives placement, time offsets and playback rates. For dozens or hundreds of animated instances, bake the clip once with `67js/addons/crowd-skinning` so they draw as a few instanced batches (see `docs/retarget.md`, Crowds); the plain path needs one skin pose per instance.--The [retarget contract](https://github.com/codepawl/67.js/blob/main/docs/retarget.md) covers rest-pose assumptions and measured cost. Mixamo clips need the owner's Adobe login to download; the code only consumes files.--## Baked liquid--`67js/addons/spume` is available from version `0.10.0`. It plays back liquid baked offline by the Spume GPU solver.--1. Bake a cache folder with Spume's `run.py` or its Blender add-on.-2. Load it with `loadSpumeCache(folderUrl)`.-3. Create a player with `createSpumePlayer(cache, { matrix, loop })`.-4. Each frame, call `renderer.replaceParts([...sceneParts, ...player.partsAt(seconds)])` and render.--The [Spume contract](https://github.com/codepawl/67.js/blob/main/docs/spume.md) covers the cache format, whitewater options, placement and measured cost.-Skinned and morphed parts may stay in the list unchanged while frames are replaced; adding or removing them needs a new renderer. Textured parts can be added and removed once `import '@codepawlhq/67js/addons/replace-textured-parts'` has run; without it they must stay unchanged. The addon deletes a texture when no part uses it any more and keeps one another part still binds.--## Wheeled vehicles--`67js/addons/vehicle` is available from version `0.10.0`. It drives a car or truck along a path with deterministic, physics-free animation.--1. Describe the vehicle with `createVehicleRig({ parts })`; wheels are found by name and geometry. Check `rig.wheels` or `findWheelCandidates(names)`.-2. Build a drive with `driveAlongPath(rig, points, { cruiseSpeed, groundHeight })`.-3. Each frame, call `applyVehiclePoseToParts(rig, drive.sample(seconds))` and then `renderer.updateParts(parts)`.--Wheel spin, front-wheel steering, roll, pitch and ground following come from the path.-Scene-graph nodes and imported GLBs use `applyVehiclePoseToSceneGraph` and `applyVehiclePoseToAsset`.-The [vehicle contract](https://github.com/codepawl/67.js/blob/main/docs/vehicle.md) lists conventions, options and limits.--## Trees and grass--`67js/addons/vegetation` is available from version `0.10.0`. It grows seeded trees with real branches and leaf cards, grass and flowers.--1. Scatter with `scatterVegetation({ area, density, terrainHeight, slopeLimit, exclusion, mix, leaves })`.-2. Pick detail per camera with `selectLevels(forest.instances, eye)`.-3. Pass `view.parts` to `Renderer67` and rebuild when `view.changed` is true.--Cards (`leaves: 'cards'`, the default) are single-sided, alpha-masked, double-sided `foliage` textures that sway with the wind; `'mesh'` leaves cost about twice the triangles. Pass `barkTextures` for scanned bark and `leafTextures: await loadLeafTextures('/textures/leaves')` for scanned leaf cards. The [Vegetation contract](https://github.com/codepawl/67.js/blob/main/docs/vegetation.md) lists species, triangle budgets and limits.--## Grass fields--`67js/addons/grass` is in the current source and not yet published. It covers terrain with clumped, curved blades in tiles with distance level of detail.--1. Create the field with `createGrass({ preset, area, terrainHeight, mask, eye })`. Presets are `meadow`, `lawn`, `savanna` and `wild`.-2. Add `grass.parts` to the renderer and pass `wind` in each frame.-3. When `grass.update(eye)` returns true, call `renderer.replaceParts` with the new part list.--Tint the ground under the grass with `grass.groundColor`. The [Grass contract](https://github.com/codepawl/67.js/blob/main/docs/grass.md) lists options, levels, measured budgets and limits.--`67js/addons/clutter` is in the current source and not yet published. It scatters pebbles, rocks, twigs and fallen leaves over terrain in the same streamed tiles. Pick a `preset` (`meadow`, `forest`, `riverbed`, `scree`), pass `terrainHeight`, and add `clutter.parts` next to the grass. The [Clutter contract](https://github.com/codepawl/67.js/blob/main/docs/clutter.md) lists options and budgets.--`67js/addons/light-bake` is in the current source and not yet published. `bakeLightmaps(parts, lights, options)` ray traces the light that bounces between surfaces into a lightmap texture for each part and returns new parts that draw it through the emissive map; direct light stays live. The [Light bake contract](https://github.com/codepawl/67.js/blob/main/docs/light-bake.md) lists the limits.--## Terrain and ground textures--`67js/addons/terrain` and `67js/addons/texture-set` are available from version `0.10.0`. They build a chunked, seeded heightfield with distance level of detail and a four-layer ground material (slope, height and mask rules, height-based edge blending, triplanar cliffs, anti-tiling) from CC0 scans.