3D engine that AI agents drive
JavaScript 3d3d-engineagentic-aiai-agentsblenderclaude-codegame-enginegaussian-splattinggltfjavascriptmcpmcp-servermodel-context-protocolpath-tracerpath-tracingpbrrenderingthreejs-alternativewebgl2webgpu
Release 0.21.0
Nguyen Xuan Ana850b9a0b47c
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.jsonindex 24332552..7196bfe7 100644--- a/.claude-plugin/marketplace.json+++ b/.claude-plugin/marketplace.json@@ -13,7 +13,7 @@"source": "./skills","strict": false,"description": "The Crucible MCP server for WebGL2 and path-traced scenes, and the build-crucible-scene and render-crucible-media skills.",- "version": "0.20.0",+ "version": "0.21.0","category": "development","homepage": "https://crucible.codepawl.com","repository": "https://github.com/codepawl/crucible",diff --git a/CHANGELOG.md b/CHANGELOG.mdindex c7563157..e92a6e46 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -1,5 +1,313 @@# Changelog+## 0.21.0++- Breaking: `bake.irradianceGrid` takes 8 to 40 nodes, so a value that passes validation always fits the 65,536 node cap; 41 to 48 is no longer accepted (use `true`, which still reaches 48 along the longest side). The recipe bake also honours the sun link `environment.keyAffects` (#1218)+- Breaking: an HDRI (a file, `environment.hdri`, a bundled look) now faces as in three.js and Cycles, a quarter turn from before; add `Math.PI / 2` to `environmentRotation` to get an old scene's look back (#1371)+- Breaking: a grid volume's `blackbodyIntensity` now gives the physical brightness and Planck colour of Blender's Principled Volume, so existing grid fires render 3 to 4 times dimmer; raise `blackbodyIntensity` or the exposure to get the old look back (#1733)+- A local volume that emits now feeds bloom in the rasterizer: the volumes are drawn before bloom extraction for a frame with bloom on (#1046).+- A local volume takes `emissionRamp`, a blackbody (`kelvin`) or a colour ramp (`stops`) read by density, so one fireball grades from a white core to an orange edge in the rasterizer and the CPU path tracer (#1046)+- A local volume with `lightsScene: true` adds a point light at its emission, so a fireball lights the floor and props around it in the rasterizer (#1046).+- A scattering local volume takes `scatterOctaves` from 1 to 4, the octave sum of `volumeFog` for the key light, in the rasterizer and the CPU path tracer; 1 draws as before (#1047).+- A box or sphere local volume can shade the opaque surfaces under it from the key light, in the rasterizer and the CPU path tracer, when `castShadow: true` is set; it is off by default so the default frame does not change (#1047).+- A scattering box or sphere volume with `shadowDetail: true` shades itself by its own noisy density along the sun, in the rasterizer and the CPU path tracer (#1047).+- `backend: "webgpu"` draws box and sphere `environment.localVolumes` (smoke, glow, fire) at the same look as the CPU tracer, 23 times faster at 480 by 300 and 13 times at 1080p (#1048)+- `scene_apply` `add-explosion` writes a whole blast into plain recipe sections from one call: a gas-sim fireball and smoke volume with a white-hot core that lights the floor, sparks, flash, shock ring, dust, rubble, scorch, camera shake and optional fractured debris; particle systems take a `delay`, and exported recipes and render jobs need no extra files (#1050)+- A scene node takes `rayVisibility` in `scene_create` and `scene_apply`, with `'glass'` as a shorthand that lets shadow rays through a window pane so the path tracer samples sun and sky through it; emitters hidden from bounce rays stay refused, with the reason in `docs/path-tracer.md` (#1081)+- The showcase page shows three Cook path-traced stills with their credits (#1083)+- Cook reads a tangent-space normal map as Cycles' Normal Map node does, so a normal-mapped mesh on a non-uniformly scaled object no longer renders brighter or darker than in Cycles, and `render-cycles.py` builds a Normal Map node for parity scenes (#1111)+- Native film frames get their 8-bit picture from Cook's GPU post stage and a faster PNG encoder, taking a 1080p frame from traced film to PNG in 22 ms instead of 276 ms (#1123)+- Ambient occlusion no longer draws vertical bands on flat walls seen at a slant (#1126)+ +- Guides tag fields newer than the last npm release with {since X.Y.Z}, a test fails when one is left untagged, and an MCP refusal of an unknown field names the installed version (#1131)+- `parseGaussianSplatSpz` and `parseGaussianSplatSpzAsync` read gzip SPZ splat files (versions 2 and 3, as written by Niantic tools and World Labs Atlas) into the same data as the PLY parser, with the same limits (#1140)+- `parseGaussianSplatPlyAsync` and `createGaussianSplatsAsync` load large splat files in slices with progress and abort, so the page keeps responding (#1142)+- The WebGPU path tracer samples the sky through light portals (areaLights with portal: true) with the same density as the CPU tracer and Cook; only the WebGL tracer still ignores them (#1144).+- The CPU path tracer samples the sky through light portals (areaLights with portal: true), so a room lit through a window converges faster at the same mean; the WebGL and WebGPU tracers still ignore them (#1144).+- The Blender add-on and the Cycles parity bridge now multiply a light's energy by its Emission node's constant Strength and its colour by a constant colour or a constant-temperature Blackbody, and report other light node setups as not reproduced with the light name (#1153).+- A chunked native trace packs up to the unique triangles the adapter reports (`maxUniqueTriangles`, 4.79 million on a 12 GB RTX 3060) instead of a fixed 4,000,000, up to a ceiling of 16,000,000, and a scene over the bound is refused with the count, the limit and how many triangles too many (#1158)+- A native render job and `render:file-cook` check a scene against the adapter capacity before exporting it, fail with one error naming the count, the limit and what to drop, and report the capacity used (#1158)+- A native trace of zero parts is refused with a message that says to pass a part, and the docs name the tracers that still keep 1,024 parts (#1158)+- A native trace takes `maximumParts` (up to 1,000,000) instead of 1,024 parts or distinct geometries, so scenes of thousands of separate meshes export without merging, and a scene past it is refused with the count and what to do (#1158)+- The native `cycles` tracer now starts camera rays at `camera.clipStart`, so a wall between the eye and the clip start is not drawn, and `report.ignored` no longer lists it (#1159).+- Add `camera.clipStart` to the CPU and WebGPU path tracers, so camera rays start that far along the view axis and a wall between the eye and the clip start is not drawn; the WebGL tracer and native kernel name it in `report.ignored` (#1159).+- MCP render jobs, scene_create and scene_apply draw grid volumes (`.67vg` files in the grid folder, still or a sequence): the server hands the grid bytes to the preview page by name (#1161)+- `generateGeometryLods` keeps uvs, colours and normals, so a textured part with levels draws, and `addLodsToDenseParts(parts, { pixels })` gives levels to the big parts of a scene with more than one triangle per pixel (#1163)+- `denseSamples` is judged on the same part list at construction, on resize and after `replaceParts` (a scene near the threshold no longer flips on its first resize because the ground was counted, and a scene that gains or loses triangles is re-judged) (#1163).+- `denseSamples` renders scenes with more than one triangle per pixel with fewer MSAA samples (default off), which took a 1.6 million triangle frame from about 3.7 to 3.0 ms at 2 samples (#1163).+- The colour shader computes the sun shadow coordinate from the world position instead of carrying it as a varying, which cuts about 15 percent of the opaque pass on scans with about one triangle per pixel (#1163)+- The Cook path tracer matches Cycles more closely: bokeh corners sit at Cycles' blade rotation, metals and glass mixes keep their energy, and soft shadows from many lights clean up faster (#1172)+- A glTF with its base color map on `TEXCOORD_1` beside maps on `TEXCOORD_0` keeps its base color: the base color map, layer maps and shadow alpha pass now read the second UV set when the map asks for it (#1175)+- A glTF map whose `texCoord` is above 1 (a normal map on a third UV set, for example) is dropped with a warning naming the material instead of refusing the whole file; a normal map on `TEXCOORD_1` already loads (#1177)+- Flat water that transmits light (glTF `KHR_materials_transmission` with an `ior` near 1.33) now shows the water reflection of the scene around it (#1178)+- A glTF file with lights no longer gets the renderer's studio key and fill lights on top of its own: on a test scene mean luma went from 224 to 204, within 1 level of three.js with the same ACES curve (#1178)+- The home page hero says "67js is now Crucible" under the hero buttons, with the new package name and a link to the changelog (#1183)+- Native render jobs get 640 MiB of fixed memory allowance instead of 256, sized from measured native + OIDN video jobs, so a 960 x 540 film no longer runs within 50 MiB of the guard (#1186)+- A native render job on a software Vulkan adapter (lavapipe on a Linux runner without a GPU) gets 2304 MiB more memory allowance, so the first 960 x 540 job no longer fails the memory guard (#1186)+- Native path traced video jobs with OIDN get a memory allowance of 256 MiB plus 1 KiB a pixel, capped at 2 GiB, and the memory guard error names the limit and the observed and allowed MiB (#1186)+- The browser GPU path tracer (WebGL2 and WebGPU) takes as many triangles as the device's texture or buffer limits hold, up to 2,000,000, so a skinned crowd over 400,000 triangles can be traced; a scene past the limit is refused with its triangle count, the limit and what to do (#1188).+- The agent skills match the code again: path traced `stills` and `video` jobs, the 1920x1080 pixel cap, all pass names, and the mesh and walkmap tools (#1191).+- `scene_capabilities` reports `jobs.maximumFps` as 60 and `jobs.maximumFrames` from the same constants the render job validator uses, instead of a stale 30 (#1193)+- Fixed render jobs with passes (`depth`, `normal`, `albedo`, `objectId`, `emission`) failing with WebGL error 1282 on a scene with a smooth metal part: the reflection probe now looks up its uniforms again when a pass links a new shader program (#1196).+- `exportGlb` and `scene_export_glb` keep the JPEG, PNG or WebP bytes of an imported texture instead of re-encoding it as an uncompressed PNG, so an imported 1.8 MB catalog asset no longer exports at about 6.5 times its size (#1197)+- `scene_look` steps a busy PNG down by 0.75 until it fits the 1 MiB budget and reports the size it returned, and `closed-room-without-bake` no longer fires on a lone object whose bounds centre is inside it (#1197)+- The site wears the Crucible brand: the pot mark and Plus Jakarta Sans Bold wordmark, Paper and charcoal themes with a light and dark switch, and the shared brand fonts (#1199)+ +- A render job with `passes` no longer fails with GL error 1282 on a scene that has a baked irradiance grid: the grid looks its uniforms up again for the program a pass links (#1204).+- glTF accessors are now decoded one primitive at a time during conversion and released after their last use, and `maxDecodedBytes` sets the decoded-bytes budget (#1210)+- A native film or stills job waits one event-loop turn after each frame, so the addon's frame planes are freed and the Node process no longer grows about 9 to 13 MB per frame (#1211).+- A big thin ground box beside dense props no longer fails `scene_create` with "Set maximum size exceeded": the flush search stops at its tested-pair limit and reports the part as unchecked, and a 2 cm slab is no longer flagged `collapsed-node` (#1216)+- A placed asset material override (and a recipe material) takes an `emissionMask` ramp along an object axis, so a part glows only at one end while its texture and shading stay (#1217)+- A placed asset's `asset.materials[]` override can take an `id` that `motion` material tracks drive, so `emission` and `emissiveColor` fade a heat glow without rebuilding the scene (#1217).+- A `scene_apply` whose rebake cannot run (part cap, path work limit) now keeps the edit, drops the old bake and returns a `bake-skipped` warning with how to fit, and the irradiance grid and probe part errors say they trace on the CPU (#1218)+- A pose edit through `scene.apply` at 961 parts now takes about 2 ms instead of about 60 ms, and a game frame that sends the same poses as the last one skips `scene.apply` (#1219)+- A render job whose frame is cut to fit the render budget now names the budget the full frame needs and the `--memory` value to restart with, instead of only the smaller size (#1221)+- Line art can be set per node: `environment.lines.nodes` takes `{ "<node id>": { lines: false | silhouetteOnly, color } }`, and the `objectId` pass metrics carry `objectNodes` (#1223)+- The WebGL2 GPU path tracer draws every part again: on ANGLE D3D11 it drew only the first part of a scene, because a nested texel fetch read every later material as shadow-only (#1228)+- An `image` render job takes up to 3840x2160 pixels instead of 1920x1080, and an image frame the `--memory` budget would shrink is refused with the `--memory` it needs; stills and video keep 1920x1080 (#1241)+- The WebGPU real-time tracer and the GPU light bake bound their scene by the device's triangle limit, up to 2,000,000, like the still render, instead of 400,000 (#1242).+- The path tracers, native Cook included, read a map on `TEXCOORD_1` with the second UV set, so a glTF base color on `TEXCOORD_1` beside a `TEXCOORD_0` normal map no longer samples UV (0, 0); a part with a map on `uvChannel` 1 and no `geometry.lightUvs` is refused with what to set (#1243)+- `crucible-mcp` stops at startup with the real message for a bad `--packs`, `--assets`, `--audio`, `--ffmpeg`, `--browser` or `--branches`/`--relations`/`--findings` value, `scene_restore` rejects two targets, and terrain and geometry-budget errors name the range, id and numbers (#1244)+- Docs and llms.txt now say the `quality` render setting ships, the checkpoint limit is 16, and the tool-cost total points to `scripts/measure-tool-cost.mjs` instead of a stale number; a test keeps the limit in step (#1245).+- A failed job release, status write or browser launch no longer breaks the MCP runtime: a retry finishes the release, `close` always frees the store lock, and a recipe over the 64 MiB asset limit is refused before any file is read (#1246)+- Scene builders refuse bad input when it is set: text outside printable ASCII names the character and font, motion tracks are checked for out-of-range values, and a terrain `seed` must be a whole number from 0 to 2147483647 (#1247)+- glTF import names the field and index of a `null` or non-object entry in `nodes`, `meshes`, `accessors` and the other arrays instead of a "Cannot read properties" TypeError (#1248)+- glTF import and export reject more bad files by name: dangling indices name their field, meshopt buffer views count against `maxDecodedBytes`, and `exportGlb` throws when a light map or `uvChannel: 1` map has no `geometry.lightUvs` (#1248)+- The site build now publishes every guide that `llms.txt` and the other guides link to, the install guide no longer links a release the site dropped, the docs say the three optional packages are source-checkout only, and the addon check accepts release candidate versions (#1249)+- A render job's deadline counts render time only, its queue wait is bounded apart, and `scene_create` starts a fresh runtime after a failed one without a server restart (#1256)+- A render job that hits its elapsed deadline under a shared GPU now gives the scene page 120 seconds to come back, and its failure says to submit the job again (#1256, #1224).+- `crucible-mcp` takes `--gpu-queue` and `--gpu-queue-wait`: an optional queue command that looks, reviews, bakes and render jobs wait for before they take the GPU, so the server can share a card queue (#1258)+- The path tracer export keeps an area light aimed along `up` as a rectangle, reads texture `wrapS` and `wrapT` per axis on every tracer, and turns skinned normals correctly (#1259)+- A client or authority restarted on the same peer id now resyncs on its own instead of staying silent (#1262)+- `startGame` no longer moves a player or NPC under a translated parent on the first frame, and the `restart` rule action, the R key and the new `game.restart()` put the game back at its start, so a won or lost exported game can be played again (#1263)+- Render jobs count frames without float noise, `render-media` refuses audio shorter than the video, and a tiny `audio.offset` reaches ffmpeg as a plain decimal (#1264)+- `crucible-addon-check` resolves the entry like the server, rejects unsatisfiable peer ranges, accepts `--version=X` and has `--help`; `crucible-mcp --help` lists every flag (#1265)+- Disposing a renderer made with `deferProgramLink` now frees the programs still linking, and `resize()` on a lost context returns instead of throwing (#1266)+- A looping clip retargeted with `retargetClip` or exported from an authored rig now ends on its end pose instead of the first pose, and `createCrowd` throws a `RangeError` for bad `columns` and `size` (#1267)+- Volume fog detail fields, the volume fog density motion track (now 0 to 10), a mesh volume's missing part and a second grid volume fail when the scene or an edit is built, with the render-time messages (#1268)+- Mesh `bevel` caps `profile` at 0.99 and shrinks the overlap clamp for concave profiles, `inset_region` clamps thickness at the face limit and reports it, and `lintMesh` with `geometry: true` flags `folded-face` (#1269).+- The games runtime calls `onEvent` for every event and caps only the stored log, a rules part starts as hidden as the scene node it is named after unless it sets `visible`, and `reportContact` is ignored after the game is won or lost (#1281).+- Asset loading counts a textured pack GLB at its embedded size, and the PLY splat loader skips elements before `vertex`, accepts CRLF headers, and rejects non-finite scales and a non-positive `maximumCount` with a clear error (#1282).+- `remesh` no longer returns an inflated or non-manifold shell for a cube, `remesh` and `boolean` reject an inside-out mesh and name `flip_normals`, and `subdivide` no longer fails on a loose vertex (#1283)+- `render-media` keeps the audio on the browser-encoded H.264 path and ends PCM and AAC audio with the picture, fails a video that lost its audio, writes 8 distinct stills as `frame-0` to `frame-7`, and rounds a derived video size to even (#1284)+- `remesh` no longer returns an edge shared by four quads on thin, bumpy surfaces; when it cannot separate the sheets, it throws and names a different `voxelSize` (#1291)+- The WebGL and WebGPU path tracers read each map of a material with its own UV set, so a floor with a `TEXCOORD_0` base colour and a `TEXCOORD_1` emissive map matches the CPU tracer; a device with limited texture or buffer size now holds about 10 percent fewer triangles (#1300)+- An `sdf` geometry now splits saddle cell vertices the way `remesh` does, so every edge joins exactly two quads when parts touch along an edge; a saddle it cannot separate throws an error that names the fix (#1302).+- A placed asset takes `flatten: { color?, unlit? }`, which sets one flat colour or unlit on every material of that asset in one call and leaves the rest of the scene unchanged (#1303)+- The smokes `film-timeline`, `object-motion-blur` and `path-tracer` ask `scene_render_result` for the detailed answer, and the render docs and skill say file digests come with `response_format: "detailed"` (#1304).+- `scene_create` and `set-environment` accept everything `scene_export_recipe` writes, including all joint fields and motion track `extrapolation`, and a malformed motion track names its bad field and the allowed ones (#1311)+- The shipped guides and the build-crucible-scene skill now import only public subpaths, and new addon entries `crucible/addons/reflection-probes`, `fracture`, `morph` and `skinning` back the examples that had none (#1312)+- Memory and size limits now agree with their messages: an `.hdr` over 8 MiB is refused like a `.cube` over 4 MiB, and `scene_asset_catalog` skips an oversized `.hdr`, `.cube` or image with a reason in `skipped` instead of failing the listing (#1313)+- Pack zips and the `sha256` in `/packs/index.json` no longer change with the checkout's line endings: `.gitattributes` pins text files to LF, and `createZip` turns CRLF into LF in text entries, so a Windows checkout and CI publish the same archive hash for the same SHA (#1314).