--1. `const set = await loadTextureSet('/textures/leafy_grass')` per layer, then `material.layers = [{ ...set, scale: 3 }, ...]` and `material.splat = { rules }`.-2. `const terrain = createTerrain({ size, chunks, resolution, levels, seed })`; give every `terrain.parts` entry the one shared material.-3. Pass `terrain.heightAt` to `scatterVegetation` and vehicles, and call `renderer.replaceParts` when `terrain.update(eye)` returns parts.--The [Terrain contract](https://github.com/codepawl/67.js/blob/main/docs/terrain.md) lists every option, the texture-array memory and the limits.--## Retained surface attachments--`createSurfaceAttachment` is introduced in `0.4.1`. Published npm `0.4.0` does not include this addon.-It binds a rigid accessory to one triangle through world transforms, relative morphs and four-influence skinning.-The accessory uses the attachment's retained world matrix. Its positive Z axis follows the posed triangle's geometric winding.--1. Choose an original triangle number and barycentric coordinates for the contact point.-2. Bind the target with `createSurfaceAttachment` from `67js/addons/surface-attachment` using `0.4.1` or a current source build.-3. Assign `attachment.matrix` to the accessory part before creating the renderer.-4. Update the target matrix, morph weights and skin palette for the frame.-5. Call `attachment.update()` before `renderer.updateParts()`.--The [attachment contract](https://github.com/codepawl/67.js/blob/main/docs/surface-attachment.md) defines retained identities, world-unit clearance and unsupported shader deformation.-Updates sample the selected triangle without rebuilding a BVH.-Static raycast hits can supply binding coordinates, but query indices cannot raycast a visibly skinned or morphed surface.-An attachment positions one rigid frame. It does not conform the accessory footprint or prevent every intersection.--## Renderer reference--`new Renderer67(canvas, parts, options)` owns GPU geometry, texture and target allocations. Constructor options:--| Option | Default | Purpose |-| --- | --- | --- |-| batching | true | Share instance batches by geometry and map identity |-| pixelRatio | 1.5 | Render scale relative to CSS size. Not automatically devicePixelRatio |-| ground | true | Add the studio floor at y 0.019, or 2% up from the lowest point of a scene it would bury more than a tenth of (small objects, scenes below y 0). Disable for worlds, 2D and assets |-| world | false | Enable outdoor rendering and world clipping defaults |-| samples | 4 | Requested MSAA count. Negotiated against hardware |-| hdr | false | Request floating-point color targets |-| (compact colour) | off | `useCompactColor(renderer)` from `@codepawlhq/67js/addons/compact-color` stores the HDR frame and its multisample colour as R11F_G11F_B10F instead of RGBA16F: 39.55 MB less at 1920 x 1080 with 4 samples (110.92 to 71.37 MB). The format has no alpha, so it needs `hdr: true` and an opaque canvas and throws otherwise; it also throws when the device does not support the renderer's sample count in that format. Mean FLIP against the default is 0.002 to 0.030 (median 0.021) over the 26 reference scenes. Without the call nothing changes. `renderer.memoryReport()` lists the formats |-| memoryBudgetMiB | sized to the scene | Estimated render allocation budget. Without it the renderer takes room for the scene and one full frame: at least 256 MiB, at most 1024 MiB on a device reporting 8 GiB of memory or more (512 MiB otherwise). A scene over the budget fails with an error that names the bytes and the `memoryBudgetMiB` that fits. |-| environment | {} | Sky, light and fog settings |-| shadowBounds | {} | Custom extent, eye and target for shadow framing |-| benchmark | false | Diagnostic flat shader. Not the appearance renderer |-| shading | 'full' | `'light'` is a cheaper picture for weak machines and software rendering: at most 2 MSAA samples, one shadow fetch, a fitted BRDF, the sky reflected by its blurred irradiance, no second order irradiance. `'full'` is unchanged |--`render(frame)` draws synchronously. `resize()` replans targets after CSS size or `pixelRatio` changes. `updateParts(parts, { shadows: true })` uploads changed matrices and materials using retained buffers. It does not add new part, geometry or texture identities.--`updateTexture(descriptor, { image?, region? })` pushes new pixels into a texture the renderer already uploaded, such as a live canvas or a simulation. Import `@codepawlhq/67js/addons/texture-update` first. `docs/material-maps.md` has the options.