+- An image whose size cannot be read from its header is refused before decoding with an error that names the file and the frame limits, and a JPEG with a restart marker before its frame header can no longer slip past the size limit (#1319)+- MCP errors name what to change: a review camera lists the fields it rejects, a stale revision gives the current one, a path traced job far over the 400000 triangle cap of the cpu tracer is refused at `scene_render_submit` with the count, and an unknown `scene_capabilities` topic suggests the nearest one (#1325)+- The docs and the render jobs skill now say `npm install @codepawlhq/crucible-denoise-oidn` instead of "not on npm yet", and the guides document `jobs.maximumFps`, `jobs.maximumFrames`, `emissionFresnel` and the adaptive resolution `raiseBelow` option (#1327)+- `@codepawlhq/crucible-denoise-oidn` is now published to npm beside Crucible, so OIDN installs with an npm install of Crucible (#1327)+- The website matches codepawl.com: light by default with a dark mode on the header button, a plain header bar, captions under the showcase pictures, 8 px controls and the pot's metal as the only orange (#1333)+- A closed room's walls, ceiling and props now get the sky they can see traced into their vertices, so a sealed room renders near black like the path tracer instead of as bright as a room with a window (#1334).+- The wavelet denoiser leaves far less blotchy grain on flat walls at 256 samples: its luminance guide ignores a lone noisy pixel (#1335)+- A recipe path such as `environment.lightDirection` written as a motion track (`target`, `keys`) now fails with an error that names the `sun` track (`direction`, `color`, `intensity`) that moves the key light under any sky; the docs say so (#1342).+- The flush-face search keeps its triangles in flat arrays instead of one object each, so a 200k triangle GLB peaks at about half the browser heap while a scene loads, and `scripts/measure-glb-heap.mjs` measures it (#1343)+- `loadGltf` keeps a `KHR_materials_ior` above 2.5 (exporters write 1000) at 2.5 with a warning instead of refusing the file (#1347)+- The WebGPU path tracer megakernel now compiles with FXC, so the bundled Chromium on Windows can run the GPU tests, and the render stays byte for byte the same on DXC (#1350).+- A camera `clipStart` at or past `far` is refused with a message that says what to pass; the MCP preview's native export and both real-time tracers now report that they ignore `clipStart` (#1351)+- A recipe scene that leaves `shadow` out and is wider than the default 7 m view now gets a sun shadow view fitted to its bounds (three automatic cascades past 40 m), so a sealed room stops leaking sun; a recipe that sets `shadow` is unchanged (#1359)+- Look change: the light an HDRI gives a surface is now an exact cosine convolution of the map, so shade next to a strong sun no longer comes out up to 50 percent off and matches three.js within 5 percent (#1374)+- The site header gives Marketplace an icon, live-example code sits on a tinted surface, and the sliders in docs examples show their value with brand colours and focus states (#1375)+- Glass that transmits light now reflects the reflection probe placed on it, and a blended part keeps its specular at full strength when its opacity drops, so a glass wall and a faint blend glass show the room and the sky instead of fading with alpha (#1385)+- A rules input trigger on an action the game does not bind (anything but jump or run) is now refused by set-game with the actions to use, instead of crashing startGame on its first step (#1387)+- Game rules arithmetic whose result overflows or is not a number now gives 0 instead of leaving a variable at Infinity or NaN (#1388)+- Live physics fixedUpdate rejects a NaN step, and raycast, shapeCast and overlap give the query options error for null, both naming what to pass (#1389)+- A film timeline frame that lands on a shot cut now shows the later shot even when shot lengths such as 0.1 s and 0.2 s sum with float error (#1393)+- A particle emitter from a scaled mesh now emits from vertices at the stated `normalSpeed` and orients instances with unit normals (#1396)+- A sound with `pitchVariance` no longer throws a range error from `play` at low pitches (#1397)+- `motionBlur.objects` now rejects negative, NaN and text values instead of ignoring them (#1398)+- A curve through a repeated point, with a duplicated end control point or with repeated trailing points (poly and NURBS) now builds instead of throwing "Cannot read properties of null"; the direction there comes from the nearest distinct point (#1399)+- A glTF accessor or sparse block with a negative or fractional byteOffset now fails with an error naming byteOffset instead of reading the wrong bytes (#1400).+- A glTF accessor that sets normalized on floats or 32-bit integers now keeps its values instead of decoding to NaN (#1401).+- Library calls `bevelEdges`, `subdivideMesh`, `loopCut`, `bisectPlane`, `resampleCurve`, `resampleSamples` and `filletCurve` refuse NaN, missing and unbounded inputs with an error that names the parameter, the value and the allowed range, and share their limits with the recipe layer (#1405)+- Make `fitOrthographicHeight` throw for a view whose up is parallel to the view direction, whose eye equals its target or whose vectors are not finite, instead of returning 0 or a meaningless height (#1406)+- Make `createCameraSampler` reject empty frames, a non-positive fps, malformed frames and a non-finite sample time with errors that say what to pass (#1407)+- A glTF node with a `matrix`, `translation`, `rotation` or `scale` of the wrong shape, or a zero quaternion, is refused with the field named instead of reported as a singular node (#1409)+- A glTF data URI accepts a capitalised `;BASE64` marker, and a bad payload says to use standard base64 instead of throwing a bare decoder message (#1410)+- A glTF TRIANGLE_STRIP, TRIANGLE_FAN or unknown primitive mode is refused with its name, mesh and primitive and the fix (#1411)+- Editing an imported mesh keeps the UVs of a face that has a real corner at UV 0,0 and rebuilds only faces with no UVs (#1416).+- `scatterVegetation`, `createGrassClump` and `createFlowerPatch` check their options and throw a RangeError with a valid example instead of crashing, hanging or running out of memory (#1417).+- The MCP catalog decode refuses an external glTF buffer or image uri instead of fetching it, with an error that names the uri (#1418).+- Fix the material snapshot so a fast-path update with a changed `emissionFresnel` or `emissionMask` throws instead of being ignored (#1419).+- Terrain ground normals on the outer border now use the real span between samples, `height` accepts plain and Float64 arrays and rejects other non-object values, and `update` rejects a non-finite eye position (#1420)+- GLB export now rejects lightUvs of the wrong length and empty, non-finite or unsorted animation keyframes with errors that name the fix, and writes a Float32Array light map tint as a JSON array (#1425, #1426, #1428).+- Navmesh and crowd bake refuse a reversed obstacle footprint, non-finite or out-of-range geometry (#1427, #1429) and a palette of more than 7200 frames (#1430), instead of building a silently wrong or oversized result.+- Cloth `step` takes zero time as a no-op and rejects a negative, NaN or over one second `deltaSeconds`, `bakeCloth` rejects a bad `frameCount` or `framesPerSecond`, `sampleClothFrames` rejects a NaN time, and `wakeTrail` no longer returns NaN points for a zero lifetime or a missing speed (#1431)+- Pose solvers refuse a zero aim `forward` or two-bone `bendToward`, and a morph-only layer refuses a non-finite `reference`, instead of silently doing nothing or writing NaN weights (#1436, #1437, #1438)+- GLB export no longer crashes on a skinned part without a matrix, folds an emission below 1 into emissiveFactor instead of an invalid emissiveStrength, and keeps the negative scale of an animated mirrored node (#1441, #1442, #1443).+- A scene query now hits a part after its world matrix moves in place, `raycast` throws on a NaN or non-number `maxDistance` instead of reporting a miss, and a rig with a zero-scale `rootMatrix` throws instead of filling the pose with NaN (#1444, #1445, #1446).+- Fix `trianglesIntersect` for a zero-area first triangle, `floatToHalf` rounding of subnormal halves, and make `RenderTargetLedger.record` throw on an unknown format or bad size instead of reporting NaN bytes (#1450, #1448, #1449).+- A light linked to a group node is now reported as an unknown name instead of lighting nothing (#1453)+- Kernel offload no longer reuses another call's cached result for look-alike inputs, no longer hangs when a worker answer cannot be read, and runs a self-referencing input inline (#1455, #1456, #1457).+- Fix three silent faults in the fog and local volume code: a moved fog box leaks its old sky texture, the shadow ambient cache ignores the measure it is given, and packing accepts a grid volume that was never bound (#1463)+- `scene_apply` fracture-node on a node that already has a physics body now names `node` and says to drop that body with set-physics first, instead of blaming `physics.bodies[n].node` after the node was replaced (#1466).+- A path-traced scene with a negative, fractional or missing index, or a NaN in uvs or colors, and a real-time tracer with a bad `historyLimit`, `filterIterations`, `cacheCellSize` or `resolutionDivisor`, now stop with a message that names the part or option (#1467, #1461, #1462)+- A pack JSON string that a browser reads as a remote address, such as a backslash form or a scheme with a tab inside, is now refused (#1468)+- An addon file in a folder whose name starts with two dots is no longer refused as outside its package (#1469)+- The pack index build refuses an `--out` folder that overlaps `--packs` instead of deleting the packs (#1470)+- A `.cube` look with a bad DOMAIN line or a malformed row is refused, and a NaN or infinite channel gives a finite display colour instead of a NaN pixel (#1476, #1475).+- A zero-size fog box no longer gets NaN sky grid cells, tree sky matching no longer writes NaN when leaf coverage is zero, and an empty crown says what to pass (#1477, #1478, #1479).+- Closed-mesh check for `softBody.pressure` now rejects a triangle wound the wrong way, and liquid optics keep the default for an undefined field and reject an unknown field (#1483)+- Compressed textures with a zero, negative or fractional base size or a mip chain whose levels do not halve, and texture sequences given a NaN or infinite time, now throw an error that says what to pass instead of failing later (#1489, #1490).+- Name the track and keyframe in motion validation errors for a non-array value, repeated times and an empty or malformed keyframe list (#1492, #1493, #1494).+- Object reflections are traced again when only `environment.sky` changes (#1495)+- A `meshEdits` entry that is not an object is reported as that entry instead of a character index (#1496)+- `lintMesh` with `geometry: true` no longer reports a false self-intersection on a prism over a concave outline, and the `occlusion` measure of `vertex_group` skips zero-area faces instead of throwing (#1497, #1498)+- A drive wheel that gives both `node` and `nodes` is rejected with a field error instead of silently ignoring `node` (#1502)+- `exportGlb` throws on texture sampler wrap or filter values that are not glTF enums and on a `boxMapAxis` outside 0, 1 and 2, and `exportObj` throws when `geometry.uvs` or `geometry.normals` does not match the vertex count, instead of writing an invalid file or dropping the data (#1503, #1505)+- A part that the flush pass ranks and that carries baked `lightUvs`, `skyVisibility` or `surfaceMasks` no longer reaches the path tracer with NaN lightmap coordinates (the demo room window-top had 507 of them), and the path tracer now refuses non-finite `lightUvs` (#1504)+- A mesh particle emitter no longer emits from a triangle or edge outside its vertex group when the random draw is exactly 0 (#1511)+- `maximumInstances` and `maximumVoices` must be whole numbers; a fractional limit used to let one more voice play (#1512)+- `markersCrossed` and `emitMarkers` throw for a NaN or missing clip time instead of firing no cues (#1513)+- A looping clip or a repeating cycles modifier sampled on a cycle start at a fractional frame rate, such as frame 9 of 30 fps on a 0.1 s clip, now shows the first pose instead of the last (#1514, #1515)+- An animated vehicle player now defaults to the speed drive and refuses an intentSpeed drive at set-game, instead of crashing on the first step (#1517); a physics listener that disposes the system no longer lets later listeners run on the freed world (#1518).+- `createTerrain` rejects a non-positive or non-finite `lodDistance`, and `scatterVegetation` rejects a bad `scale` maximum or an infinite `mix` weight, instead of silently building empty terrain or NaN instances (#1519, #1520, #1521).+- `crucible-addon-check` now fails a `package.json` that is not a JSON object, refuses a repeated `--version`, and fails an entry that is a folder (#1524, #1525, #1526).+- A shadow casting light with a range of 0.05 or less no longer writes a NaN or flipped shadow matrix, and cascaded shadows keep their splits ordered when `maximumDistance` is nearer than the camera near plane (#1530, #1531).+- Calling `createNetcode(...).init` twice without `dispose` no longer leaves the first transport subscription live, so messages are not handled twice (#1532).+- A `game.restart()` whose physics rebuild fails now sets `game.physicsError` and makes the physics step throw the cause, instead of leaving a silent dead world (#1533).+- A camera with an `fStop` now focuses the raster depth of field at `camera.focusDistance` when it is set (the path tracer and Blender already did), and a non-positive `fStop` throws (#1536)+- The real-time path tracer refuses a `width` or `height` that is not a whole number of pixels, and names `width` and `resolutionDivisor` when the traced size leaves 16 to 4096 (#1538)+- `edgeAngle` reads 0 next to a zero-area face, mesh volumes measure the positions and indices a part holds now, and empty or NaN positions get a clear error (#1539, #1540, #1541)+- A cached modifier stack result is checked against the current vertex and face limits, and realized instance geometry keeps one geometry record per distinct content, so the same scene builds to the same recipe (#1545, #1546).+- `knife_project` no longer fails on a line along existing edges, `grid_fill` handles long loops within a vertex budget, and `fill_holes` stops reporting an Infinity limit (#1547, #1548, #1549)+- `meshFromGeometry` refuses uvs or colors of the wrong length, a face tag with values above 1 now combines correctly in `selectFaces`, and a sculpt edit with non-array strokes names the field (#1552, #1553, #1554)+- The splat PLY loader drops a splat with any non-finite field (a scale that overflows included, which used to throw) and reports `skipped: { count, fields }`, throwing only when no splat is usable; createGaussianSplats refuses a covariance that overflows 32-bit floats, naming the splat (#1558)+- A glTF sparse accessor with a fractional, negative, missing or too large count, or without indices or values, now fails with an error naming sparse.count, sparse.indices or sparse.values instead of a typed array error or a silent skip (#1563).+- Light profiles with a near zero axis value, IES headers with a fractional angle count and GPU bakes of spot cones narrower than cosine precision no longer give Infinity, a wrong error or a NaN edge (#1566, #1567, #1568).+- An animation graph state with `loop: false` over a looping clip now holds the clip's last frame instead of snapping back to its first frame (#1569)+- glTF import normalizes the direction of a scaled spot or directional light, skips one scaled to a zero direction with a warning that names the node, and ignores a non-numeric light range with a warning (#1573)+- createCameraSampler rejects a frame whose eye equals its target and an fps that makes the duration infinite (#1574)+- A kernel worker request with an unregistered name (including `toString` or `constructor`) now answers an unknown kernel error, `encodeVolumeGrid` refuses channel names and sizes that `decodeVolumeGrid` rejects, and a grid header that is not a JSON object says so (#1575, #1576, #1577).+- Vegetation scatter errors and clearings are exact: an area off the ground names `vegetation.scatter.area`, a node exclusion follows the node's world position, and vertex group errors in particles, cloth and exclusions name their list index (#1579, #1580, #1581)+- A motion track on the environment kind with keyIntensity, keyColor or keyDirection now says the key light is the sun kind and names the sun track to use (#1585)+- Refuse mismatched OIDN buffers, release the OrbitControls listeners on dispose and refuse short or out-of-range geometry in toThreeObject (#1586, #1587, #1588).+- The Blender add-on keeps a Roughness, Metallic or Alpha constant when its image has no pixels instead of exporting an empty texture, and notes an anamorphic aperture ratio it cannot render (#1589, #1591).+- Terrain hole errors and `scene_inspect` path query errors now name the exact field (`terrain.holes[1].radius`, `path.from`, `path.avoid[1]`) instead of a bare message or a generic `id` (#1592)+- glTF import clamps negative skin weights to 0 and renormalises each vertex (an all-negative vertex binds to joint 0), with one warning per primitive that counts them, instead of refusing the file (#1595)+- Grass and clutter fields refuse an area that needs over 262144 tiles, `filletCurve` rejects an arc segment count of 0, a fraction or over 64, and field `div` and `mod` give 0 for a divisor so small the result overflows (#1596, #1597, #1598)+- three-compat sums several DirectionalLights, honours every rotation.order and rejects a torus radius of 0; three-bridge clamps an InstancedMesh count to its matrix array with a warning (#1599)+- A vegetation exclusion that mixes min/max with center, radius or node, a hair `swayRate` without `sway`, and a `radius` on a box particle emitter are now refused with an error that names the fix instead of being ignored (#1603, #1604, #1605).+- The Blender add-on exports the camera's own clip end, says so when every mesh has no triangles, expands `//` paths for the tracer folder, and refuses a NaN or infinite environment cap (#1608, #1609, #1610, #1611).+- `triangulate`, `createStreaming`, `createInput` and the crowd `setTimes` refuse bad input with a named error, `crowdFrameBlend` shows the first frame for a non-finite time, and input `dispose` finishes every source before rethrowing (#1617, #1618, #1619, #1620)+- Physics collider limit errors name the node by its list index, `nodes[1].geometry`, instead of putting the node ID in the brackets (#1622)+- A crowd `farDistance` is measured in the space of the crowd instances, so a crowd under a moved parent switches to the far rate by its real distance from the camera (#1623)+- The live runtime retries a system whose init threw, keeps time finite after one NaN clock reading, and an arc shape refuses a span wider than half a turn (#1625, #1626, #1627).+- Look up a part's matrix by the part before its name in the closed-room, sky openness, ground bounce and sun bounce tests, so glTF parts that share a name keep their own matrix (#1628)+- Crowd `farDistance` follows a parent group that has a motion track: the camera distance is measured from where the parent is at that time, not where it was when the scene was built (#1629)+- Mesh normals: weighted_normal falls back to the face normal, then +Z, for a zero-area face instead of writing a zero normal (#1631), and cornerNormals refuses a NaN, negative or oversized sharpAngle (#1632).+- A ragdoll rig with a later parent, a repeated bone id or a bone that is its own parent is refused when built, camera rigs ignore a time step that is not above zero, and `buildSprayGeometry` refuses a bad `maxPoints` or `radius` (#1633, #1634, #1636)+- `projectUv`, `packUvIslands`, `unwrapUv` and `bisectPlane` refuse an unknown mode or axis, a margin outside 0 to 0.2 and a bad distance with an error that names the field (#1638, #1640, #1641).+- Fix three edge cases from a bug hunt: the render target ledger refuses negative sizes, assembly layouts ignore inherited object keys, and a clutter mask returning NaN places nothing (#1642, #1643, #1644).+- A sealed room whose glTF parts share one name is now recognised as a closed room by the floor sky trace, and object motion blur gives each part of a shared name its own motion (#1645, #1646)+- Radiance .hdr files in the original format, where a 1,1,1 pixel repeats the previous pixel, now load instead of reading wrong or failing as truncated (#1650)+- `createClutter` refuses a seed that is not a finite number, as grass and terrain already do, instead of scattering one fixed layout or joining a string seed into its hashes (#1652).+- Refuse a NaN, Infinity or non-number seed in `createGrass` and `createTerrain` (noise height), and a non-finite budget in `createSlicer`, with errors that name the field; fractional seeds still truncate (#1652)+- Fix unsubdivide losing every face tag, and boolean cutterFaces counting faces of earlier booleans when a cutTag is reused (#1653)+- Spume whitewater skips points with NaN or infinite values and `partsForFrame` clamps an out-of-range frame instead of throwing (#1655).+- Hair roots stay finite when a surface vertex has a NaN or infinite position or normal (#1656).+- Game setup no longer hangs on scene nodes whose parents form a cycle, the NPC path follower system skips the loop's zero-length first frame, and `createLookAt` and `createFootPlacement` ignore a non-finite pose time instead of staying NaN (#1660, #1661, #1662)+- `uploadTextureInSlices` refuses short RGBA data and a zero size before any GL call, and `effectsWithFStop` accepts a camera without a target like the renderer does (#1664, #1663)+- Refuse a NaN or out-of-range `minVertices` and `angle` in `triangulateFaces` and `splitSharpEdges` when called directly, instead of returning an unchanged mesh (#1666)+- Fixed three agent-facing bugs: a branch snapshot now copies the forked recipe instead of sharing it, a tail light with a non-numeric `off` or `on` emission is refused instead of driving NaN, and an image sequence time that overflows frame arithmetic is refused instead of picking frame NaN (#1669)+- Refuse an irradiance grid resolution under 2 or samples under 1, an odd probe map width and an unreachable adaptive `raiseBelow`, instead of writing NaN grids, dropping probe rows or pinning the resolution at its minimum (#1672)+- A non-finite look-at target or foot groundOffset no longer throws each frame: it keeps the last good value and warns once per controller (#1675)+- `flatten` without a normal now refuses a closed mesh instead of squashing it onto a random plane (#1679)+- `buildWithKernels` no longer hangs when its kernel worker goes silent: a new `kernelTimeoutMilliseconds` option (default 60000, above 0, at most 3600000) ends the worker and finishes the build inline (#1681).+- Fix three LOD input bugs: `generateMeshLods` gives `relativeError` 0 instead of NaN for a zero-extent mesh (#1682), `addLodsToDenseParts` checks `ratios` on a sparse scene too (#1683), and `generateAssetLods` names a part that has no geometry (#1685).+- split_edges and delete_loose (and the other edits that rebuild from polygons) now keep vertex groups and other point attributes: a split vertex copies its source vertex's values and removed vertices drop theirs (#1684)+- `box()` now refuses a bevel too small to move its core instead of returning NaN normals, and `text()` draws a glyph whose overlapping contours start inside each other instead of throwing a TypeError (#1686, #1687)+- A world renderer with no `shadowBounds` target or extent fits the sun shadow view to its shadow casters, so the shadow bias stays thinner than 0.2 m walls and a closed room no longer shows sun lines along its seams (#1695).+- The MCP server decodes Draco-compressed GLBs in the asset catalog and packs with the optional `draco3d` package, so an agent can place a Sketchfab or NASA download as it is (#1698)+- `subdivide` and `loop_cut` keep point attributes such as vertex groups: new points blend their sources, and integer values take the first one (#1704)+- The MCP server transcodes KTX2 textures of catalog GLBs to PNG with the optional `three` package, so an agent can place an optimised download as it is (#1709)+- Add focusDistance (metres, with fStop) to the recipe camera so an agent can focus on a subject without moving the camera target (#1711)+- Dispersive glass keeps signed spectral weights as Cycles does, so prism fringes match its saturation, and negative pixels are clamped before OIDN and the display transform (#1714)+- Glass can fringe into colour: a material `dispersion` from 0 to 10 reaches the native path tracer, and glTF `KHR_materials_dispersion` is read instead of ignored (#1714)+- Two soft volumes that overlap are marched together, so a soft edge no longer shows a hard slab where the other volume starts (#1722)+- `scene_look` no longer fails on a mesh local volume that names a node added by `scene_apply`: the renderer now keeps its part list after a topology change, which also keeps volume shadow ambient and irradiance grid fog current after an edit (#1723).+- Local volumes take `noiseOctaves` (1 to 5, default 1): finer octaves of noise on top of the large lumps give smoke and fire a ragged edge, in the rasterizer and the CPU path tracer (#1725)+- Native video jobs that blend frames over time trace natively again: the temporal blend accepted only a 320-byte kernel parameter block, but Cook sends 336 since its anamorphic aperture ratio, so every such job fell back to another backend on its first frame (#1729).+- `bakeGasSim` takes `detail` (0 to 1, default 0): a noise upres of each frame's smoke density, so a coarse sim breaks into puffs and gaps below its cell size; 0 keeps the bytes of every existing bake (#1735)+- The gas-sim bake keeps smoke and fuel mass constant between sources and sinks: advection no longer swells the plume (a 3 s blast used to roughly double its smoke) (#1735).+- `bakeGasSim` bakes an explosion from a grid gas simulation (advection, pressure projection, vorticity, fuel burning into flame and heat) into grid volume frames, in `src/gas-sim.js` (#1735)+- add-explosion fireball emission is 9.5 instead of 14 at the reference size, so the core no longer clips to white and the edge falloff reaches 0.79 of the Cycles frame on a flat background, with the RMSE against it down from 0.030 to 0.020 (#1736)+- add-explosion bakes with turbulence 1 and vorticity 40, so the fireball outline breaks up more (silhouette roughness 5.4 to 6.1 and edge falloff 0.74 to 0.82 against a Cycles reference 10.1 and 1.06); `scripts/measure-explosion-breakup.mjs` scores a blast against it (#1736)+- `vdb-to-grid.py` reads Blosc compressed VDBs, openvdb's default, when run with `uv run --no-project --with blosc` (#1742)+- `createGaussianSplats` takes `palette` (2 to 256 colours) and `dither` (`bayer4` or `bayer8`) to snap `alphaCut` splats to a palette for a pixel-art look (#1744)+- `createGaussianSplats` takes `alphaCut` to draw opaque, unblended discs and `parseGaussianSplatPly` takes `upAxis: 'z'` to load Z-up files such as Blender splats upright (#1744)+- A grid volume with `castShadow: true` now shades the surfaces under it from the sun, so an explosion no longer floats over a lit floor (#1748)+- `render()` takes `linearOutput: true` with a float `outputFramebuffer` to write linear scene radiance after fog and volumes, with no view transform or clamp, and `compare-volume-render.mjs` reads it as a `.f32` candidate (#1749)+- Grid fires no longer show vertical streaks, cross lines between voxels or a hard flat bottom edge in Glaze and the CPU path tracer (#1752)+- An agent can place a part by its top, bottom or centre face with a new `level` node relation (`{ "kind": "level", "at": "top", "y": 0.5 }`), so it no longer adds half the thickness by hand. Existing recipes are unchanged. (#1755)+- Particle systems of `type: "sparks"` throw emissive streaks that cool along the blackbody curve with gravity and drag, and `preset: "explosion"` makes a blast (#1758)+- A grid volume blurs along the gas velocity when the scene's motion blur shutter is on: bake with `velocity: true` and a fireball or smoke smears along its expansion in Glaze and the CPU path tracer, as in Cycles (#1759)+- The shadow a grid volume casts on surfaces blurs along the gas velocity when the motion blur shutter is on, so a blurred blast no longer casts a sharp shadow (#1763)+- A render job or still no longer draws its first frame with the fallback post program while the driver builds the real one, so a cold first frame matches a warm one (#1767)+- The fireball bake that every explosion scene sends with `scene_create` runs about a third to a half faster (72 frames at 32 cubed, same bytes): the gas solver locates each traced position once per step instead of once per field (#1768)+- The MCP server keeps baked add-explosion frames in a cache folder, so a second process that needs the same blast loads it instead of simulating for about 5 seconds (#1768)+- A `scene_create` whose recipe has a grid volume may run 60 s instead of 30 s, the same limit `add-explosion` already had, because the first draw compiles its shader in the graphics driver (#1768)+- The MCP server keeps Chrome's compiled GPU shaders between runs, so the first volume `scene_create` of a new server process takes about 0.8 s instead of 4.9 s; `CRUCIBLE_SHADER_CACHE=0` turns it off (#1768)+- A scene with only grid volumes, or only box and sphere volumes, builds the shader of that kind alone, about half the shader build of the combined program, with the same pixels (#1768)+- A glowing or shadow-casting grid volume no longer stalls the first draw: its shader builds in about 8 s instead of 45 s on ANGLE d3d11, and a cold `scene_create` fits well inside the 30 s preview limit (#1768)+- Environment variables now take the `CRUCIBLE_` name first (`CRUCIBLE_NO_UPDATE_CHECK`, `CRUCIBLE_MCP_LOG`); the `SIXTYSEVENJS_` names keep working as aliases (#1770)+- `scene_apply` refuses an add-node that would take the scene's unique assets past the 64 MiB combined limit, with a clear message, instead of letting the later render job crash the page ('Target page closed') (#1771)+- `add-explosion` takes `sparks` and `motionBlur` options, on by default: spark streaks and a velocity motion blur of the fireball (#1775)+- `crucible-mcp --memory` goes up to half of the machine RAM (at most 4096 MiB) and the render job memory guard and the native tracer allowance follow it, so 1080p and native path trace jobs of heavier scenes finish (#1776, #1696)+- Particle emitter systems take `start` (the same field as `delay`) and `stop` in seconds, so a system with a `rate` can begin partway through a film and stop before it ends; nothing is born at or after `stop`, and systems without them are unchanged (#1777)+- The WebGPU path tracer now draws local volumes with `emissionRamp` and `noiseOctaves`, matching the CPU tracer; before, backend "webgpu" threw for both (#1785).+- The WebGPU path tracer now draws grid volumes (the #1735 blast, smoke and fire baked to a grid): density, flame, baked blackbody and, with `volumeShutter`, the motion blur along the velocity, matching the CPU tracer; before, backend "webgpu" threw and named "cpu" (#1785).+- Native sequence jobs keep one Cook session through moving geometry such as particles: new triangles rebuild only the changed meshes, and each frame's route reports `changedMeshes` (#1786)+- `backend: "native"` path traces a grid volume with Cook, and the Cook pin has grid volumes (#1789)+- The CPU tracer no longer traces a hidden part as a wall in a scene with local volumes, so a soft volume next to an idle hidden part is not cut by a hard edge any more (#1793)+- The CPU path tracer never traces a part marked `visible: false` any more, so a hidden box casts no shadow and is not drawn, as in Glaze, while a mesh volume that names a hidden part still takes its shape from it (#1806)+- `scripts/render-file-native.mjs` serves a GLB above 32 MiB over a local HTTP server instead of a Playwright route, so a single-mesh GLB of 4.8 million triangles (about 82 MiB) no longer closes the loader tab and reaches the native trace's triangle-count error (#1811).+- A light list over its cap now throws an error that names the cap, says it comes from the 32 bit light-linking mask (1 sun, 16 point, 8 spot, 7 area) and points to merging lights, an HDRI or emissive parts, or `backend: native` (#1812)+- Materials take `retro: { vertexSnap: [width, height], affineUV: true }` for the PS1 look in the raster renderer: vertices snap to a screen grid and UVs interpolate without perspective correction. Materials without it render as before.+- Blender export reports a light's max bounces below the scene's as not reproduced (Cook has one global bounce limit, no per-light field) (#1825)+- The path tracer lights a scene with `environment.keyIntensity` (a scene with no sun, sky or lamp traces black) and applies `effects.exposure` to the traced picture, as the rasterizer does (#1833)+- A post pass or material fragment is measured at 4096 bytes of code: comments and extra white space no longer count, and the error gives both sizes (#1834).+- A post pass reads its `time` from the motion time when the frame has one (a still at `motion.start: 6.5` gets 6.5, as the video frame does), instead of the renderer time, which stayed 0 in a still (#1835)+- A recipe places a catalog GLB by file name, `asset: { "name": "sheep_body.glb" }`, in `scene_create`, `scene_apply` and (as `assetName`) `scene_create_asset`, so a re-exported file no longer breaks it; the scene keeps the resolved SHA-256 (#1843)+- Native "cycles" exports keep a part with a base or coat normal map and a non-uniform matrix as an instance, so Cook tilts the normal map from the matrix like Cycles instead of from baked unit normals (#1844).+- `scene_apply`, `scene_render_submit` and `scene_review` count the scene's grid (.67vg) files toward the 64 MiB combined cap, and a scene already over it fails with the cap message instead of crashing the page (#1848)+- A folder of more than 256 consecutively numbered images is now one `sequences` entry in `scene_asset_catalog` (pattern, first, last, frames, size) instead of one entry per frame, and a material `map.sequence` can name it with `catalog` instead of listing the frame IDs; the frames are hashed when named, not at startup (#1852).+- A material `map.sequence` may list up to 4096 frames of 2048 pixels: a sequence over 64 frames streams, decoding the frame of each time and keeping two, so a canvas that changes over a whole film can be textured (#1852)+- A `stills` render job takes `stillCount` (2 to 300) and writes that many numbered PNGs, one per 1/fps s of the motion clock, so a transparent overlay comes from one job; `scene_capabilities` jobs also states the largest square frame (#1853).+- A local volume with `noiseOctaves` above 1 takes its march steps from its finest noise cell (up to 96), so its edge no longer comes out wider than the real density field (#1855)+- A native render job and `npm run render:file-cook` read the scene tables from the page as one Blob over the DevTools stream, which cuts the per-frame export of a dense scene with 1,000 particles from about 125 ms to about 64 ms (#1861)+- `scene_capabilities` takes `topic` as a list of up to 8 names and answers each one in one call, keyed by name (#1872)+- A native render that falls back because the installed Cook addon is stale now says so, with Cook's own fix (`npx cook fetch-addon`), instead of blaming a missing Vulkan ray query adapter (#1892)+- Shape and grid volumes weight each march step's glow by the exact integral over the step, `(1 - exp(-tau)) / tau` per channel, in the rasterizer, the CPU path tracer and the WebGPU path tracer, so a dense glowing volume is dimmer and its edge narrower: the 1-octave sphere edge goes from 0.2926 to 0.2295 sphere radii against Cycles 0.2330 (#1894)+- The shipped scene skill now matches the tools: the scene contract says up to seven area lights, `light.md` says an area light casts a shadow with `castShadow: true`, and `look.md` gives 8 MiB as the `.hdr` file limit (#1897)+- `createGaussianSplats` and `createGaussianSplatsAsync` now refuse a fractional `count`, a missing or short `positions`, `scales`, `rotations` or `colors` array, and `matrix: null` or `options: null`, each with an error that names the field, the expected length and what to pass (#1899)+- `simplifyGeometry` throws a `RangeError` that names the bad index, its place in `indices` and the vertex count (or the positions length when it is not a multiple of three) instead of crashing in the collapser, and treats `options: null` as the defaults (#1900)+- A `scene_apply` or `scene_create` call with several distinct `add-explosion` operations is now refused at once when its grids pass the 64 MiB combined limit: the check counts each blast at its exact size before baking, instead of after minutes of synchronous baking (#1903).+- `scene_capabilities` no longer answers names inherited from Object.prototype (`environment.constructor`, `parts.toString`) as fields or entries; a dotted suffix that is not an entry of a plain topic now fails with the list of entries (#1904)+- `EXT_meshopt_compression` views are checked before decoding: a bad `byteStride`, `byteOffset`, `byteLength`, `count`, `mode` or `filter` now throws an error naming the buffer view, the field, the value and what to pass, instead of `Malformed buffer data: -1`.+- A Draco GLB whose point or face count disagrees with its accessor `count`, or whose decoded buffer would pass the asset byte limit, is refused with an error naming the primitive, the attribute and both numbers before the decoder heap allocates it (#1913)+- `readGlb` in the catalog scripts now refuses a GLB with a wrong header version or total length, an unaligned or past-the-end chunk length, a zero or second BIN chunk, and a bad JSON chunk, each with an error that names the chunk and the lengths (#1919).+- A glTF primitive compressed with KHR_draco_mesh_compression now applies a sparse block on its accessors on top of the decoded data instead of ignoring it (#1921).+- `scene_render_status` of a finished job now returns its record while a newer job runs, instead of failing with "The renderer is leased".+- A render job with `animation` on a scene that has no clips now fails with an error that says to drop `animation` or add a clip, and a missing clips list counts as empty instead of a TypeError (#1928)+- `scene_capabilities` with a topic list answers the valid topics and reports unknown names (at most 4) with suggestions in `unknown`, and an unknown op name suggests the dotted `meshEdit.<op>` form.+- The shipped skills now cover review links with `reviewFile`, placing a GLB by file name, catalog image sequences, topic lists with `unknown`, `stillCount` stills jobs and the no-clips `animation` error (#1938)+- The agent bench passes the skill prompt to `claude` as a file, so `--skill-mode references` runs on Windows instead of failing with `spawn ENAMETOOLONG` (#1939)+- `detachedPartFlight` now settles a part with a `size` on its most stable face (the largest it can reach by falling or tipping with its landing energy) instead of the nearest face, so a bumper no longer ends standing on its narrow edge (#1971).+- A finished path traced job reports `noise` (relative noise before denoising and a one-line verdict) in `scene_render_result` (#905).+- `scene_look` and render jobs name the strongest lights, with the field to lower, in an `exposure-clipped` finding, so an over-exposed light can be fixed without opening the image (#905).+- `scene_ask_user` can show its annotate view inline in chat apps that support MCP Apps, with the same marks as the browser flow (#923)+- A scene can be exported as a phone review page (`scene_export_web` with `review: true`) whose saved marks `scene_ask_user` reads back with `reviewFile` (#924)+- `hingedDoorSwing` in `crucible/addons/vehicle` swings a damaged door open on its hinge: a torque-driven, damped swing that stops at `maxAngle` with a small bounce and rests (#983)+- `detachedPartFlight` in `crucible/addons/vehicle` gives a loose part torn off by a crash a seeded flight: launch from the impulse above its threshold, tumble, bounce and rest on the ground (#983)+- `crucible/addons/vehicle` dents vehicle parts on impact, finds the moment a drive meets a wall and lists the parts that tear off, and recipes get a `dented` geometry (#983)+## 0.20.0- 67js is now Crucible: the package is `@codepawlhq/crucible` (addons `@codepawlhq/crucible-*`), the commands are `crucible-mcp` and `crucible-addon-check`, the renderer class is `CrucibleRenderer` and the skills are `build-crucible-scene` and `render-crucible-media`. `67js-mcp`, `67js-addon-check` and `Renderer67` keep working for one release and print a rename notice, and recipes, assets and packs saved with the old `67js.scene`, `67js.asset`, `67js.pack` and `67js-pack.json` names still load. (#1183)diff --git a/bench/guide-fields/released.json b/bench/guide-fields/released.jsonindex 4d39c9ff..66c1b3f9 100644--- a/bench/guide-fields/released.json+++ b/bench/guide-fields/released.json@@ -1,5 +1,5 @@{- "release": "0.20.0",+ "release": "0.21.0","game": ["camera","ground",@@ -49,6 +49,7 @@"scene_apply operations[].blend.radius","scene_apply operations[].castShadow","scene_apply operations[].cells",+ "scene_apply operations[].center","scene_apply operations[].clip","scene_apply operations[].cloth","scene_apply operations[].color",@@ -69,6 +70,7 @@"scene_apply operations[].fStop","scene_apply operations[].far","scene_apply operations[].focalLength",+ "scene_apply operations[].focusDistance","scene_apply operations[].fov","scene_apply operations[].friction","scene_apply operations[].game",@@ -106,6 +108,7 @@"scene_apply operations[].node.label","scene_apply operations[].node.material","scene_apply operations[].node.parent",+ "scene_apply operations[].node.rayVisibility","scene_apply operations[].node.receiveShadow","scene_apply operations[].node.relations","scene_apply operations[].node.rotation",@@ -125,6 +128,7 @@"scene_apply operations[].physics","scene_apply operations[].postPasses","scene_apply operations[].projection",+ "scene_apply operations[].rayVisibility","scene_apply operations[].receiveShadow","scene_apply operations[].recursive","scene_apply operations[].relations",@@ -143,6 +147,7 @@"scene_apply operations[].shadowCatcher","scene_apply operations[].shadowOnly","scene_apply operations[].shift",+ "scene_apply operations[].size","scene_apply operations[].skin","scene_apply operations[].skin.armature","scene_apply operations[].start",@@ -160,6 +165,7 @@"scene_apply sessionId","scene_ask_user maxMarks","scene_ask_user prompt",+ "scene_ask_user reviewFile","scene_ask_user sessionId","scene_ask_user timeoutSeconds","scene_asset_catalog material",@@ -177,6 +183,7 @@"scene_create recipe.camera.fStop","scene_create recipe.camera.far","scene_create recipe.camera.focalLength",+ "scene_create recipe.camera.focusDistance","scene_create recipe.camera.fov","scene_create recipe.camera.near","scene_create recipe.camera.orthoScale",@@ -225,6 +232,7 @@"scene_create recipe.nodes[].label","scene_create recipe.nodes[].material","scene_create recipe.nodes[].parent",+ "scene_create recipe.nodes[].rayVisibility","scene_create recipe.nodes[].receiveShadow","scene_create recipe.nodes[].relations","scene_create recipe.nodes[].rotation",@@ -253,11 +261,13 @@"scene_create_asset recipe.animation.clip","scene_create_asset recipe.animation.time","scene_create_asset recipe.assetId",+ "scene_create_asset recipe.assetName","scene_create_asset recipe.camera","scene_create_asset recipe.camera.eye","scene_create_asset recipe.camera.fStop","scene_create_asset recipe.camera.far","scene_create_asset recipe.camera.focalLength",+ "scene_create_asset recipe.camera.focusDistance","scene_create_asset recipe.camera.fov","scene_create_asset recipe.camera.near","scene_create_asset recipe.camera.orthoScale",@@ -305,6 +315,7 @@"scene_export_glb motionTime","scene_export_glb sessionId","scene_export_recipe sessionId",+ "scene_export_web review","scene_export_web sessionId","scene_export_web title","scene_inspect ids",@@ -363,6 +374,7 @@"scene_render_submit settings.rendererTime","scene_render_submit settings.rendererTime.rate","scene_render_submit settings.rendererTime.start",+ "scene_render_submit settings.stillCount","scene_render_submit settings.timeline","scene_render_submit settings.timeline.start","scene_render_submit settings.transparent",@@ -387,6 +399,7 @@"scene_review shots[].camera.fStop","scene_review shots[].camera.far","scene_review shots[].camera.focalLength",+ "scene_review shots[].camera.focusDistance","scene_review shots[].camera.fov","scene_review shots[].camera.near","scene_review shots[].camera.orthoScale",diff --git a/bench/size/baseline.json b/bench/size/baseline.jsonindex eea5913a..391dfc69 100644--- a/bench/size/baseline.json+++ b/bench/size/baseline.json@@ -14,29 +14,29 @@},"bundles": {"entry:core": {- "min": 303434,- "gzip": 91934,- "brotli": 79071+ "min": 315798,+ "gzip": 95783,+ "brotli": 82259},"entry:core+lazy": {- "min": 306135,- "gzip": 93179,- "brotli": 80158+ "min": 318746,+ "gzip": 97142,+ "brotli": 83449},"entry:model": {- "min": 40093,- "gzip": 14018,- "brotli": 11556+ "min": 40135,+ "gzip": 14065,+ "brotli": 11578},"entry:nature": {- "min": 55133,- "gzip": 19229,- "brotli": 16367+ "min": 55149,+ "gzip": 19271,+ "brotli": 16319},"entry:motion": {- "min": 44022,- "gzip": 15350,- "brotli": 13040+ "min": 44129,+ "gzip": 15437,+ "brotli": 13029},"entry:canvas2d": {"min": 27190,@@ -44,19 +44,19 @@"brotli": 7331},"entry:gltf": {- "min": 168099,- "gzip": 55095,- "brotli": 47960+ "min": 177468,+ "gzip": 58203,+ "brotli": 50677},"entry:gltf+lazy": {- "min": 227992,- "gzip": 74665,- "brotli": 65517+ "min": 242075,+ "gzip": 79286,+ "brotli": 69559},"entry:modeling": {- "min": 43613,- "gzip": 16309,- "brotli": 14614+ "min": 46667,+ "gzip": 17536,+ "brotli": 15661},"entry:sculpt": {"min": 5937,@@ -66,167 +66,167 @@"entry:surface": {"min": 29893,"gzip": 9838,- "brotli": 7937+ "brotli": 7989},"entry:mesh-query": {- "min": 45588,- "gzip": 15651,- "brotli": 13072+ "min": 46056,+ "gzip": 15804,+ "brotli": 13264},"entry:scene-query": {- "min": 50989,- "gzip": 17629,- "brotli": 14875+ "min": 51198,+ "gzip": 17704,+ "brotli": 14979},"entry:mesh": {- "min": 305172,- "gzip": 102418,- "brotli": 85771+ "min": 322712,+ "gzip": 108776,+ "brotli": 90664},"entry:shrinkwrap": {- "min": 56422,- "gzip": 17941,- "brotli": 15130+ "min": 56486,+ "gzip": 17956,+ "brotli": 15080},"entry:surface-attachment": {"min": 42282,- "gzip": 13430,- "brotli": 11160+ "gzip": 13428,+ "brotli": 11153},"entry:physics": {"min": 62229,"gzip": 20428,- "brotli": 17502+ "brotli": 17530},"entry:physics+lazy": {"min": 71325,"gzip": 24467,- "brotli": 21147+ "brotli": 21179},"entry:settle": {"min": 70851,"gzip": 23442,- "brotli": 20173+ "brotli": 20195},"entry:settle+lazy": {"min": 79947,"gzip": 27481,- "brotli": 23818+ "brotli": 23844},"entry:spume": {- "min": 36194,- "gzip": 12855,- "brotli": 10829+ "min": 38995,+ "gzip": 13774,+ "brotli": 11638},"entry:retarget": {- "min": 56211,- "gzip": 19308,- "brotli": 16498+ "min": 56722,+ "gzip": 19473,+ "brotli": 16623},"entry:crowd-skinning": {- "min": 69820,- "gzip": 21386,- "brotli": 18403+ "min": 71894,+ "gzip": 21988,+ "brotli": 18819},"entry:skeleton-pose": {- "min": 75822,- "gzip": 25209,- "brotli": 21846+ "min": 76258,+ "gzip": 25379,+ "brotli": 21929},"entry:vehicle": {- "min": 47322,- "gzip": 16359,- "brotli": 13938+ "min": 60314,+ "gzip": 21066,+ "brotli": 18155},"entry:vegetation": {- "min": 85511,- "gzip": 30386,- "brotli": 26015+ "min": 86569,+ "gzip": 30805,+ "brotli": 26393},"entry:terrain": {- "min": 29683,- "gzip": 10932,- "brotli": 9829+ "min": 30429,+ "gzip": 11175,+ "brotli": 9999},"entry:heightfield": {"min": 2951,"gzip": 1387,- "brotli": 1232+ "brotli": 1237},"entry:export-mesh": {- "min": 5610,- "gzip": 2490,- "brotli": 2240+ "min": 5833,+ "gzip": 2587,+ "brotli": 2332},"entry:navmesh": {- "min": 7470,- "gzip": 3373,- "brotli": 3041+ "min": 7889,+ "gzip": 3509,+ "brotli": 3163},"entry:grass": {- "min": 28185,- "gzip": 10279,- "brotli": 9275+ "min": 28474,+ "gzip": 10398,+ "brotli": 9328},"entry:clutter": {- "min": 30801,- "gzip": 11229,- "brotli": 10086+ "min": 31091,+ "gzip": 11342,+ "brotli": 10204},"entry:light-bake": {- "min": 81534,- "gzip": 29797,- "brotli": 26255+ "min": 82422,+ "gzip": 30192,+ "brotli": 26588},"entry:light-bake-gpu": {- "min": 118432,- "gzip": 42412,- "brotli": 37090+ "min": 132946,+ "gzip": 47835,+ "brotli": 42583},"entry:light-bake-gpu+lazy": {- "min": 283066,- "gzip": 89435,- "brotli": 78164+ "min": 327261,+ "gzip": 103174,+ "brotli": 90821},"entry:ground-bounce": {- "min": 37273,- "gzip": 13196,- "brotli": 11956+ "min": 40571,+ "gzip": 14255,+ "brotli": 12904},"entry:floor-sky-visibility": {- "min": 48883,- "gzip": 18271,- "brotli": 16357+ "min": 53276,+ "gzip": 19658,+ "brotli": 17599},"entry:path-tracer": {- "min": 93688,- "gzip": 35231,- "brotli": 31097+ "min": 113571,+ "gzip": 42650,+ "brotli": 37739},"entry:path-tracer+lazy": {- "min": 254338,- "gzip": 80581,- "brotli": 70769+ "min": 303904,+ "gzip": 96335,+ "brotli": 84594},"entry:path-tracer-progressive": {- "min": 130709,- "gzip": 42558,- "brotli": 37590+ "min": 148676,+ "gzip": 48801,+ "brotli": 43076},"entry:path-tracer-progressive+lazy": {- "min": 219334,- "gzip": 67040,- "brotli": 58893+ "min": 264041,+ "gzip": 80614,+ "brotli": 70721},"entry:path-tracer-realtime": {- "min": 145486,- "gzip": 46988,- "brotli": 41813+ "min": 164383,+ "gzip": 53453,+ "brotli": 47258},"entry:path-tracer-realtime+lazy": {- "min": 263752,- "gzip": 80992,- "brotli": 71635+ "min": 309978,+ "gzip": 95102,+ "brotli": 83679},"entry:texture-set": {"min": 1837,@@ -234,89 +234,89 @@"brotli": 762},"entry:agent-scene": {- "min": 244810,- "gzip": 87216,- "brotli": 76828+ "min": 266385,+ "gzip": 95084,+ "brotli": 83874},"entry:agent-scene+lazy": {- "min": 1443012,- "gzip": 525143,- "brotli": 467238+ "min": 1588963,+ "gzip": 579604,+ "brotli": 515309},"entry:agent-asset": {- "min": 401803,- "gzip": 141219,- "brotli": 124347+ "min": 425316,+ "gzip": 149805,+ "brotli": 132056},"entry:agent-asset+lazy": {- "min": 1479868,- "gzip": 537348,- "brotli": 477152+ "min": 1625229,+ "gzip": 591668,+ "brotli": 525001},"entry:hdri": {- "min": 19740,- "gzip": 6797,- "brotli": 6092+ "min": 22615,+ "gzip": 7745,+ "brotli": 6915},"entry:camera-effects": {- "min": 25669,- "gzip": 9029,- "brotli": 7334+ "min": 25966,+ "gzip": 9125,+ "brotli": 7412},"entry:water-reflection": {- "min": 38768,- "gzip": 13566,- "brotli": 12279+ "min": 42173,+ "gzip": 14674,+ "brotli": 13261},"entry:render-passes": {- "min": 30598,- "gzip": 10852,- "brotli": 9001+ "min": 30779,+ "gzip": 10932,+ "brotli": 9048},"entry:user-pass": {- "min": 22388,- "gzip": 9088,- "brotli": 8113+ "min": 22714,+ "gzip": 9205,+ "brotli": 8240},"entry:line-art": {- "min": 55616,- "gzip": 16212,- "brotli": 14467+ "min": 61040,+ "gzip": 17833,+ "brotli": 15939},"entry:mesh-blend": {"min": 12261,- "gzip": 4553,- "brotli": 4137+ "gzip": 4554,+ "brotli": 4140},"entry:coplanar-overlap": {- "min": 28353,- "gzip": 11642,- "brotli": 10384+ "min": 29432,+ "gzip": 12059,+ "brotli": 10769},"entry:reversed-depth": {"min": 23479,"gzip": 8245,- "brotli": 6609+ "brotli": 6605},"entry:material-layer-shaders": {- "min": 11822,- "gzip": 3473,- "brotli": 3187+ "min": 12594,+ "gzip": 3679,+ "brotli": 3352},"entry:view-transform-shader": {- "min": 5251,- "gzip": 2463,- "brotli": 2187+ "min": 5266,+ "gzip": 2469,+ "brotli": 2182},"entry:view-transform": {- "min": 3715,- "gzip": 1973,- "brotli": 1739+ "min": 4054,+ "gzip": 2114,+ "brotli": 1861},"entry:liquid-shader": {- "min": 11751,- "gzip": 4217,- "brotli": 3831+ "min": 14111,+ "gzip": 4967,+ "brotli": 4504},"entry:world-material-shader": {"min": 19406,@@ -326,32 +326,42 @@"entry:splat-shader": {"min": 22719,"gzip": 6820,- "brotli": 6152+ "brotli": 6145},"entry:subsurface-shader": {"min": 3871,"gzip": 1617,- "brotli": 1467+ "brotli": 1470+ },+ "entry:retro-shader": {+ "min": 2311,+ "gzip": 1147,+ "brotli": 1017},"entry:fog-shader": {- "min": 55506,- "gzip": 15991,- "brotli": 14408+ "min": 86449,+ "gzip": 25490,+ "brotli": 22609+ },+ "entry:linear-output-shader": {+ "min": 701,+ "gzip": 433,+ "brotli": 358},"entry:light-linking": {- "min": 2539,- "gzip": 1133,- "brotli": 1019+ "min": 2616,+ "gzip": 1156,+ "brotli": 1042},"entry:cascaded-shadows": {- "min": 33377,- "gzip": 11024,- "brotli": 9121+ "min": 33399,+ "gzip": 11035,+ "brotli": 9120},"entry:light-shapes": {- "min": 7782,- "gzip": 3270,- "brotli": 2969+ "min": 7950,+ "gzip": 3331,+ "brotli": 3019},"entry:compact-color": {"min": 1430,@@ -359,69 +369,69 @@"brotli": 563},"entry:gaussian-splats": {- "min": 31374,- "gzip": 11306,- "brotli": 9423+ "min": 46888,+ "gzip": 16285,+ "brotli": 13950},"entry:replace-textured-parts": {- "min": 54072,- "gzip": 15368,- "brotli": 13627+ "min": 56267,+ "gzip": 16022,+ "brotli": 14176},"entry:replace-textured-parts+lazy": {- "min": 56773,- "gzip": 16612,- "brotli": 14714+ "min": 59215,+ "gzip": 17381,+ "brotli": 15359},"entry:texture-update": {- "min": 55852,- "gzip": 16048,- "brotli": 14193+ "min": 58047,+ "gzip": 16704,+ "brotli": 14750},"entry:texture-update+lazy": {- "min": 58553,- "gzip": 17293,- "brotli": 15279+ "min": 60995,+ "gzip": 18063,+ "brotli": 15933},"entry:cloth": {- "min": 20666,- "gzip": 8547,- "brotli": 7530+ "min": 21497,+ "gzip": 8780,+ "brotli": 7767},"entry:runtime": {- "min": 6799,- "gzip": 2660,- "brotli": 2387+ "min": 6824,+ "gzip": 2673,+ "brotli": 2392},"entry:streaming": {- "min": 11373,- "gzip": 4362,- "brotli": 3933+ "min": 11507,+ "gzip": 4399,+ "brotli": 3959},"entry:streaming+lazy": {- "min": 239259,- "gzip": 78884,- "brotli": 69390+ "min": 253476,+ "gzip": 83547,+ "brotli": 73446},"entry:netcode": {- "min": 23825,- "gzip": 7446,- "brotli": 6770+ "min": 25052,+ "gzip": 7861,+ "brotli": 7139},"entry:input": {- "min": 15165,- "gzip": 5289,- "brotli": 4768+ "min": 15281,+ "gzip": 5348,+ "brotli": 4834},"entry:physics-runtime": {- "min": 65915,- "gzip": 21521,- "brotli": 18548+ "min": 66161,+ "gzip": 21573,+ "brotli": 18567},"entry:physics-runtime+lazy": {- "min": 75011,- "gzip": 25560,- "brotli": 22193+ "min": 75257,+ "gzip": 25612,+ "brotli": 22216},"entry:physics-jolt": {"min": 10790,@@ -429,69 +439,89 @@"brotli": 4172},"entry:audio": {- "min": 14833,- "gzip": 5612,- "brotli": 5098+ "min": 15147,+ "gzip": 5691,+ "brotli": 5174},"entry:animation-graph": {- "min": 38234,- "gzip": 13546,- "brotli": 12274+ "min": 38525,+ "gzip": 13717,+ "brotli": 12488},"entry:player": {- "min": 22042,- "gzip": 7154,- "brotli": 6425+ "min": 22059,+ "gzip": 7157,+ "brotli": 6429},"entry:game-rules": {- "min": 12108,- "gzip": 4553,- "brotli": 4100+ "min": 12311,+ "gzip": 4641,+ "brotli": 4189},"entry:game": {- "min": 594781,- "gzip": 203053,- "brotli": 178569+ "min": 630675,+ "gzip": 216558,+ "brotli": 190600},"entry:game+lazy": {- "min": 1771292,- "gzip": 634958,- "brotli": 563983+ "min": 1928017,+ "gzip": 693916,+ "brotli": 615988},"entry:update-part-vertices": {- "min": 1254,- "gzip": 693,- "brotli": 609+ "min": 1309,+ "gzip": 720,+ "brotli": 638+ },+ "entry:reflection-probes": {+ "min": 58164,+ "gzip": 20972,+ "brotli": 18655+ },+ "entry:fracture": {+ "min": 6444,+ "gzip": 2914,+ "brotli": 2630+ },+ "entry:morph": {+ "min": 28623,+ "gzip": 9633,+ "brotli": 7887+ },+ "entry:skinning": {+ "min": 23742,+ "gzip": 8356,+ "brotli": 6705},"scene:lit-box": {- "min": 206994,- "gzip": 60017,- "brotli": 51126+ "min": 214146,+ "gzip": 62149,+ "brotli": 52915},"scene:lit-box+lazy": {- "min": 209695,- "gzip": 61262,- "brotli": 52213+ "min": 217094,+ "gzip": 63506,+ "brotli": 54097},"scene:gltf-box": {- "min": 299609,- "gzip": 92245,- "brotli": 79886+ "min": 313911,+ "gzip": 96806,+ "brotli": 83647},"scene:gltf-box+lazy": {- "min": 359502,- "gzip": 111812,- "brotli": 97439+ "min": 378523,+ "gzip": 117892,+ "brotli": 102532},"scene:agent-scene": {- "min": 402505,- "gzip": 132802,- "brotli": 116588+ "min": 430990,+ "gzip": 143161,+ "brotli": 125891},"scene:agent-scene+lazy": {- "min": 1562207,- "gzip": 560275,- "brotli": 497555+ "min": 1706408,+ "gzip": 614882,+ "brotli": 545946}}}diff --git a/bench/size/budgets.json b/bench/size/budgets.jsonindex 8845aebd..96a5b968 100644--- a/bench/size/budgets.json+++ b/bench/size/budgets.json@@ -1,106 +1,112 @@{- "archiveBytes": 1764128,+ "archiveBytes": 1951497,"goals": {"lit-box": 50000,"gltf-box": 80000,"agent-scene": 100000},"budgets": {- "entry:core": 93800,- "entry:core+lazy": 95100,- "entry:model": 14300,+ "entry:core": 97700,+ "entry:core+lazy": 99100,+ "entry:model": 14400,"entry:nature": 19700,- "entry:motion": 15700,+ "entry:motion": 15800,"entry:canvas2d": 9400,- "entry:gltf": 56200,- "entry:gltf+lazy": 76200,- "entry:modeling": 16700,+ "entry:gltf": 59400,+ "entry:gltf+lazy": 80900,+ "entry:modeling": 17900,"entry:sculpt": 2500,"entry:surface": 10100,- "entry:mesh-query": 16000,- "entry:scene-query": 18000,- "entry:mesh": 104500,- "entry:shrinkwrap": 18300,+ "entry:mesh-query": 16200,+ "entry:scene-query": 18100,+ "entry:mesh": 111000,+ "entry:shrinkwrap": 18400,"entry:surface-attachment": 13700,"entry:physics": 20900,"entry:physics+lazy": 25000,"entry:settle": 24000,"entry:settle+lazy": 28100,- "entry:spume": 13200,- "entry:retarget": 19700,- "entry:crowd-skinning": 21900,- "entry:skeleton-pose": 25800,- "entry:vehicle": 16700,- "entry:vegetation": 31000,- "entry:terrain": 11200,+ "entry:spume": 14100,+ "entry:retarget": 19900,+ "entry:crowd-skinning": 22500,+ "entry:skeleton-pose": 25900,+ "entry:vehicle": 21500,+ "entry:vegetation": 31500,+ "entry:terrain": 11400,"entry:heightfield": 1500,- "entry:export-mesh": 2600,- "entry:navmesh": 3500,- "entry:grass": 10500,- "entry:clutter": 11500,- "entry:light-bake": 30400,- "entry:light-bake-gpu": 43300,- "entry:light-bake-gpu+lazy": 91300,- "entry:ground-bounce": 13500,- "entry:floor-sky-visibility": 18700,- "entry:path-tracer": 36000,- "entry:path-tracer+lazy": 82200,- "entry:path-tracer-progressive": 43500,- "entry:path-tracer-progressive+lazy": 68400,- "entry:path-tracer-realtime": 48000,- "entry:path-tracer-realtime+lazy": 82700,+ "entry:export-mesh": 2700,+ "entry:navmesh": 3600,+ "entry:grass": 10700,+ "entry:clutter": 11600,+ "entry:light-bake": 30800,+ "entry:light-bake-gpu": 48800,+ "entry:light-bake-gpu+lazy": 105300,+ "entry:ground-bounce": 14600,+ "entry:floor-sky-visibility": 20100,+ "entry:path-tracer": 43600,+ "entry:path-tracer+lazy": 98300,+ "entry:path-tracer-progressive": 49800,+ "entry:path-tracer-progressive+lazy": 82300,+ "entry:path-tracer-realtime": 54600,+ "entry:path-tracer-realtime+lazy": 97100,"entry:texture-set": 900,- "entry:agent-scene": 89000,- "entry:agent-scene+lazy": 535700,- "entry:agent-asset": 144100,- "entry:agent-asset+lazy": 548100,- "entry:hdri": 7000,- "entry:camera-effects": 9300,- "entry:water-reflection": 13900,- "entry:render-passes": 11100,- "entry:user-pass": 9300,- "entry:line-art": 16600,+ "entry:agent-scene": 97000,+ "entry:agent-scene+lazy": 591200,+ "entry:agent-asset": 152900,+ "entry:agent-asset+lazy": 603600,+ "entry:hdri": 7900,+ "entry:camera-effects": 9400,+ "entry:water-reflection": 15000,+ "entry:render-passes": 11200,+ "entry:user-pass": 9400,+ "entry:line-art": 18200,"entry:mesh-blend": 4700,- "entry:coplanar-overlap": 11900,+ "entry:coplanar-overlap": 12400,"entry:reversed-depth": 8500,- "entry:material-layer-shaders": 3600,+ "entry:material-layer-shaders": 3800,"entry:view-transform-shader": 2600,- "entry:view-transform": 2100,- "entry:liquid-shader": 4400,+ "entry:view-transform": 2200,+ "entry:liquid-shader": 5100,"entry:world-material-shader": 6700,"entry:splat-shader": 7000,"entry:subsurface-shader": 1700,- "entry:fog-shader": 16400,+ "entry:retro-shader": 1200,+ "entry:fog-shader": 26000,+ "entry:linear-output-shader": 500,"entry:light-linking": 1200,"entry:cascaded-shadows": 11300,"entry:light-shapes": 3400,"entry:compact-color": 700,- "entry:gaussian-splats": 11600,- "entry:replace-textured-parts": 15700,- "entry:replace-textured-parts+lazy": 17000,- "entry:texture-update": 16400,- "entry:texture-update+lazy": 17700,- "entry:cloth": 8800,+ "entry:gaussian-splats": 16700,+ "entry:replace-textured-parts": 16400,+ "entry:replace-textured-parts+lazy": 17800,+ "entry:texture-update": 17100,+ "entry:texture-update+lazy": 18500,+ "entry:cloth": 9000,"entry:runtime": 2800,"entry:streaming": 4500,- "entry:streaming+lazy": 80500,- "entry:netcode": 7600,- "entry:input": 5400,- "entry:physics-runtime": 22000,- "entry:physics-runtime+lazy": 26100,+ "entry:streaming+lazy": 85300,+ "entry:netcode": 8100,+ "entry:input": 5500,+ "entry:physics-runtime": 22100,+ "entry:physics-runtime+lazy": 26200,"entry:physics-jolt": 4700,- "entry:audio": 5800,- "entry:animation-graph": 13900,- "entry:player": 7300,- "entry:game-rules": 4700,- "entry:game": 207200,- "entry:game+lazy": 647700,+ "entry:audio": 5900,+ "entry:animation-graph": 14000,+ "entry:player": 7400,+ "entry:game-rules": 4800,+ "entry:game": 220900,+ "entry:game+lazy": 707800,"entry:update-part-vertices": 800,- "scene:lit-box": 61300,- "scene:lit-box+lazy": 62500,- "scene:gltf-box": 94100,- "scene:gltf-box+lazy": 114100,- "scene:agent-scene": 135500,- "scene:agent-scene+lazy": 571500+ "entry:reflection-probes": 21400,+ "entry:fracture": 3000,+ "entry:morph": 9900,+ "entry:skinning": 8600,+ "scene:lit-box": 63400,+ "scene:lit-box+lazy": 64800,+ "scene:gltf-box": 98800,+ "scene:gltf-box+lazy": 120300,+ "scene:agent-scene": 146100,+ "scene:agent-scene+lazy": 627200}}diff --git a/bench/tool-cost-budgets.json b/bench/tool-cost-budgets.jsonindex 054c6c4d..9a43f8ba 100644--- a/bench/tool-cost-budgets.json+++ b/bench/tool-cost-budgets.json@@ -1,18 +1,18 @@{- "schemaTotalBytes": 74990,+ "schemaTotalBytes": 76470,"schemaToolBytes": {- "scene_apply": 23360,- "scene_create": 13410,- "scene_create_asset": 8980,- "scene_render_submit": 4760+ "scene_apply": 23900,+ "scene_create": 13540,+ "scene_create_asset": 9130,+ "scene_render_submit": 4830},"schemaOtherToolBytes": 4000,"resultBytes": {- "scene_capabilities": 1630,+ "scene_capabilities": 1730,"scene_create": 1200,"scene_inspect": 1150,- "mesh_inspect": 1110,- "mesh_edit": 2220,+ "mesh_inspect": 1130,+ "mesh_edit": 2240,"scene_apply": 100,"scene_export_recipe": 2730}diff --git a/changes/1046-bloom.md b/changes/1046-bloom.mddeleted file mode 100644index 30041dc0..00000000--- a/changes/1046-bloom.md+++ /dev/null@@ -1 +0,0 @@-- A local volume that emits now feeds bloom in the rasterizer: the volumes are drawn before bloom extraction for a frame with bloom on (#1046).diff --git a/changes/1046-bloom.size.md b/changes/1046-bloom.size.mddeleted file mode 100644index b93917fa..00000000--- a/changes/1046-bloom.size.md+++ /dev/null@@ -1,12 +0,0 @@-- entry:fog-shader: the composite pass (its own fragment shader and target) that draws a glowing volume before bloom extraction; it compiles only for a frame with bloom and a glowing volume.-- scene:agent-scene: the composite path in the frame loop and its release hooks.-- scene:agent-scene+lazy: the composite code loads with the volume code.-- goal:agent-scene: the composite path in the frame loop.-- goal:lit-box: the guarded composite call in the frame path.-- goal:gltf-box: the same guarded call.-- entry:core: the hook in the renderer that draws glowing local volumes before bloom extraction, so a fireball blooms.-- entry:core+lazy: same hook.-- scene:lit-box+lazy: same hook in a scene that uses the renderer.-- scene:gltf-box+lazy: same hook.-- scene:lit-box: the guarded composite call and release hooks in the frame path, which do nothing unless a glowing volume and bloom are both on.