--Construct a new renderer for a new graph. `dispose()` releases owned resources.--Read `memoryPlan` for actual width, height, estimatedBytes and reduced status. `capabilities` for HDR and MSAA negotiation. `triangles`, `batches.length` and `cpuMilliseconds` for diagnostics. CPU submission time is not GPU frame time.--## Local lights and volume fog--Version `0.4.0` supports local lighting through the renderer's `environment` option and per-frame environment overrides.-`pointLights` accepts at most 16 world-space lights with linear RGB colors.-`keyIntensity`, `fillIntensity` and `ambientIntensity` control the existing studio lighting.--1. Set the environment's lighting controls when creating the renderer.-2. Supply the complete `frame.environment.pointLights` list to animate local lights.-3. Inspect exposure and material response at a fixed sample time.--The [lighting contract](/lighting.md) defines intensity, radius, range and validation limits.-Point lights have no local shadows and can illuminate through walls.-Radius regularizes attenuation and does not create an area emitter.--The [volume fog contract](https://github.com/codepawl/67.js/blob/main/docs/volumetric.md) covers depth-clipped directional scattering in version `0.4.0`.-Active fog requires `hdr: true`, `depth: true` and supported floating-point targets.-Its depth texture contributes to the render memory estimate.--## Deterministic animation--```js-import { createAnimationClip, createAnimationPlayer } from '67js';-const clip = createAnimationClip({ duration: 2, tracks: [{- target: 'Rounded cube', property: 'translation', interpolation: 'SMOOTH',- keyframes: [{ time: 0, value: [0,0,0] }, { time: 2, value: [0,1,0] }],-}] });-const player = createAnimationPlayer(clip, parts);-player.apply(1);-renderer.updateParts(parts);-renderer.render({ eye: [3,2,4], target: [0,0,0], time: 1 });-```--Tracks target unique part names. Supported transform tracks are translation, rotation (quaternion XYZW) and scale. Material tracks include metallic, roughness and emission. Interpolation supports `LINEAR`, `STEP`, `SMOOTH` and `CUBICSPLINE`. A keyframe can add `easing` (`easeOutBack`, `easeInOutCubic`, `cubic-bezier(...)`) to shape its segment.--`sampleAnimationClip(clip, time)` returns samples without binding parts. Preserve the original graph and apply absolute time so seeking backwards reproduces the scene. Use `requestAnimationFrame` for playback, pause when hidden, and use a fixed sample time for export.--## Animation curves and mixing--Version `0.4.0` supports `CUBICSPLINE` tracks with explicit `inTangent` and `outTangent` values on every keyframe.-Tangents represent component changes per second. Quaternion tangents are component derivatives.-The [curve contract](https://github.com/codepawl/67.js/blob/main/docs/animation-curves.md) defines interpolation, overshoot constraints and retained player buffers.--`createAnimationMixer(parts)` combines clips through reusable actions and absolute-time control timelines.-Actions support weights, playback speed, fades and crossfades.-The [mixer contract](https://github.com/codepawl/67.js/blob/main/docs/animation-mixer.md) defines blending, ownership and control ordering.--1. Create clips and bind their actions through `mixer.clipAction(clip)`.-2. Author `play`, `fadeTo` or `crossFade` controls in nondecreasing time order.-3. Call `mixer.apply(time)` for the desired absolute time.-4. Upload the changed parts with `renderer.updateParts(parts)`.--Playback can seek backward independently of control authoring order.-The mixer owns its registered properties and captures their base values during binding.--## Skeletal and morph animation--Version `0.4.0` supports four-influence GPU skinning and relative position/normal morph targets.-`createSkinPose` supplies a retained joint palette. `updateSkinPose` calculates mesh-local joint matrices from world joints and inverse binds.-`createMorphPose` supplies retained weights for the geometry's relative targets.-The renderer applies morph deltas before skinning.--1. Define static skin influences or relative morph channels before renderer construction.-2. Assign the matching retained `skin` and `morph` poses to each part.-3. Update joint palette values and morph weights for the desired frame.-4. Call `renderer.updateParts(parts)` to validate and upload the poses.--The [skinning contract](https://github.com/codepawl/67.js/blob/main/docs/skinning.md) defines supported transforms, four influences and identity requirements.-The [morph contract](https://github.com/codepawl/67.js/blob/main/docs/morph-targets.md) defines channel shapes, weight bounds and resource estimates.-Changing topology, influences or pose identities requires a new renderer.