-- scene:gltf-box: the same guarded call and release hooks.diff --git a/changes/1046-ramp.md b/changes/1046-ramp.mddeleted file mode 100644index b33d783b..00000000--- a/changes/1046-ramp.md+++ /dev/null@@ -1 +0,0 @@-- A local volume takes `emissionRamp`, a blackbody (`kelvin`) or a colour ramp (`stops`) read by density, so one fireball grades from a white core to an orange edge in the rasterizer and the CPU path tracer (#1046)diff --git a/changes/1046-ramp.size.md b/changes/1046-ramp.size.mddeleted file mode 100644index cee4a1bf..00000000--- a/changes/1046-ramp.size.md+++ /dev/null@@ -1,5 +0,0 @@-- entry:path-tracer: the CPU tracer bakes and reads the emission ramp (the blackbody curve against the CIE colour matching functions is about 1.2 KB gzip); it buys a fireball that grades from a white core to an orange edge in one volume-- entry:fog-shader: the rasterizer packs and reads the same ramp (the shader lookup and the shared blackbody bake); the other half of the same feature-- scene:agent-scene: 161 bytes gzip for the emissionRamp help text in scene_capabilities and the on-demand module entry; the ramp checks and the blackbody live in a lazy module, not the core-- scene:agent-scene+lazy: the lazy chunks carry the ramp checks, the blackbody bake and the shader lookup-- goal:agent-scene: 161 bytes gzip, all of it the scene_capabilities text an agent reads to find emissionRamp; the code sits in a lazy modulediff --git a/changes/1046.md b/changes/1046.mddeleted file mode 100644index 9ca04846..00000000--- a/changes/1046.md+++ /dev/null@@ -1 +0,0 @@-- A local volume with `lightsScene: true` adds a point light at its emission, so a fireball lights the floor and props around it in the rasterizer (#1046).diff --git a/changes/1046.size.md b/changes/1046.size.mddeleted file mode 100644index 09d1bb68..00000000--- a/changes/1046.size.md+++ /dev/null@@ -1,6 +0,0 @@-- entry:fog-shader: the fire light sums a grid's flame and blackbody light per frame and places a point light; without it an emitting fireball cannot light the floor or props around it, and the code loads only with local volumes.-- scene:agent-scene: the new lightsScene field and its check, plus the capabilities line, so an agent can switch fire light on.-- scene:agent-scene+lazy: the fire-light module loads with the volume code that grid scenes already carry (grid decode and blackbody).-- goal:lit-box: one guarded call in the frame path (about 50 bytes gzip), which does nothing unless a local volume sets lightsScene.-- goal:gltf-box: same guarded call in the frame path.-- goal:agent-scene: the lightsScene field and its check.diff --git a/changes/1047-octaves.md b/changes/1047-octaves.mddeleted file mode 100644index 6965ef97..00000000--- a/changes/1047-octaves.md+++ /dev/null@@ -1 +0,0 @@-- A scattering local volume takes `scatterOctaves` from 1 to 4, the octave sum of `volumeFog` for the key light, in the rasterizer and the CPU path tracer; 1 draws as before (#1047).diff --git a/changes/1047-octaves.size.md b/changes/1047-octaves.size.mddeleted file mode 100644index 41957889..00000000--- a/changes/1047-octaves.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: the scatterOctaves field and its check.-- scene:agent-scene+lazy: the field and the CPU tracer's octave sum, which ship with the local volume code.-- goal:agent-scene: the scatterOctaves field and its check, so an agent can ask for multiple scattering.diff --git a/changes/1047-surface-shadow.md b/changes/1047-surface-shadow.mddeleted file mode 100644index 88ee7925..00000000--- a/changes/1047-surface-shadow.md+++ /dev/null@@ -1 +0,0 @@-- A box or sphere local volume can shade the opaque surfaces under it from the key light, in the rasterizer and the CPU path tracer, when `castShadow: true` is set; it is off by default so the default frame does not change (#1047).diff --git a/changes/1047-surface-shadow.size.md b/changes/1047-surface-shadow.size.mddeleted file mode 100644index 31d2c0ea..00000000--- a/changes/1047-surface-shadow.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the castShadow field and its check, and the CPU tracer's surface-shadow function; both ship with the local volume code that carries them.-- goal:agent-scene: the castShadow field and its check, so an agent can switch the surface shadow on.diff --git a/changes/1047.md b/changes/1047.mddeleted file mode 100644index bb40984b..00000000--- a/changes/1047.md+++ /dev/null@@ -1 +0,0 @@-- A scattering box or sphere volume with `shadowDetail: true` shades itself by its own noisy density along the sun, in the rasterizer and the CPU path tracer (#1047).diff --git a/changes/1047.size.md b/changes/1047.size.mddeleted file mode 100644index b2857b90..00000000--- a/changes/1047.size.md+++ /dev/null@@ -1,6 +0,0 @@-- entry:fog-shader: the second packed component, one more argument on the light depth and the tap constant; the shadow detail has to run in the post shader that already holds the light depth.-- scene:agent-scene: the shadowDetail field and its check.-- scene:agent-scene+lazy: the field and the CPU path tracer's extra tap count.-- goal:lit-box: the field check shared by every local volume scene.-- goal:gltf-box: same shared check.-- goal:agent-scene: the shadowDetail field and its check.diff --git a/changes/1048.md b/changes/1048.mddeleted file mode 100644index 2654401f..00000000--- a/changes/1048.md+++ /dev/null@@ -1 +0,0 @@-- `backend: "webgpu"` draws box and sphere `environment.localVolumes` (smoke, glow, fire) at the same look as the CPU tracer, 23 times faster at 480 by 300 and 13 times at 1080p (#1048)diff --git a/changes/1048.size.md b/changes/1048.size.mddeleted file mode 100644index 1c457f88..00000000--- a/changes/1048.size.md+++ /dev/null@@ -1,4 +0,0 @@-- entry:light-bake-gpu+lazy: the WGSL march of box and sphere local volumes (about 3 KB gzip) sits in the shared kernel module, which every WebGPU entry bundles; it buys backend "webgpu" volume shots 13 to 23 times faster than the CPU tracer-- entry:path-tracer+lazy: same WGSL march, loaded only with the WebGPU tracer chunk-- entry:path-tracer-progressive+lazy: same WGSL march, loaded only with the WebGPU tracer chunk-- entry:path-tracer-realtime+lazy: the realtime entry carries about 3 KB of WGSL it never runs (it still refuses local volumes); accepted so the march stays in one shared kernel modulediff --git a/changes/1050.md b/changes/1050.mddeleted file mode 100644index b0b55420..00000000--- a/changes/1050.md+++ /dev/null@@ -1 +0,0 @@-- `scene_apply` `add-explosion` writes a whole blast into plain recipe sections from one call: a gas-sim fireball and smoke volume with a white-hot core that lights the floor, sparks, flash, shock ring, dust, rubble, scorch, camera shake and optional fractured debris; particle systems take a `delay`, and exported recipes and render jobs need no extra files (#1050)diff --git a/changes/1050.size.md b/changes/1050.size.mddeleted file mode 100644index db70cc19..00000000--- a/changes/1050.size.md+++ /dev/null@@ -1,6 +0,0 @@-- scene:agent-scene: the operation's validation entry, the module markers and a short capability text (366 bytes gzip); the core entry is +530 bytes gzip now that the capability text names the grid files; the 4 KB expansion itself is its own on-demand chunk-- goal:agent-scene: add-explosion needs its dispatch and markers in the core; the expansion code is already lazy, so the core cannot shrink further without dropping the operation-- tool-total: one operation schema of four fields (373 bytes); every other field lives in scene_capabilities topic explosion-- tool:scene_apply: same operation schema, the minimum an agent needs to ask for a blast in one call-- scene:agent-scene+lazy: src/agent-explosion.js loads only when a recipe uses add-explosion and replaces about six hand-built sections per blast; the chunk now carries the gas-sim solver (src/gas-sim.js), the grid encoder and src/explosion-grids.js, +9.4 KB gzip measured (agent-scene+lazy 556,186 to 565,538); it loads only when a recipe uses add-explosion. This is the cost of the approved gas-sim look, which a noise box cannot give: a baked simulation lit by physical blackbody emission, with a floor shadow-- tool:scene_apply: the scene_capabilities topic text for explosion changed (12 bytes more in the topic result, no schema change; the 373 bytes are the operation schema as before), so the Haiku bench is pending from the leaddiff --git a/changes/1081.md b/changes/1081.mddeleted file mode 100644index 9552ac9d..00000000--- a/changes/1081.md+++ /dev/null@@ -1 +0,0 @@-- A scene node takes `rayVisibility` in `scene_create` and `scene_apply`, with `'glass'` as a shorthand that lets shadow rays through a window pane so the path tracer samples sun and sky through it; emitters hidden from bounce rays stay refused, with the reason in `docs/path-tracer.md` (#1081)diff --git a/changes/1081.size.md b/changes/1081.size.mddeleted file mode 100644index bf4cb37f..00000000--- a/changes/1081.size.md+++ /dev/null@@ -1,4 +0,0 @@-- scene:agent-scene: about 420 bytes gzip for the one capability that lets an agent-built interior path trace sun and sky through a window pane (validation of `rayVisibility` with errors that say what to pass, the `'glass'` shorthand, part wiring). Before this the field existed only in the library.-- scene:agent-scene+lazy: the same bytes, nothing else is added to the lazy modules.-- goal:agent-scene: the field is an agent-facing capability with no smaller form; the schema side was cut from 870 to 57 bytes by leaving the value open in the tool listing and checking it in the scene, whose errors name what to pass.-- tool-total: three listings of the `rayVisibility` key, 57 bytes, within the allowance.diff --git a/changes/1083-showcase-cook-stills.md b/changes/1083-showcase-cook-stills.mddeleted file mode 100644index 57dbfe9b..00000000--- a/changes/1083-showcase-cook-stills.md+++ /dev/null@@ -1 +0,0 @@-- The showcase page shows three Cook path-traced stills with their credits (#1083)diff --git a/changes/1111.md b/changes/1111.mddeleted file mode 100644index 3be2a564..00000000--- a/changes/1111.md+++ /dev/null@@ -1 +0,0 @@-- Cook reads a tangent-space normal map as Cycles' Normal Map node does, so a normal-mapped mesh on a non-uniformly scaled object no longer renders brighter or darker than in Cycles, and `render-cycles.py` builds a Normal Map node for parity scenes (#1111)diff --git a/changes/1123.md b/changes/1123.mddeleted file mode 100644index c3a67c32..00000000--- a/changes/1123.md+++ /dev/null@@ -1 +0,0 @@-- Native film frames get their 8-bit picture from Cook's GPU post stage and a faster PNG encoder, taking a 1080p frame from traced film to PNG in 22 ms instead of 276 ms (#1123)diff --git a/changes/1126.md b/changes/1126.mddeleted file mode 100644index 6e9cb619..00000000--- a/changes/1126.md+++ /dev/null@@ -1,2 +0,0 @@-- Ambient occlusion no longer draws vertical bands on flat walls seen at a slant (#1126)- media: https://raw.githubusercontent.com/codepawl/crucible/evidence-pr-1239/after.png The corner scene with ambient occlusion 0.5: the walls are smooth, with darkening only at the crates, the sphere and the wall corner.diff --git a/changes/1126.size.md b/changes/1126.size.mddeleted file mode 100644index 57a158d3..00000000--- a/changes/1126.size.md+++ /dev/null@@ -1,4 +0,0 @@-- goal:lit-box: 52 bytes gzip fix vertical bands that ambient occlusion drew on flat walls (#1126), a visible glitch.-- goal:gltf-box: 53 bytes gzip for the same fix.-- goal:agent-scene: 56 bytes gzip for the same fix.-- scene:agent-scene+lazy: 87 bytes gzip for the same fix (561227 to 561314 against the base branch), a visible glitch.diff --git a/changes/1131.md b/changes/1131.mddeleted file mode 100644index 2c71db77..00000000--- a/changes/1131.md+++ /dev/null@@ -1 +0,0 @@-- Guides tag fields newer than the last npm release with {since X.Y.Z}, a test fails when one is left untagged, and an MCP refusal of an unknown field names the installed version (#1131)diff --git a/changes/1140.md b/changes/1140.mddeleted file mode 100644index 7662bbcb..00000000--- a/changes/1140.md+++ /dev/null@@ -1 +0,0 @@-- `parseGaussianSplatSpz` and `parseGaussianSplatSpzAsync` read gzip SPZ splat files (versions 2 and 3, as written by Niantic tools and World Labs Atlas) into the same data as the PLY parser, with the same limits (#1140)diff --git a/changes/1140.size.md b/changes/1140.size.mddeleted file mode 100644index 75ff5958..00000000--- a/changes/1140.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:gaussian-splats: the optional addon grows 1517 bytes gzip for the SPZ reader (versions 2 and 3), the format World Labs Atlas returns and about 10 times smaller than the PLY, so a hosted splat page downloads a tenth of the bytes; the core bundle is unchanged.diff --git a/changes/1142.md b/changes/1142.mddeleted file mode 100644index ef085f27..00000000--- a/changes/1142.md+++ /dev/null@@ -1 +0,0 @@-- `parseGaussianSplatPlyAsync` and `createGaussianSplatsAsync` load large splat files in slices with progress and abort, so the page keeps responding (#1142)diff --git a/changes/1142.size.md b/changes/1142.size.mddeleted file mode 100644index 46a4f344..00000000--- a/changes/1142.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:gaussian-splats: the optional addon grows 599 bytes gzip for async parse and pack with progress and abort, which cuts the longest main-thread block for 1,000,000 splats from about 275 ms to about 22 ms; the core bundle is unchanged.diff --git a/changes/1144-webgpu.md b/changes/1144-webgpu.mddeleted file mode 100644index f3801537..00000000--- a/changes/1144-webgpu.md+++ /dev/null@@ -1 +0,0 @@-- The WebGPU path tracer samples the sky through light portals (areaLights with portal: true) with the same density as the CPU tracer and Cook; only the WebGL tracer still ignores them (#1144).diff --git a/changes/1144.md b/changes/1144.mddeleted file mode 100644index b497df3b..00000000--- a/changes/1144.md+++ /dev/null@@ -1 +0,0 @@-- The CPU path tracer samples the sky through light portals (areaLights with portal: true), so a room lit through a window converges faster at the same mean; the WebGL and WebGPU tracers still ignore them (#1144).diff --git a/changes/1153.md b/changes/1153.mddeleted file mode 100644index cea91f0d..00000000--- a/changes/1153.md+++ /dev/null@@ -1 +0,0 @@-- The Blender add-on and the Cycles parity bridge now multiply a light's energy by its Emission node's constant Strength and its colour by a constant colour or a constant-temperature Blackbody, and report other light node setups as not reproduced with the light name (#1153).diff --git a/changes/1158-adapter-bound.md b/changes/1158-adapter-bound.mddeleted file mode 100644index c6ea2e86..00000000--- a/changes/1158-adapter-bound.md+++ /dev/null@@ -1 +0,0 @@-- A chunked native trace packs up to the unique triangles the adapter reports (`maxUniqueTriangles`, 4.79 million on a 12 GB RTX 3060) instead of a fixed 4,000,000, up to a ceiling of 16,000,000, and a scene over the bound is refused with the count, the limit and how many triangles too many (#1158)diff --git a/changes/1158-capacity.md b/changes/1158-capacity.mddeleted file mode 100644index 18432f31..00000000--- a/changes/1158-capacity.md+++ /dev/null@@ -1 +0,0 @@-- A native render job and `render:file-cook` check a scene against the adapter capacity before exporting it, fail with one error naming the count, the limit and what to drop, and report the capacity used (#1158)diff --git a/changes/1158-followup.md b/changes/1158-followup.mddeleted file mode 100644index 74940468..00000000--- a/changes/1158-followup.md+++ /dev/null@@ -1 +0,0 @@-- A native trace of zero parts is refused with a message that says to pass a part, and the docs name the tracers that still keep 1,024 parts (#1158)diff --git a/changes/1158.md b/changes/1158.mddeleted file mode 100644index 00897ec4..00000000--- a/changes/1158.md+++ /dev/null@@ -1 +0,0 @@-- A native trace takes `maximumParts` (up to 1,000,000) instead of 1,024 parts or distinct geometries, so scenes of thousands of separate meshes export without merging, and a scene past it is refused with the count and what to do (#1158)diff --git a/changes/1159-native.md b/changes/1159-native.mddeleted file mode 100644index 127181a4..00000000--- a/changes/1159-native.md+++ /dev/null@@ -1 +0,0 @@-- The native `cycles` tracer now starts camera rays at `camera.clipStart`, so a wall between the eye and the clip start is not drawn, and `report.ignored` no longer lists it (#1159).diff --git a/changes/1159.md b/changes/1159.mddeleted file mode 100644index 8cae5595..00000000--- a/changes/1159.md+++ /dev/null@@ -1 +0,0 @@-- Add `camera.clipStart` to the CPU and WebGPU path tracers, so camera rays start that far along the view axis and a wall between the eye and the clip start is not drawn; the WebGL tracer and native kernel name it in `report.ignored` (#1159).diff --git a/changes/1161.md b/changes/1161.mddeleted file mode 100644index 7454248f..00000000--- a/changes/1161.md+++ /dev/null@@ -1 +0,0 @@-- MCP render jobs, scene_create and scene_apply draw grid volumes (`.67vg` files in the grid folder, still or a sequence): the server hands the grid bytes to the preview page by name (#1161)diff --git a/changes/1163-dense-lods.md b/changes/1163-dense-lods.mddeleted file mode 100644index 982d83bc..00000000--- a/changes/1163-dense-lods.md+++ /dev/null@@ -1 +0,0 @@-- `generateGeometryLods` keeps uvs, colours and normals, so a textured part with levels draws, and `addLodsToDenseParts(parts, { pixels })` gives levels to the big parts of a scene with more than one triangle per pixel (#1163)diff --git a/changes/1163-dense-lods.size.md b/changes/1163-dense-lods.size.mddeleted file mode 100644index 6ab629f1..00000000--- a/changes/1163-dense-lods.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:mesh: `generateGeometryLods` could not reduce a textured geometry (its levels lost the uvs and failed to draw), and the attribute-keeping edge collapse that fixes it is about 2.1 KB gzip; on the Monster scan it cuts the frame from 5.5 to 1.7 ms, so the bytes buy that speed gain for scenes that call addLodsToDenseParts.-- scene:agent-scene+lazy: the agent scene loads the mesh engine lazily, and the same edge collapse (about 2.1 KB gzip) comes with it; it fixes level reduction of textured geometry and cuts the Monster scan frame from 5.5 to 1.7 ms.diff --git a/changes/1163-dense-parts.md b/changes/1163-dense-parts.mddeleted file mode 100644index 90a1fd11..00000000--- a/changes/1163-dense-parts.md+++ /dev/null@@ -1 +0,0 @@-- `denseSamples` is judged on the same part list at construction, on resize and after `replaceParts` (a scene near the threshold no longer flips on its first resize because the ground was counted, and a scene that gains or loses triangles is re-judged) (#1163).diff --git a/changes/1163-dense-parts.size.md b/changes/1163-dense-parts.size.mddeleted file mode 100644index b041b0ca..00000000--- a/changes/1163-dense-parts.size.md+++ /dev/null@@ -1,3 +0,0 @@-- goal:lit-box: the renderer constructor, resize() and replaceParts are core code every scene bundles, so the density fix cannot move out of it; about 25 to 36 bytes gzip.-- goal:gltf-box: the renderer constructor, resize() and replaceParts are core code every scene bundles, so the density fix cannot move out of it; about 25 to 36 bytes gzip.-- goal:agent-scene: the renderer constructor, resize() and replaceParts are core code every scene bundles, so the density fix cannot move out of it; about 25 to 36 bytes gzip.diff --git a/changes/1163-dense-resize.size.md b/changes/1163-dense-resize.size.mddeleted file mode 100644index 4fbad90c..00000000--- a/changes/1163-dense-resize.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- entry:core+lazy: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- scene:agent-scene: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- scene:agent-scene+lazy: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- scene:gltf-box: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- scene:gltf-box+lazy: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- scene:lit-box: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- scene:lit-box+lazy: re-judging denseSamples on resize (a canvas that crosses one triangle per pixel changes its MSAA count) is an agent-visible fix to an opt-in option, about 75 to 85 bytes gzip.-- goal:agent-scene: the renderer constructor and resize() are core code every scene bundles, so the fix cannot move out of it.-- goal:gltf-box: the renderer constructor and resize() are core code every scene bundles, so the fix cannot move out of it.-- goal:lit-box: the renderer constructor and resize() are core code every scene bundles, so the fix cannot move out of it.diff --git a/changes/1163-dense-samples.md b/changes/1163-dense-samples.mddeleted file mode 100644index 6aefdff8..00000000--- a/changes/1163-dense-samples.md+++ /dev/null@@ -1 +0,0 @@-- `denseSamples` renders scenes with more than one triangle per pixel with fewer MSAA samples (default off), which took a 1.6 million triangle frame from about 3.7 to 3.0 ms at 2 samples (#1163).diff --git a/changes/1163-dense-samples.size.md b/changes/1163-dense-samples.size.mddeleted file mode 100644index 6a4b0a37..00000000--- a/changes/1163-dense-samples.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: the denseSamples option and its triangle-per-pixel test, 130 bytes gzip, for a measured 18 percent shorter frame on a 1.6 million triangle mesh at 2 samples (#1163).-- entry:core+lazy: the same code (#1163).-- scene:lit-box: the same code, 112 bytes (#1163).-- scene:lit-box+lazy: the same code (#1163).-- scene:gltf-box: the same code, 125 bytes (#1163).-- scene:gltf-box+lazy: the same code (#1163).-- scene:agent-scene: the same code, 112 bytes (#1163).-- scene:agent-scene+lazy: the same code (#1163).-- goal:lit-box: the option lives in the renderer constructor, which every scene carries; it is the only place the sample count is chosen (#1163).-- goal:gltf-box: the same (#1163).-- goal:agent-scene: the same (#1163).diff --git a/changes/1163-light-fragment-path.md b/changes/1163-light-fragment-path.mddeleted file mode 100644index 22c04925..00000000--- a/changes/1163-light-fragment-path.md+++ /dev/null@@ -1 +0,0 @@-- The colour shader computes the sun shadow coordinate from the world position instead of carrying it as a varying, which cuts about 15 percent of the opaque pass on scans with about one triangle per pixel (#1163)diff --git a/changes/1163-light-fragment-path.size.md b/changes/1163-light-fragment-path.size.mddeleted file mode 100644index 46479ded..00000000--- a/changes/1163-light-fragment-path.size.md+++ /dev/null@@ -1,9 +0,0 @@-- entry:core: 70 bytes gzip so the colour program carries one varying less, which takes 2.0 ms (15 percent) off the opaque pass of a 5.2 million triangle scan (#1163), a measured speed gain.-- entry:core+lazy: 71 bytes gzip for the same change.-- scene:lit-box: 80 bytes gzip for the same change.-- scene:lit-box+lazy: 80 bytes gzip for the same change.-- scene:agent-scene: 88 bytes gzip for the same change.-- scene:agent-scene+lazy: 84 bytes gzip for the same change.-- goal:lit-box: 80 bytes gzip for the same change; goal:light work, the cost is two preprocessor branches in each shader.-- goal:gltf-box: 56 bytes gzip for the same change.-- goal:agent-scene: 88 bytes gzip for the same change.diff --git a/changes/1172-cook-main.md b/changes/1172-cook-main.mddeleted file mode 100644index bdfade51..00000000--- a/changes/1172-cook-main.md+++ /dev/null@@ -1 +0,0 @@-- The Cook path tracer matches Cycles more closely: bokeh corners sit at Cycles' blade rotation, metals and glass mixes keep their energy, and soft shadows from many lights clean up faster (#1172)diff --git a/changes/1172.size.md b/changes/1172.size.mddeleted file mode 100644index 69f47352..00000000--- a/changes/1172.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 9 bytes gzip; the same +9 shows on unrelated pull requests (#1262, #1263, #1266), and this change only shortens src/camera.js, so it is drift against the size base, not this change (#1172)diff --git a/changes/1175.md b/changes/1175.mddeleted file mode 100644index 77407b7e..00000000--- a/changes/1175.md+++ /dev/null@@ -1 +0,0 @@-- A glTF with its base color map on `TEXCOORD_1` beside maps on `TEXCOORD_0` keeps its base color: the base color map, layer maps and shadow alpha pass now read the second UV set when the map asks for it (#1175)diff --git a/changes/1175.size.md b/changes/1175.size.mddeleted file mode 100644index 7d3bf29d..00000000--- a/changes/1175.size.md+++ /dev/null@@ -1,5 +0,0 @@-- entry:material-layer-shaders: layer maps now pick their UV set through the shared mapUV helper, a fixed missing-texture bug for glTF files with maps on TEXCOORD_1 (70 bytes gzip).-- scene:agent-scene+lazy: the same change, 77 bytes gzip, a functional fix for glTF base colour on TEXCOORD_1.-- goal:agent-scene: the same change, 14 bytes gzip.-- goal:lit-box: the same change, 23 bytes gzip, the depth alpha pass and base colour map read the second UV set.-- scene:gltf-box+lazy: the same change, 145 bytes gzip, the second-set promotion rule and the identity transform test in the glTF fallbacks.diff --git a/changes/1177.md b/changes/1177.mddeleted file mode 100644index b4b18e10..00000000--- a/changes/1177.md+++ /dev/null@@ -1 +0,0 @@-- A glTF map whose `texCoord` is above 1 (a normal map on a third UV set, for example) is dropped with a warning naming the material instead of refusing the whole file; a normal map on `TEXCOORD_1` already loads (#1177)diff --git a/changes/1178-water-reflection.md b/changes/1178-water-reflection.mddeleted file mode 100644index 4b4b21b8..00000000--- a/changes/1178-water-reflection.md+++ /dev/null@@ -1 +0,0 @@-- Flat water that transmits light (glTF `KHR_materials_transmission` with an `ior` near 1.33) now shows the water reflection of the scene around it (#1178)diff --git a/changes/1178-water-reflection.size.md b/changes/1178-water-reflection.size.mddeleted file mode 100644index b35cc64e..00000000--- a/changes/1178-water-reflection.size.md+++ /dev/null@@ -1,5 +0,0 @@-- entry:liquid-shader: the liquid pass carries the mirror lookup (about 580 bytes gzip) so water that transmits light shows the building and sky in it, as the lit program already does for opaque water-- entry:spume: spume bundles the liquid shader, so it takes the same 580 bytes-- scene:gltf-box+lazy: a glTF with transmission loads the liquid shader, so the scene pays the same 600 bytes-- scene:agent-scene+lazy: the lazy shader code includes the liquid shader, the same 610 bytes-- goal:agent-scene: the scene was over its goal before this change (136589 bytes), and the 10 bytes it grows in the core bundle come from the shared mirror lookup moving into its own modulediff --git a/changes/1178.md b/changes/1178.mddeleted file mode 100644index 8140dc4b..00000000--- a/changes/1178.md+++ /dev/null@@ -1 +0,0 @@-- A glTF file with lights no longer gets the renderer's studio key and fill lights on top of its own: on a test scene mean luma went from 224 to 204, within 1 level of three.js with the same ACES curve (#1178)diff --git a/changes/1178.size.md b/changes/1178.size.mddeleted file mode 100644index aa55a49a..00000000--- a/changes/1178.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:gltf-box: 24 bytes gzip so a glTF file with lights stops getting the studio key and fill lights on top of its own, a visible brightening and wash-out (#1178).diff --git a/changes/1183.md b/changes/1183.mddeleted file mode 100644index 97334625..00000000--- a/changes/1183.md+++ /dev/null@@ -1 +0,0 @@-- The home page hero says "67js is now Crucible" under the hero buttons, with the new package name and a link to the changelog (#1183)diff --git a/changes/1183.size.md b/changes/1183.size.mddeleted file mode 100644index ac5871cb..00000000--- a/changes/1183.size.md+++ /dev/null@@ -1,5 +0,0 @@-- scene:agent-scene+lazy: the name Crucible is 3 characters longer than 67js in every message and identifier of the bundle, plus the deprecated Renderer67 export and the old format names for one release (173 bytes gzip)-- goal:gltf-box: same rename cost, 10 bytes gzip, with no feature change-- goal:agent-scene: same rename cost, 102 bytes gzip, with no feature change-- scene:agent-scene: the rename and the code that keeps the old 67js.scene and 67js.asset format names loading for one release (102 bytes gzip); both go away with the aliases-- goal:lit-box: same rename cost, 9 bytes gzip, with no feature changediff --git a/changes/1186-measured.md b/changes/1186-measured.mddeleted file mode 100644index f62c89e4..00000000--- a/changes/1186-measured.md+++ /dev/null@@ -1 +0,0 @@-- Native render jobs get 640 MiB of fixed memory allowance instead of 256, sized from measured native + OIDN video jobs, so a 960 x 540 film no longer runs within 50 MiB of the guard (#1186)diff --git a/changes/1186-software-adapter.md b/changes/1186-software-adapter.mddeleted file mode 100644index bb8d9700..00000000--- a/changes/1186-software-adapter.md+++ /dev/null@@ -1 +0,0 @@-- A native render job on a software Vulkan adapter (lavapipe on a Linux runner without a GPU) gets 2304 MiB more memory allowance, so the first 960 x 540 job no longer fails the memory guard (#1186)diff --git a/changes/1186.md b/changes/1186.mddeleted file mode 100644index 809caaf3..00000000--- a/changes/1186.md+++ /dev/null@@ -1 +0,0 @@-- Native path traced video jobs with OIDN get a memory allowance of 256 MiB plus 1 KiB a pixel, capped at 2 GiB, and the memory guard error names the limit and the observed and allowed MiB (#1186)diff --git a/changes/1188.md b/changes/1188.mddeleted file mode 100644index f9745850..00000000--- a/changes/1188.md+++ /dev/null@@ -1 +0,0 @@-- The browser GPU path tracer (WebGL2 and WebGPU) takes as many triangles as the device's texture or buffer limits hold, up to 2,000,000, so a skinned crowd over 400,000 triangles can be traced; a scene past the limit is refused with its triangle count, the limit and what to do (#1188).diff --git a/changes/1188.size.md b/changes/1188.size.mddeleted file mode 100644index 66b79c70..00000000--- a/changes/1188.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:path-tracer: +829 bytes gzip buy a device-sized triangle bound (WebGL2 texture and WebGPU buffer limits) and an error that names the count, the limit and the fix, so a 43-avatar skinned crowd can be traced in the page (#1188).-- entry:path-tracer-progressive: +860 bytes gzip come from the same device-sized triangle bound and the upload out-of-memory check the progressive tracer now shares with the still renderer (#1188).diff --git a/changes/1191.md b/changes/1191.mddeleted file mode 100644index 8cba0025..00000000--- a/changes/1191.md+++ /dev/null@@ -1 +0,0 @@-- The agent skills match the code again: path traced `stills` and `video` jobs, the 1920x1080 pixel cap, all pass names, and the mesh and walkmap tools (#1191).diff --git a/changes/1193.md b/changes/1193.mddeleted file mode 100644index d7b76f7d..00000000--- a/changes/1193.md+++ /dev/null@@ -1 +0,0 @@-- `scene_capabilities` reports `jobs.maximumFps` as 60 and `jobs.maximumFrames` from the same constants the render job validator uses, instead of a stale 30 (#1193)diff --git a/changes/1196.md b/changes/1196.mddeleted file mode 100644index 19618fa6..00000000--- a/changes/1196.md+++ /dev/null@@ -1 +0,0 @@-- Fixed render jobs with passes (`depth`, `normal`, `albedo`, `objectId`, `emission`) failing with WebGL error 1282 on a scene with a smooth metal part: the reflection probe now looks up its uniforms again when a pass links a new shader program (#1196).diff --git a/changes/1197-export-keeps-image-encoding.md b/changes/1197-export-keeps-image-encoding.mddeleted file mode 100644index 75f61311..00000000--- a/changes/1197-export-keeps-image-encoding.md+++ /dev/null@@ -1 +0,0 @@-- `exportGlb` and `scene_export_glb` keep the JPEG, PNG or WebP bytes of an imported texture instead of re-encoding it as an uncompressed PNG, so an imported 1.8 MB catalog asset no longer exports at about 6.5 times its size (#1197)diff --git a/changes/1197-export-keeps-image-encoding.size.md b/changes/1197-export-keeps-image-encoding.size.mddeleted file mode 100644index 534db5ea..00000000--- a/changes/1197-export-keeps-image-encoding.size.md+++ /dev/null@@ -1,3 +0,0 @@-- entry:core: an imported GLB texture keeps its JPEG, PNG or WebP bytes on export, and an imported 1.8 MB asset no longer exports at about 6.5 times its size (#1197)-- entry:core+lazy: same change as entry:core (#1197)-- goal:gltf-box: the same 53 bytes, the price of the export fix (#1197)diff --git a/changes/1197.md b/changes/1197.mddeleted file mode 100644index 5dba0406..00000000--- a/changes/1197.md+++ /dev/null@@ -1 +0,0 @@-- `scene_look` steps a busy PNG down by 0.75 until it fits the 1 MiB budget and reports the size it returned, and `closed-room-without-bake` no longer fires on a lone object whose bounds centre is inside it (#1197)diff --git a/changes/1197.size.md b/changes/1197.size.mddeleted file mode 100644index 63c48b69..00000000--- a/changes/1197.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: the closed-room test now starts from a point outside every part, which stops a false "closed room" warning on a lone object (about 180 bytes gzip).-- scene:agent-scene+lazy: same change, about 170 bytes gzip.-- goal:agent-scene: a fixed false positive in scene_inspect, a functional gain for agents; the growth is 181 bytes.diff --git a/changes/1199.md b/changes/1199.mddeleted file mode 100644index 26e04b65..00000000--- a/changes/1199.md+++ /dev/null@@ -1,2 +0,0 @@-- The site wears the Crucible brand: the pot mark and Plus Jakarta Sans Bold wordmark, Paper and charcoal themes with a light and dark switch, and the shared brand fonts (#1199)- media: /og.jpg The Crucible mark over a container port rendered by Cruciblediff --git a/changes/1204.md b/changes/1204.mddeleted file mode 100644index efece2b7..00000000--- a/changes/1204.md+++ /dev/null@@ -1 +0,0 @@-- A render job with `passes` no longer fails with GL error 1282 on a scene that has a baked irradiance grid: the grid looks its uniforms up again for the program a pass links (#1204).diff --git a/changes/1204.size.md b/changes/1204.size.mddeleted file mode 100644index 92d52cdd..00000000--- a/changes/1204.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the grid looks its uniforms up again for each program, which fixes GL error 1282 in render pass jobs on baked scenes (80 bytes gzip).-- goal:agent-scene: the same fix, 9 bytes gzip, a functional bug fix for agent render jobs.diff --git a/changes/1210.md b/changes/1210.mddeleted file mode 100644index 247da8ca..00000000--- a/changes/1210.md+++ /dev/null@@ -1 +0,0 @@-- glTF accessors are now decoded one primitive at a time during conversion and released after their last use, and `maxDecodedBytes` sets the decoded-bytes budget (#1210)diff --git a/changes/1210.size.md b/changes/1210.size.mddeleted file mode 100644index d2c86143..00000000--- a/changes/1210.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: lazy per-primitive accessor decoding lowers the peak of a large GLB load by about 8 percent (502 to 462 MiB for a 213 MiB file), and maxDecodedBytes gives the decoded-bytes budget its own named option, for 108 bytes gzip.-- scene:gltf-box+lazy: the same 107 bytes gzip, one load path shared with the glTF-only scene.-- goal:gltf-box: the scene was already over its goal, and the 108 bytes buy a measured peak-memory drop on every glTF load that the goal:light work targets.diff --git a/changes/1211.md b/changes/1211.mddeleted file mode 100644index bb3df146..00000000--- a/changes/1211.md+++ /dev/null@@ -1 +0,0 @@-- A native film or stills job waits one event-loop turn after each frame, so the addon's frame planes are freed and the Node process no longer grows about 9 to 13 MB per frame (#1211).diff --git a/changes/1216.md b/changes/1216.mddeleted file mode 100644index 9462683d..00000000--- a/changes/1216.md+++ /dev/null@@ -1 +0,0 @@-- A big thin ground box beside dense props no longer fails `scene_create` with "Set maximum size exceeded": the flush search stops at its tested-pair limit and reports the part as unchecked, and a 2 cm slab is no longer flagged `collapsed-node` (#1216)diff --git a/changes/1216.size.md b/changes/1216.size.mddeleted file mode 100644index b5ae5934..00000000--- a/changes/1216.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: grows 152 bytes gzip for the tested-pair limit, the named part in the unchecked leftover and the in-plane pruning of the flush search, which turns a crash (and a 5 minute stall) on a 400 m ground beside dense props into a pass of about 7 s.diff --git a/changes/1217-mask.md b/changes/1217-mask.mddeleted file mode 100644index e38b6f30..00000000--- a/changes/1217-mask.md+++ /dev/null@@ -1 +0,0 @@-- A placed asset material override (and a recipe material) takes an `emissionMask` ramp along an object axis, so a part glows only at one end while its texture and shading stay (#1217)diff --git a/changes/1217-mask.size.md b/changes/1217-mask.size.mddeleted file mode 100644index e09295e5..00000000--- a/changes/1217-mask.size.md+++ /dev/null@@ -1,8 +0,0 @@-- entry:material-layer-shaders: the GLSL of the mask ramp and its uniform bind, loaded only by scenes that use a layer shader (142 bytes gzip).-- entry:gltf-box+lazy: the same chunk, which a glTF scene loads with its layer shaders (196 bytes gzip).-- scene:agent-scene: the field, the loader marker and the lazy validator hook in the agent scene (185 bytes gzip; the validator itself sits in a module that loads only for a recipe with emissionMask).-- scene:agent-scene+lazy: the same change (1035 bytes gzip, including that validator module).-- goal:agent-scene: the same change, 185 bytes gzip, a functional capability for agents.-- scene:gltf-box+lazy: the same chunk as entry:gltf-box+lazy, 196 bytes gzip.-- goal:lit-box: the renderer feature check and the layer state values, 51 bytes gzip, the whole core cost of the mask.-- goal:gltf-box: the same core cost plus the layer shader chunk, 54 bytes gzip.diff --git a/changes/1217.md b/changes/1217.mddeleted file mode 100644index 14f4b84c..00000000--- a/changes/1217.md+++ /dev/null@@ -1 +0,0 @@-- A placed asset's `asset.materials[]` override can take an `id` that `motion` material tracks drive, so `emission` and `emissiveColor` fade a heat glow without rebuilding the scene (#1217).diff --git a/changes/1217.size.md b/changes/1217.size.mddeleted file mode 100644index 2e34dac6..00000000--- a/changes/1217.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: a handle on an asset material override and a live material view let motion tracks fade emission, an agent-facing capability that otherwise costs a scene_create per heat state (186 bytes gzip).-- scene:agent-scene+lazy: the same change (526 bytes gzip).-- goal:agent-scene: the same change, 186 bytes gzip, a functional capability for agents.diff --git a/changes/1218-rebake-warning.md b/changes/1218-rebake-warning.mddeleted file mode 100644index c5c4ffa1..00000000--- a/changes/1218-rebake-warning.md+++ /dev/null@@ -1 +0,0 @@-- A `scene_apply` whose rebake cannot run (part cap, path work limit) now keeps the edit, drops the old bake and returns a `bake-skipped` warning with how to fit, and the irradiance grid and probe part errors say they trace on the CPU (#1218)diff --git a/changes/1218-rebake-warning.size.md b/changes/1218-rebake-warning.size.mddeleted file mode 100644index 710c6bf4..00000000--- a/changes/1218-rebake-warning.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: 67 bytes gzip of core wire the skipped-bake reason into the staged state and the inspect diagnostics; the catch and the message live in the lazy bake module.-- scene:agent-scene+lazy: 424 bytes gzip for the bake-skipped diagnostic with its fix text and the refusal wrapper, so a size limit stops failing a whole scene_apply on a big scene.-- goal:agent-scene: the growth is the hook that makes an over-limit rebake a warning instead of a failed apply; nothing else in the core moved.diff --git a/changes/1218.md b/changes/1218.mddeleted file mode 100644index c5ca076b..00000000--- a/changes/1218.md+++ /dev/null@@ -1 +0,0 @@-- Breaking: `bake.irradianceGrid` takes 8 to 40 nodes, so a value that passes validation always fits the 65,536 node cap; 41 to 48 is no longer accepted (use `true`, which still reaches 48 along the longest side). The recipe bake also honours the sun link `environment.keyAffects` (#1218)diff --git a/changes/1218.size.md b/changes/1218.size.mddeleted file mode 100644index be970c2c..00000000--- a/changes/1218.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the bake now honours the environment sun link (`keyAffects`) and the irradiance grid range check is tied to the node cap, about 72 bytes gzip; without them a sun link is silently ignored by the baked bounce and a validated grid size can fail at build time (#1218)-- goal:agent-scene: 8 bytes gzip from the same sun-link and grid-range checks; the scene bundle is already over its goal (#1218)diff --git a/changes/1219-apply-graph-cache.size.md b/changes/1219-apply-graph-cache.size.mddeleted file mode 100644index 2f06daad..00000000--- a/changes/1219-apply-graph-cache.size.md+++ /dev/null@@ -1,5 +0,0 @@-- entry:core: 112 bytes gzip for `replaceLocalMatrices`, the scene graph method that moves nodes in place; it removes the graph rebuild (about 2.5 ms of 4.9 ms) from every pose edit at 961 parts.-- entry:core+lazy: the same 112 bytes, no other module grew.-- scene:agent-scene: 564 bytes gzip buy a pose edit at 961 parts of about 2 ms instead of about 5 ms (graph moved in place, checked nodes not checked again, reflection key compared as numbers), which leaves a 60 fps game frame most of its 16.7 ms.-- scene:agent-scene+lazy: the same growth, 658 bytes with the reflection stage.-- goal:agent-scene: the growth is a per-frame cost cut (pose edit at 961 parts from about 5 ms to about 2 ms), the same case the earlier #1219 size notes accepted; the scene is over its goal for other reasons.diff --git a/changes/1219-apply-graph.size.md b/changes/1219-apply-graph.size.mddeleted file mode 100644index 348d6089..00000000--- a/changes/1219-apply-graph.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: 164 bytes gzip buy a pose edit at 961 parts that takes about 9 ms instead of about 20 ms (a whole-recipe clone, a validation compile and an undo copy fewer), which a per-frame game scene system needs.-- scene:agent-scene+lazy: the same 164 bytes as the core scene, no other module grew.-- goal:agent-scene: the growth is the shared-node field edit copy; the scene is over its goal for other reasons and this change moves a frame loop from 14 fps toward 60.diff --git a/changes/1219.md b/changes/1219.mddeleted file mode 100644index 9540343f..00000000--- a/changes/1219.md+++ /dev/null@@ -1 +0,0 @@-- A pose edit through `scene.apply` at 961 parts now takes about 2 ms instead of about 60 ms, and a game frame that sends the same poses as the last one skips `scene.apply` (#1219)diff --git a/changes/1219.size.md b/changes/1219.size.mddeleted file mode 100644index e86be512..00000000--- a/changes/1219.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: 153 bytes gzip buy a threefold faster pose edit at 961 parts (about 60 ms down to about 20 ms), which a per-frame game scene system needs.