--The glTF addon imports supported skins and relative targets, including joint and weight animation.-`asset.applyAnimation(time, clipIndex)` updates retained parts and poses for upload.-The [glTF morph contract](https://github.com/codepawl/67.js/blob/main/docs/gltf-morph.md) and [asset contract](/gltf.md) define the supported asset subset.--## Recorded physics--Version `0.4.0` provides recorded rigid-body simulation through `67js/addons/physics`.-The addon requires the separate optional peer dependency `@dimforge/[email protected]`.-It bakes body contacts, joints, limits and bounded motors into a timeline.-Playback samples that recording without rerunning the solver.--1. Define named bodies, colliders, joints and tick events.-2. Call `planPhysicsScene` to validate the definition and allocation estimate.-3. Await `createPhysicsScene` to bake the bounded recording.-4. Bind visual parts through `physics.bindParts`.-5. Apply bindings at absolute time before uploading changed parts.-6. Call `physics.dispose()` when the recording is no longer needed.--The [physics contract](/physics.md) defines supported colliders, joint controls, timing and disposal.-Recorded poses can seek forward or backward. Interpolation between recorded frames can show small penetrations.-The bake can block the main thread after WASM initialization.-Cloth, soft bodies, mesh colliders and direct skeletal bindings are unsupported.--## Assemblies and interaction--Explodable parts include `assembly: { layer: 'exterior' | 'interior', section, offset: [x,y,z], detailOffset: [x,y,z] }`. Base matrices stay assembled. A frame's `assemblyAmount` goes from 0 to 1. Section offsets separate first, small detail offsets separate later.--`assemblyOffset(part, amount)` returns the same offset for export. Filter parts by layer and recreate the renderer after disposing its predecessor.--The object studio implements pointer capture for orbit, wheel zoom, keyboard arrows and a slider. Render only after a change when the scene is static. Use a ResizeObserver to call `resize()` and redraw. Offer keyboard controls and honor `prefers-reduced-motion`.--## Images and films--Offline media commands run from the repository checkout, not from an installed npm package. Install checkout dependencies and provide Chrome and FFmpeg. The exporter reports missing tools. On Windows it uses Chrome's standard install path and `~/.local/bin/ffmpeg.exe`.--```sh-npm run render -- --scene motion --kind image --backend gpu --time 3 --output artifacts/frame.png-npm run render -- --scene motion --kind video --backend gpu --encoder cpu --preset balanced --duration 12 --fps 24 --memory 384 --output artifacts/film.mp4-npm run render -- --scene motion --kind video --backend cpu --preset cpu --aspect portrait --duration 2 --output artifacts/cpu-trial.mp4-```--Scenes: object, nature, film, motion, view2d and assets. Use `--assembly 1` for object inspection. `--asset path/model.glb` with assets. `--start` for a timeline range.--`--aspect` supports landscape, portrait and square. `--width` and `--height` set dimensions (even dimensions for video). `--encoder gpu` requires NVENC. GPU rasterization and GPU encoding are independent.--Select original film audio with `--audio`. Other scenes default to silent.--Presets: preview 960×540 / 2 samples / 24fps / 256MiB. Balanced 1280×720 / 4 / 24 / 384MiB. High 1920×1080 / 4 / 24 / 512MiB. Cpu 640×360 / 1 / 12 / 192MiB.--Export one short trial before a full film. Inspect output JSON and FFprobe for backend, dimensions, duration, frame count and audio.--## Performance and memory--Reuse geometry and decoded textures. Do not create a renderer per frame. Change existing transforms through `updateParts`. Use lower scene detail before increasing the budget.--Pause hidden scenes and avoid simultaneous GPU exports. For static previews, redraw on demand.--The planner includes geometry, textures and mipmaps, shadows, MSAA, HDR, bloom and one capture staging frame. It reduces render dimensions when necessary. The exporter keeps one pending capture and respects encoder backpressure. This is an estimated allocation budget, not a hard process RAM/VRAM cap.--Driver/browser allocations can still fail. Sampled browser JS heap and Node RSS do not measure total graphics memory.-## Addons and limitsCore exports are available from `67js`:diff --git a/public/docs-media/fog-volume.webp b/public/docs-media/fog-volume.webpindex bab05e33..cb253c63 100644Binary files a/public/docs-media/fog-volume.webp and b/public/docs-media/fog-volume.webp differdiff --git a/public/docs-media/shrinkwrap-grid.webp b/public/docs-media/shrinkwrap-grid.webpindex 06eab638..85efe757 100644Binary files a/public/docs-media/shrinkwrap-grid.webp and b/public/docs-media/shrinkwrap-grid.webp differ