-- scene:agent-scene+lazy: the same 263 bytes as the core scene, no other module grew.-- goal:agent-scene: the growth is the cached part bounds, the string size cache and the field-edit fast path; the scene is over its goal for other reasons and this change makes a frame loop playable.diff --git a/changes/1221.md b/changes/1221.mddeleted file mode 100644index f9a48052..00000000--- a/changes/1221.md+++ /dev/null@@ -1 +0,0 @@-- A render job whose frame is cut to fit the render budget now names the budget the full frame needs and the `--memory` value to restart with, instead of only the smaller size (#1221)diff --git a/changes/1221.size.md b/changes/1221.size.mddeleted file mode 100644index d86176ed..00000000--- a/changes/1221.size.md+++ /dev/null@@ -1,6 +0,0 @@-- scene:lit-box: `planRenderSize` returns `requiredMemoryMiB` (about 16 bytes gzip), so a cut frame can say which `--memory` renders it at full size-- goal:lit-box: the same field; without it an agent sees 1532x862 for a 1920x1080 job and cannot tell which budget fixes it-- scene:gltf-box: the same `requiredMemoryMiB` field in the shared budget planner-- goal:gltf-box: the same field, for the same reason-- scene:agent-scene: the same `requiredMemoryMiB` field in the shared budget planner-- goal:agent-scene: the same field; it turns a silent downscale into a message that names the fixdiff --git a/changes/1223.md b/changes/1223.mddeleted file mode 100644index c53917fc..00000000--- a/changes/1223.md+++ /dev/null@@ -1 +0,0 @@-- Line art can be set per node: `environment.lines.nodes` takes `{ "<node id>": { lines: false | silhouetteOnly, color } }`, and the `objectId` pass metrics carry `objectNodes` (#1223)diff --git a/changes/1223.size.md b/changes/1223.size.mddeleted file mode 100644index 1dcee00a..00000000--- a/changes/1223.size.md+++ /dev/null@@ -1,4 +0,0 @@-- entry:line-art: per-node line settings add a node-table shader path, validation and the table upload (1047 bytes gzip); the addon loads only for scenes with lines, and sketch-to-3D composites cannot be drawn without it (#1223).-- scene:agent-scene+lazy: the same line art addon is part of this lazy bundle (1110 bytes gzip), plus the capabilities text and the part group field.-- goal:agent-scene: 37 bytes gzip of capabilities text so an agent finds lines.nodes and metrics.objectNodes without reading the docs (#1223).-- scene:agent-scene: the check that fails a lines.nodes ID no node carries (about 50 bytes gzip), so a typo in a node ID stops with a message instead of silently doing nothing (#1223).diff --git a/changes/1228.md b/changes/1228.mddeleted file mode 100644index e062146e..00000000--- a/changes/1228.md+++ /dev/null@@ -1 +0,0 @@-- The WebGL2 GPU path tracer draws every part again: on ANGLE D3D11 it drew only the first part of a scene, because a nested texel fetch read every later material as shadow-only (#1228)diff --git a/changes/1241.md b/changes/1241.mddeleted file mode 100644index b495cbf4..00000000--- a/changes/1241.md+++ /dev/null@@ -1 +0,0 @@-- An `image` render job takes up to 3840x2160 pixels instead of 1920x1080, and an image frame the `--memory` budget would shrink is refused with the `--memory` it needs; stills and video keep 1920x1080 (#1241)diff --git a/changes/1241.size.md b/changes/1241.size.mddeleted file mode 100644index cdf955ca..00000000--- a/changes/1241.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: two constants for the video and stills size limits add 13 bytes gzip; they let an image job ask for 2880x1620 (#1241).diff --git a/changes/1242.md b/changes/1242.mddeleted file mode 100644index 6f456183..00000000--- a/changes/1242.md+++ /dev/null@@ -1 +0,0 @@-- The WebGPU real-time tracer and the GPU light bake bound their scene by the device's triangle limit, up to 2,000,000, like the still render, instead of 400,000 (#1242).diff --git a/changes/1242.size.md b/changes/1242.size.mddeleted file mode 100644index 14dcaef6..00000000--- a/changes/1242.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 9 bytes gzip; the same +9 shows on three unrelated pull requests (#1262, #1263, #1242), so it is drift against the size base, not this change (#1242)diff --git a/changes/1243.md b/changes/1243.mddeleted file mode 100644index 66e1eae4..00000000--- a/changes/1243.md+++ /dev/null@@ -1 +0,0 @@-- The path tracers, native Cook included, read a map on `TEXCOORD_1` with the second UV set, so a glTF base color on `TEXCOORD_1` beside a `TEXCOORD_0` normal map no longer samples UV (0, 0); a part with a map on `uvChannel` 1 and no `geometry.lightUvs` is refused with what to set (#1243)diff --git a/changes/1243.size.md b/changes/1243.size.mddeleted file mode 100644index 2fda0207..00000000--- a/changes/1243.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:path-tracer: a glTF base colour on TEXCOORD_1 traced as UV (0, 0) before; the second UV set, per-map set selection and the lightUvs check add under 1 kB gzip.-- entry:path-tracer-progressive: same change as entry:path-tracer, loaded by the progressive tracer.diff --git a/changes/1244.md b/changes/1244.mddeleted file mode 100644index 0a00e4ed..00000000--- a/changes/1244.md+++ /dev/null@@ -1 +0,0 @@-- `crucible-mcp` stops at startup with the real message for a bad `--packs`, `--assets`, `--audio`, `--ffmpeg`, `--browser` or `--branches`/`--relations`/`--findings` value, `scene_restore` rejects two targets, and terrain and geometry-budget errors name the range, id and numbers (#1244)diff --git a/changes/1244.size.md b/changes/1244.size.mddeleted file mode 100644index 6ebd65a6..00000000--- a/changes/1244.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: the geometry preflight error now names the geometry id and the vertex and byte numbers, and terrain reports its real range, so an agent fixes the recipe in one call (127 bytes gzip).-- scene:agent-scene+lazy: the same change (106 bytes gzip).-- goal:agent-scene: the same change, 127 bytes gzip, an agent-facing error that otherwise costs a second failed call.diff --git a/changes/1245.md b/changes/1245.mddeleted file mode 100644index 45f92c9f..00000000--- a/changes/1245.md+++ /dev/null@@ -1 +0,0 @@-- Docs and llms.txt now say the `quality` render setting ships, the checkpoint limit is 16, and the tool-cost total points to `scripts/measure-tool-cost.mjs` instead of a stale number; a test keeps the limit in step (#1245).diff --git a/changes/1246.md b/changes/1246.mddeleted file mode 100644index 2c45a303..00000000--- a/changes/1246.md+++ /dev/null@@ -1 +0,0 @@-- A failed job release, status write or browser launch no longer breaks the MCP runtime: a retry finishes the release, `close` always frees the store lock, and a recipe over the 64 MiB asset limit is refused before any file is read (#1246)diff --git a/changes/1247.md b/changes/1247.mddeleted file mode 100644index 2283bbb1..00000000--- a/changes/1247.md+++ /dev/null@@ -1 +0,0 @@-- Scene builders refuse bad input when it is set: text outside printable ASCII names the character and font, motion tracks are checked for out-of-range values, and a terrain `seed` must be a whole number from 0 to 2147483647 (#1247)diff --git a/changes/1247.size.md b/changes/1247.size.mddeleted file mode 100644index cae3e059..00000000--- a/changes/1247.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the set-time track sweep, the glyph error text and the seed check turn crashes in the middle of a render into errors at create time that name the field and the time (+412 bytes gzip)diff --git a/changes/1248-null-entries.md b/changes/1248-null-entries.mddeleted file mode 100644index 84c7ee62..00000000--- a/changes/1248-null-entries.md+++ /dev/null@@ -1 +0,0 @@-- glTF import names the field and index of a `null` or non-object entry in `nodes`, `meshes`, `accessors` and the other arrays instead of a "Cannot read properties" TypeError (#1248)diff --git a/changes/1248-null-entries.size.md b/changes/1248-null-entries.size.mddeleted file mode 100644index 7e532ce4..00000000--- a/changes/1248-null-entries.size.md+++ /dev/null@@ -1,4 +0,0 @@-- scene:gltf-box: named errors instead of TypeErrors on malformed glTF arrays, 92 bytes gzip.-- scene:gltf-box+lazy: the same code as scene:gltf-box.-- goal:gltf-box: 92 bytes for named errors instead of TypeErrors on malformed glTF arrays (#1248)-- goal:agent-scene: 9 bytes gzip of drift against the size base that every recent pull request shows (#1248)diff --git a/changes/1248.md b/changes/1248.mddeleted file mode 100644index 27c3ff66..00000000--- a/changes/1248.md+++ /dev/null@@ -1 +0,0 @@-- glTF import and export reject more bad files by name: dangling indices name their field, meshopt buffer views count against `maxDecodedBytes`, and `exportGlb` throws when a light map or `uvChannel: 1` map has no `geometry.lightUvs` (#1248)diff --git a/changes/1248.size.md b/changes/1248.size.mddeleted file mode 100644index 7f6d1afd..00000000--- a/changes/1248.size.md+++ /dev/null@@ -1,5 +0,0 @@-- entry:core: validation of untrusted files and exports (a lightUvs check, prerelease versions) with error messages that tell an agent what to pass instead.-- entry:core+lazy: the same code as entry:core.-- scene:gltf-box: field-named errors for dangling glTF indices, byteStride and sparse index checks and the meshopt size bound replace raw TypeErrors and a 4.47 GiB allocation from a 691-byte file.-- scene:gltf-box+lazy: the same code as scene:gltf-box.-- goal:gltf-box: the growth is bounds checking on untrusted input, about 580 bytes gzip, with no feature added to the render path.diff --git a/changes/1249-published-site.md b/changes/1249-published-site.mddeleted file mode 100644index 8fe760f8..00000000--- a/changes/1249-published-site.md+++ /dev/null@@ -1 +0,0 @@-- The site build now publishes every guide that `llms.txt` and the other guides link to, the install guide no longer links a release the site dropped, the docs say the three optional packages are source-checkout only, and the addon check accepts release candidate versions (#1249)diff --git a/changes/1249-published-site.size.md b/changes/1249-published-site.size.mddeleted file mode 100644index 931bf778..00000000--- a/changes/1249-published-site.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 9 bytes gzip of drift against the size base, the same +9 described in changes/1263.size.md; this change touches nothing under src/ (#1249)diff --git a/changes/1256-queue-wait.md b/changes/1256-queue-wait.mddeleted file mode 100644index c1eb3c69..00000000--- a/changes/1256-queue-wait.md+++ /dev/null@@ -1 +0,0 @@-- A render job's deadline counts render time only, its queue wait is bounded apart, and `scene_create` starts a fresh runtime after a failed one without a server restart (#1256)diff --git a/changes/1256.md b/changes/1256.mddeleted file mode 100644index 80912df2..00000000--- a/changes/1256.md+++ /dev/null@@ -1 +0,0 @@-- A render job that hits its elapsed deadline under a shared GPU now gives the scene page 120 seconds to come back, and its failure says to submit the job again (#1256, #1224).diff --git a/changes/1258.md b/changes/1258.mddeleted file mode 100644index 63d852b0..00000000--- a/changes/1258.md+++ /dev/null@@ -1 +0,0 @@-- `crucible-mcp` takes `--gpu-queue` and `--gpu-queue-wait`: an optional queue command that looks, reviews, bakes and render jobs wait for before they take the GPU, so the server can share a card queue (#1258)diff --git a/changes/1259.md b/changes/1259.mddeleted file mode 100644index 236c9153..00000000--- a/changes/1259.md+++ /dev/null@@ -1 +0,0 @@-- The path tracer export keeps an area light aimed along `up` as a rectangle, reads texture `wrapS` and `wrapT` per axis on every tracer, and turns skinned normals correctly (#1259)diff --git a/changes/1262.md b/changes/1262.mddeleted file mode 100644index 4bf64085..00000000--- a/changes/1262.md+++ /dev/null@@ -1 +0,0 @@-- A client or authority restarted on the same peer id now resyncs on its own instead of staying silent (#1262)diff --git a/changes/1262.size.md b/changes/1262.size.mddeleted file mode 100644index aa1cefe0..00000000--- a/changes/1262.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:netcode: restarted peers resync instead of staying dead; the ordered run id, the run-matched ack, the null-ack reset and the session check cost 348 bytes gzip (#1262)-- goal:agent-scene: 9 bytes gzip for the same restart resync; netcode is bundled into the agent scene, and a dead peer after restart is worse than 9 bytes (#1262)diff --git a/changes/1263.md b/changes/1263.mddeleted file mode 100644index 91173af7..00000000--- a/changes/1263.md+++ /dev/null@@ -1 +0,0 @@-- `startGame` no longer moves a player or NPC under a translated parent on the first frame, and the `restart` rule action, the R key and the new `game.restart()` put the game back at its start, so a won or lost exported game can be played again (#1263)diff --git a/changes/1263.size.md b/changes/1263.size.mddeleted file mode 100644index 20bf37b9..00000000--- a/changes/1263.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the game restart (player, camera, NPCs, animation graphs, physics world), the tap and R key restart and the parent-transform check at set-game and export add about 250 bytes gzip; without them an exported game cannot be played twice and a rotated parent silently misplaces the player.-- goal:agent-scene: 9 bytes gzip; the same +9 shows on three unrelated pull requests (#1262, #1263, #1242), so it is drift against the size base, not this change (#1263)diff --git a/changes/1264.md b/changes/1264.mddeleted file mode 100644index ae05b681..00000000--- a/changes/1264.md+++ /dev/null@@ -1 +0,0 @@-- Render jobs count frames without float noise, `render-media` refuses audio shorter than the video, and a tiny `audio.offset` reaches ffmpeg as a plain decimal (#1264)diff --git a/changes/1265.md b/changes/1265.mddeleted file mode 100644index 69350c33..00000000--- a/changes/1265.md+++ /dev/null@@ -1 +0,0 @@-- `crucible-addon-check` resolves the entry like the server, rejects unsatisfiable peer ranges, accepts `--version=X` and has `--help`; `crucible-mcp --help` lists every flag (#1265)diff --git a/changes/1266.md b/changes/1266.mddeleted file mode 100644index da2fab90..00000000--- a/changes/1266.md+++ /dev/null@@ -1 +0,0 @@-- Disposing a renderer made with `deferProgramLink` now frees the programs still linking, and `resize()` on a lost context returns instead of throwing (#1266)diff --git a/changes/1266.size.md b/changes/1266.size.mddeleted file mode 100644index d8c6a279..00000000--- a/changes/1266.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: a disposed deferred renderer no longer leaks 5 programs and 10 shaders or rebuilds targets, and resize survives a lost context; about 140 bytes gzip.-- entry:core+lazy: same fix as entry:core.-- scene:lit-box: same fix as entry:core.-- scene:lit-box+lazy: same fix as entry:core.-- scene:gltf-box: same fix as entry:core.-- scene:gltf-box+lazy: same fix as entry:core.-- scene:agent-scene: same fix as entry:core.-- scene:agent-scene+lazy: same fix as entry:core.-- goal:lit-box: a leak fix of about 140 bytes in the renderer, not a feature.-- goal:gltf-box: a leak fix of about 140 bytes in the renderer, not a feature.-- goal:agent-scene: a leak fix of about 140 bytes in the renderer, not a feature.diff --git a/changes/1267.md b/changes/1267.mddeleted file mode 100644index bef8f5fe..00000000--- a/changes/1267.md+++ /dev/null@@ -1 +0,0 @@-- A looping clip retargeted with `retargetClip` or exported from an authored rig now ends on its end pose instead of the first pose, and `createCrowd` throws a `RangeError` for bad `columns` and `size` (#1267)diff --git a/changes/1267.size.md b/changes/1267.size.mddeleted file mode 100644index 6b03e957..00000000--- a/changes/1267.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the last baked or retargeted frame of a looping clip was the first pose, and createCrowd gave NaN positions for bad columns or size; the fix is a few lines and two error messages-- goal:agent-scene: the baked GLB of a looping rig motion ended on the wrong pose, a correctness fix of 26 bytes gzipdiff --git a/changes/1268.md b/changes/1268.mddeleted file mode 100644index 47b71f8c..00000000--- a/changes/1268.md+++ /dev/null@@ -1 +0,0 @@-- Volume fog detail fields, the volume fog density motion track (now 0 to 10), a mesh volume's missing part and a second grid volume fail when the scene or an edit is built, with the render-time messages (#1268)diff --git a/changes/1268.size.md b/changes/1268.size.mddeleted file mode 100644index 8b137c2a..00000000--- a/changes/1268.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: 322 bytes gzip for the fog detail, part and one-grid checks that make a bad volume fail when the scene is built, not at the first frame-- entry:core+lazy: the same 323 bytes-- scene:lit-box: the same checks, +297 bytes-- scene:lit-box+lazy: the same checks, +296 bytes-- scene:gltf-box: the same checks, +303 bytes-- scene:gltf-box+lazy: the same checks, +302 bytes-- scene:agent-scene: the same checks, +567 bytes-- scene:agent-scene+lazy: the same checks, +154 bytes-- goal:lit-box: a bad volume now fails when the scene is built with the render message; the lead accepts agent-facing validation growth of a few hundred bytes-- goal:gltf-box: a bad volume now fails when the scene is built with the render message; the lead accepts agent-facing validation growth of a few hundred bytes-- goal:agent-scene: a bad volume now fails when the scene is built with the render message; the lead accepts agent-facing validation growth of a few hundred bytesdiff --git a/changes/1269.md b/changes/1269.mddeleted file mode 100644index 6310ddb5..00000000--- a/changes/1269.md+++ /dev/null@@ -1 +0,0 @@-- Mesh `bevel` caps `profile` at 0.99 and shrinks the overlap clamp for concave profiles, `inset_region` clamps thickness at the face limit and reports it, and `lintMesh` with `geometry: true` flags `folded-face` (#1269).diff --git a/changes/1269.size.md b/changes/1269.size.mddeleted file mode 100644index 3eeceb62..00000000--- a/changes/1269.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: fixes three ways bevel and inset_region returned inside-out or folded solids that lint missed: the inset limit search, the depth-aware bevel clamp and the folded-face lint check add about 700 bytes gzip.diff --git a/changes/1281.md b/changes/1281.mddeleted file mode 100644index 5dc80c82..00000000--- a/changes/1281.md+++ /dev/null@@ -1 +0,0 @@-- The games runtime calls `onEvent` for every event and caps only the stored log, a rules part starts as hidden as the scene node it is named after unless it sets `visible`, and `reportContact` is ignored after the game is won or lost (#1281).diff --git a/changes/1281.size.md b/changes/1281.size.mddeleted file mode 100644index d57fda78..00000000--- a/changes/1281.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 9 bytes gzip; the same +9 is already accepted in changes/1263.size.md and shows on pull requests that do not touch the scene, so it is drift against the size base, not this change (#1281)diff --git a/changes/1282.md b/changes/1282.mddeleted file mode 100644index 3903b5d7..00000000--- a/changes/1282.md+++ /dev/null@@ -1 +0,0 @@-- Asset loading counts a textured pack GLB at its embedded size, and the PLY splat loader skips elements before `vertex`, accepts CRLF headers, and rejects non-finite scales and a non-positive `maximumCount` with a clear error (#1282).diff --git a/changes/1282.size.md b/changes/1282.size.mddeleted file mode 100644index c384ea34..00000000--- a/changes/1282.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:gaussian-splats: reads PLY files with elements before vertex and CRLF headers and rejects non-finite scales and a bad maximumCount with errors that name the fix, 301 bytes gzip against 268 allowed.-- goal:agent-scene: 6 to 9 bytes gzip, the same drift against the size base seen on #1262, #1263 and #1242; this change only touches the gaussian-splat PLY reader and the asset catalog script, so it is not from this change (#1282)diff --git a/changes/1283.md b/changes/1283.mddeleted file mode 100644index 8356d2a2..00000000--- a/changes/1283.md+++ /dev/null@@ -1 +0,0 @@-- `remesh` no longer returns an inflated or non-manifold shell for a cube, `remesh` and `boolean` reject an inside-out mesh and name `flip_normals`, and `subdivide` no longer fails on a loose vertex (#1283)diff --git a/changes/1283.size.md b/changes/1283.size.mddeleted file mode 100644index a0a9c0be..00000000--- a/changes/1283.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: 149 bytes gzip for the inside-out check in remesh, the closed-output check, the on-surface sign rule that stops the cube remesh from returning a broken shell, and the loose-vertex skip in subdivide; each turns a silent bad mesh or an unnamed error into a result or a fix hint an agent can follow.-- goal:agent-scene: a few bytes gzip over the goal; the same drift shows on unrelated pull requests (#1262, #1263, #1242) against the size base, and this change adds the hole check and the optional quad-builder argument in remesh, a one-line cost (#1283)diff --git a/changes/1284.md b/changes/1284.mddeleted file mode 100644index 5e272b68..00000000--- a/changes/1284.md+++ /dev/null@@ -1 +0,0 @@-- `render-media` keeps the audio on the browser-encoded H.264 path and ends PCM and AAC audio with the picture, fails a video that lost its audio, writes 8 distinct stills as `frame-0` to `frame-7`, and rounds a derived video size to even (#1284)diff --git a/changes/1284.size.md b/changes/1284.size.mddeleted file mode 100644index 4ff53329..00000000--- a/changes/1284.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 9 bytes gzip drift against the size base, no src change (#1284)diff --git a/changes/1291.md b/changes/1291.mddeleted file mode 100644index 4f540684..00000000--- a/changes/1291.md+++ /dev/null@@ -1 +0,0 @@-- `remesh` no longer returns an edge shared by four quads on thin, bumpy surfaces; when it cannot separate the sheets, it throws and names a different `voxelSize` (#1291)diff --git a/changes/1291.size.md b/changes/1291.size.mddeleted file mode 100644index e9eaa0e9..00000000--- a/changes/1291.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 744 bytes gzip for splitting saddle cell vertices in remesh, which turns a silently returned non-manifold shell (an edge shared by four quads) on thin, bumpy surfaces into a closed manifold mesh that later edits and lint accept (#1291).diff --git a/changes/1300.md b/changes/1300.mddeleted file mode 100644index 351a14d8..00000000--- a/changes/1300.md+++ /dev/null@@ -1 +0,0 @@-- The WebGL and WebGPU path tracers read each map of a material with its own UV set, so a floor with a `TEXCOORD_0` base colour and a `TEXCOORD_1` emissive map matches the CPU tracer; a device with limited texture or buffer size now holds about 10 percent fewer triangles (#1300)diff --git a/changes/1300.size.md b/changes/1300.size.mddeleted file mode 100644index e70e910c..00000000--- a/changes/1300.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:path-tracer: the second UV set adds the per-map selection to both kernels (about 1.2 KB of shader source, kernel size limits raised by 1 KB); a mixed-set material now matches the CPU tracer, where the GPU tracers read every map with the base colour's set. The GPU triangle table grows from 9 to 10 texels (11 percent more bytes per triangle); the native Cook record stays at 9.diff --git a/changes/1302.md b/changes/1302.mddeleted file mode 100644index c660e732..00000000--- a/changes/1302.md+++ /dev/null@@ -1 +0,0 @@-- An `sdf` geometry now splits saddle cell vertices the way `remesh` does, so every edge joins exactly two quads when parts touch along an edge; a saddle it cannot separate throws an error that names the fix (#1302).diff --git a/changes/1302.size.md b/changes/1302.size.mddeleted file mode 100644index 76089ca1..00000000--- a/changes/1302.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:modeling: 845 bytes gzip because the sdf geometry now bundles the saddle splitter that only remesh used, so edge-touching sdf parts no longer return four-quad edges (#1302).-- scene:agent-scene+lazy: 85 bytes gzip for the sdf geometry to share the saddle splitter with remesh, so edge-touching sdf parts mesh with two quads on every edge (#1302).diff --git a/changes/1303.md b/changes/1303.mddeleted file mode 100644index c083c522..00000000--- a/changes/1303.md+++ /dev/null@@ -1 +0,0 @@-- A placed asset takes `flatten: { color?, unlit? }`, which sets one flat colour or unlit on every material of that asset in one call and leaves the rest of the scene unchanged (#1303)diff --git a/changes/1303.size.md b/changes/1303.size.mddeleted file mode 100644index d3180cec..00000000--- a/changes/1303.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: 98 bytes gzip for the `flatten` hint in the placedAssets capability text, which lets an agent flatten or unlit every material of an imported asset in one call instead of one override per material name (#1303).-- scene:agent-scene+lazy: 439 bytes gzip for the `flatten` validation and the flatten-aware material output in the assets module, the agent-facing capability this issue asks for (#1303).-- goal:agent-scene: 98 bytes gzip, the capability text an agent reads to find `flatten`; the code itself sits in the lazy assets module (#1303).diff --git a/changes/1304-smoke-sha.md b/changes/1304-smoke-sha.mddeleted file mode 100644index ec6273f2..00000000--- a/changes/1304-smoke-sha.md+++ /dev/null@@ -1 +0,0 @@-- The smokes `film-timeline`, `object-motion-blur` and `path-tracer` ask `scene_render_result` for the detailed answer, and the render docs and skill say file digests come with `response_format: "detailed"` (#1304).diff --git a/changes/1311.md b/changes/1311.mddeleted file mode 100644index 0b82a729..00000000--- a/changes/1311.md+++ /dev/null@@ -1 +0,0 @@-- `scene_create` and `set-environment` accept everything `scene_export_recipe` writes, including all joint fields and motion track `extrapolation`, and a malformed motion track names its bad field and the allowed ones (#1311)diff --git a/changes/1311.size.md b/changes/1311.size.mddeleted file mode 100644index fe8b5cb8..00000000--- a/changes/1311.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene+lazy: the joint fields line in scene_capabilities physics, the spring joint type and the grid base in the mesh base error, so an agent can author breakable, motorised and spring joints and decals from the capability text alone-- scene:agent-scene: same bytes (125 gzip); the capability text is what an agent reads before it writes a joint or mesh base-- goal:agent-scene: same bytes; an exported recipe loads again, and the capability text names every joint field the backend takesdiff --git a/changes/1312.md b/changes/1312.mddeleted file mode 100644index 9d0e7f0d..00000000--- a/changes/1312.md+++ /dev/null@@ -1 +0,0 @@-- The shipped guides and the build-crucible-scene skill now import only public subpaths, and new addon entries `crucible/addons/reflection-probes`, `fracture`, `morph` and `skinning` back the examples that had none (#1312)diff --git a/changes/1313.md b/changes/1313.mddeleted file mode 100644index 870ca53e..00000000--- a/changes/1313.md+++ /dev/null@@ -1 +0,0 @@-- Memory and size limits now agree with their messages: an `.hdr` over 8 MiB is refused like a `.cube` over 4 MiB, and `scene_asset_catalog` skips an oversized `.hdr`, `.cube` or image with a reason in `skipped` instead of failing the listing (#1313)diff --git a/changes/1313.size.md b/changes/1313.size.mddeleted file mode 100644index 49d46b0b..00000000--- a/changes/1313.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: the budget error and the renderer option carry the runtime's memory ceiling, so an agent is never told to pass a --memory the server refuses (about 150 bytes of message text)-- entry:core+lazy: same as entry:core-- scene:lit-box: same as entry:core-- scene:lit-box+lazy: same as entry:core-- scene:gltf-box: same as entry:core-- scene:gltf-box+lazy: same as entry:core-- scene:agent-scene: same as entry:core-- scene:agent-scene+lazy: same as entry:core-- goal:lit-box: 146 bytes of agent-facing error text that fixes an unreachable instruction-- goal:gltf-box: 171 bytes, the same text plus the shared image header reader moved out of glTF validation-- goal:agent-scene: 137 bytes, the same textdiff --git a/changes/1314.md b/changes/1314.mddeleted file mode 100644index cadaffb8..00000000--- a/changes/1314.md+++ /dev/null@@ -1 +0,0 @@-- Pack zips and the `sha256` in `/packs/index.json` no longer change with the checkout's line endings: `.gitattributes` pins text files to LF, and `createZip` turns CRLF into LF in text entries, so a Windows checkout and CI publish the same archive hash for the same SHA (#1314).diff --git a/changes/1319.md b/changes/1319.mddeleted file mode 100644index ad52afa8..00000000--- a/changes/1319.md+++ /dev/null@@ -1 +0,0 @@-- An image whose size cannot be read from its header is refused before decoding with an error that names the file and the frame limits, and a JPEG with a restart marker before its frame header can no longer slip past the size limit (#1319)diff --git a/changes/1319.size.md b/changes/1319.size.mddeleted file mode 100644index 60612625..00000000--- a/changes/1319.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:gltf-box: 30 bytes gzip for the JPEG header reader to step over restart, SOI and TEM markers, which closes a bypass of the texture size limit by a crafted JPEG (#1319).diff --git a/changes/1325.md b/changes/1325.mddeleted file mode 100644index 3083868f..00000000--- a/changes/1325.md+++ /dev/null@@ -1 +0,0 @@-- MCP errors name what to change: a review camera lists the fields it rejects, a stale revision gives the current one, a path traced job far over the 400000 triangle cap of the cpu tracer is refused at `scene_render_submit` with the count, and an unknown `scene_capabilities` topic suggests the nearest one (#1325)diff --git a/changes/1327-npm.md b/changes/1327-npm.mddeleted file mode 100644index f7617faf..00000000--- a/changes/1327-npm.md+++ /dev/null@@ -1 +0,0 @@-- The docs and the render jobs skill now say `npm install @codepawlhq/crucible-denoise-oidn` instead of "not on npm yet", and the guides document `jobs.maximumFps`, `jobs.maximumFrames`, `emissionFresnel` and the adaptive resolution `raiseBelow` option (#1327)diff --git a/changes/1327-publish.md b/changes/1327-publish.mddeleted file mode 100644index 9d820464..00000000--- a/changes/1327-publish.md+++ /dev/null@@ -1 +0,0 @@-- `@codepawlhq/crucible-denoise-oidn` is now published to npm beside Crucible, so OIDN installs with an npm install of Crucible (#1327)diff --git a/changes/1333.md b/changes/1333.mddeleted file mode 100644index eca33a43..00000000--- a/changes/1333.md+++ /dev/null@@ -1 +0,0 @@-- The website matches codepawl.com: light by default with a dark mode on the header button, a plain header bar, captions under the showcase pictures, 8 px controls and the pot's metal as the only orange (#1333)diff --git a/changes/1334.md b/changes/1334.mddeleted file mode 100644index 4742133b..00000000--- a/changes/1334.md+++ /dev/null@@ -1 +0,0 @@-- A closed room's walls, ceiling and props now get the sky they can see traced into their vertices, so a sealed room renders near black like the path tracer instead of as bright as a room with a window (#1334).diff --git a/changes/1334.size.md b/changes/1334.size.mddeleted file mode 100644index b69b5b87..00000000--- a/changes/1334.size.md+++ /dev/null@@ -1,4 +0,0 @@-- scene:agent-scene+lazy: the closed-room sky trace now covers every surface of the room, not only the floor, so a sealed room stops lighting up like a room with a window; 202 bytes gzip for the extra surface selection and the whole-room cut (#1334).-- entry:floor-sky-visibility: the wall test (level rays around the room) and the total ray budget keep a canopy or a colonnade from being traced as a closed room; 135 bytes more on top of the whole-room trace (#1334).-- scene:agent-scene: the same wall test and ray budget, 72 bytes gzip (#1334).-- goal:agent-scene: 12 bytes gzip because the closed-room helpers (worldTriangles, freeOrigin, hitsFrontFace) are now exported for the wall test, which lives in the lazy fitted-closed-room module so the core bundle carries no new logic (#1334).diff --git a/changes/1335.md b/changes/1335.mddeleted file mode 100644index c48fb526..00000000--- a/changes/1335.md+++ /dev/null@@ -1 +0,0 @@-- The wavelet denoiser leaves far less blotchy grain on flat walls at 256 samples: its luminance guide ignores a lone noisy pixel (#1335)diff --git a/changes/1342.md b/changes/1342.mddeleted file mode 100644index 795a5730..00000000--- a/changes/1342.md+++ /dev/null@@ -1 +0,0 @@-- A recipe path such as `environment.lightDirection` written as a motion track (`target`, `keys`) now fails with an error that names the `sun` track (`direction`, `color`, `intensity`) that moves the key light under any sky; the docs say so (#1342).diff --git a/changes/1342.size.md b/changes/1342.size.mddeleted file mode 100644index 806c2a20..00000000--- a/changes/1342.size.md+++ /dev/null@@ -1,6 +0,0 @@-- scene:agent-scene+lazy: a motion track written as a recipe path now fails with an error naming the sun track that does the job, about 165 bytes gzip; agents wrote the path form and got only a list of allowed fields.-- entry:agent-scene+lazy: a motion track written as a recipe path now fails with an error naming the sun track that does the job, about 165 bytes gzip; agents wrote the path form and got only a list of allowed fields.-- entry:agent-asset: a motion track written as a recipe path now fails with an error naming the sun track that does the job, about 165 bytes gzip; agents wrote the path form and got only a list of allowed fields.-- entry:agent-asset+lazy: a motion track written as a recipe path now fails with an error naming the sun track that does the job, about 165 bytes gzip; agents wrote the path form and got only a list of allowed fields.-- entry:game+lazy: a motion track written as a recipe path now fails with an error naming the sun track that does the job, about 165 bytes gzip; agents wrote the path form and got only a list of allowed fields.-- goal:agent-scene: the track error text lives next to the track validation every agent scene bundles, so it cannot move out of core.diff --git a/changes/1343.md b/changes/1343.mddeleted file mode 100644index 251ba050..00000000--- a/changes/1343.md+++ /dev/null@@ -1 +0,0 @@-- The flush-face search keeps its triangles in flat arrays instead of one object each, so a 200k triangle GLB peaks at about half the browser heap while a scene loads, and `scripts/measure-glb-heap.mjs` measures it (#1343)diff --git a/changes/1343.size.md b/changes/1343.size.mddeleted file mode 100644index fd7a2ff5..00000000--- a/changes/1343.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:coplanar-overlap: +283 bytes of minified code for the flat triangle table and its bisection, which cuts the plane search heap per triangle by about half (#1343)-- scene:agent-scene+lazy: the same +286 bytes, counted again through the lazy coplanar module (#1343)diff --git a/changes/1347.md b/changes/1347.mddeleted file mode 100644index 681f4d66..00000000--- a/changes/1347.md+++ /dev/null@@ -1 +0,0 @@-- `loadGltf` keeps a `KHR_materials_ior` above 2.5 (exporters write 1000) at 2.5 with a warning instead of refusing the file (#1347)diff --git a/changes/1347.size.md b/changes/1347.size.mddeleted file mode 100644index 596663db..00000000--- a/changes/1347.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: a valid glTF with an ior of 1000 stopped the whole page; the clamp and its warning cost about 113 bytes gzip.-- goal:gltf-box: the loader fix cannot move out of the core bundle, because the glTF material read runs there for every file.-- scene:gltf-box+lazy: the same 114 bytes counted with the lazy modules.diff --git a/changes/1350.md b/changes/1350.mddeleted file mode 100644index b3ddd0ea..00000000--- a/changes/1350.md+++ /dev/null@@ -1 +0,0 @@-- The WebGPU path tracer megakernel now compiles with FXC, so the bundled Chromium on Windows can run the GPU tests, and the render stays byte for byte the same on DXC (#1350).diff --git a/changes/1351.md b/changes/1351.mddeleted file mode 100644index fbe54a30..00000000--- a/changes/1351.md+++ /dev/null@@ -1 +0,0 @@-- A camera `clipStart` at or past `far` is refused with a message that says what to pass; the MCP preview's native export and both real-time tracers now report that they ignore `clipStart` (#1351)diff --git a/changes/1359.md b/changes/1359.mddeleted file mode 100644index b29f720d..00000000--- a/changes/1359.md+++ /dev/null@@ -1 +0,0 @@-- A recipe scene that leaves `shadow` out and is wider than the default 7 m view now gets a sun shadow view fitted to its bounds (three automatic cascades past 40 m), so a sealed room stops leaking sun; a recipe that sets `shadow` is unchanged (#1359)diff --git a/changes/1371.md b/changes/1371.mddeleted file mode 100644index 8ba85b2d..00000000--- a/changes/1371.md+++ /dev/null@@ -1 +0,0 @@-- Breaking: an HDRI (a file, `environment.hdri`, a bundled look) now faces as in three.js and Cycles, a quarter turn from before; add `Math.PI / 2` to `environmentRotation` to get an old scene's look back (#1371)diff --git a/changes/1374.md b/changes/1374.mddeleted file mode 100644index 8ab167f8..00000000--- a/changes/1374.md+++ /dev/null@@ -1 +0,0 @@-- Look change: the light an HDRI gives a surface is now an exact cosine convolution of the map, so shade next to a strong sun no longer comes out up to 50 percent off and matches three.js within 5 percent (#1374)diff --git a/changes/1374.size.md b/changes/1374.size.mddeleted file mode 100644index 78cd767b..00000000--- a/changes/1374.size.md+++ /dev/null@@ -1,16 +0,0 @@-- entry:core: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- entry:core+lazy: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- entry:ground-bounce: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- entry:floor-sky-visibility: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- entry:hdri: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- entry:water-reflection: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- entry:reflection-probes: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- scene:lit-box: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- scene:lit-box+lazy: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- scene:gltf-box: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- scene:gltf-box+lazy: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- scene:agent-scene: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- scene:agent-scene+lazy: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- goal:lit-box: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- goal:gltf-box: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)-- goal:agent-scene: the diffuse light comes from an exact 64 by 32 cosine-convolved irradiance grid (about 30 ms of CPU, 98 KB of GPU memory, no extra sampler) instead of a nine-term fit that was up to 50 percent off in the shade of a strong sun (#1374)diff --git a/changes/1375.md b/changes/1375.mddeleted file mode 100644index 15210b1b..00000000--- a/changes/1375.md+++ /dev/null@@ -1 +0,0 @@-- The site header gives Marketplace an icon, live-example code sits on a tinted surface, and the sliders in docs examples show their value with brand colours and focus states (#1375)diff --git a/changes/1385.md b/changes/1385.mddeleted file mode 100644index 9a27eca7..00000000--- a/changes/1385.md+++ /dev/null@@ -1 +0,0 @@-- Glass that transmits light now reflects the reflection probe placed on it, and a blended part keeps its specular at full strength when its opacity drops, so a glass wall and a faint blend glass show the room and the sky instead of fading with alpha (#1385)diff --git a/changes/1385.size.md b/changes/1385.size.mddeleted file mode 100644index 2e38c8f1..00000000--- a/changes/1385.size.md+++ /dev/null@@ -1,12 +0,0 @@-- entry:core: about 320 bytes gzip for the premultiplied blend output (a blended glass keeps its reflection at any opacity, a visible fix) and the shared probe lookup that glass in the liquid pass reads; the lit shader carries the specular split only through one global and four lines.-- entry:core+lazy: the same 320 bytes, the lit shader is in the core bundle.-- scene:lit-box: the lit shader is in every scene bundle, so the blend output fix adds about 320 bytes gzip.-- scene:lit-box+lazy: the same 320 bytes.-- scene:gltf-box: the same 335 bytes for the lit shader, and glass in a glTF is the case the fix is for.-- scene:agent-scene: the same 333 bytes for the lit shader.-- goal:lit-box: about 320 bytes gzip so a blended glass keeps its reflection when its opacity drops (a visible fix in the lit shader every scene carries).-- goal:gltf-box: about 335 bytes gzip for the same fix, which glass loaded from glTF needs.-- goal:agent-scene: about 333 bytes gzip for the same fix.-- entry:liquid-shader: about 157 bytes gzip for the shared probe lookup that glass drawn in the liquid pass reads, so glass there reflects the room like glass in the lit pass.-- scene:gltf-box+lazy: about 511 bytes gzip, the lit shader blend output (335) plus the lazy liquid shader's probe lookup; glass in a glTF is the case the fix is for.-- scene:agent-scene+lazy: about 789 bytes gzip, the lit shader blend output (333) plus the lazy liquid and probe code an agent scene loads; same visible glass fix.diff --git a/changes/1387.md b/changes/1387.mddeleted file mode 100644index 5408565b..00000000--- a/changes/1387.md+++ /dev/null@@ -1 +0,0 @@-- A rules input trigger on an action the game does not bind (anything but jump or run) is now refused by set-game with the actions to use, instead of crashing startGame on its first step (#1387)diff --git a/changes/1387.size.md b/changes/1387.size.mddeleted file mode 100644index 787cec56..00000000--- a/changes/1387.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 90 bytes gzip for the check that refuses a rules input trigger on an unbound action, which crashed startGame on its first step, and for finite arithmetic in game rules.diff --git a/changes/1388.md b/changes/1388.mddeleted file mode 100644index 4a78b73a..00000000--- a/changes/1388.md+++ /dev/null@@ -1 +0,0 @@-- Game rules arithmetic whose result overflows or is not a number now gives 0 instead of leaving a variable at Infinity or NaN (#1388)diff --git a/changes/1389.md b/changes/1389.mddeleted file mode 100644index 87a2be34..00000000--- a/changes/1389.md+++ /dev/null@@ -1 +0,0 @@-- Live physics fixedUpdate rejects a NaN step, and raycast, shapeCast and overlap give the query options error for null, both naming what to pass (#1389)diff --git a/changes/1393.md b/changes/1393.mddeleted file mode 100644index b5bd7690..00000000--- a/changes/1393.md+++ /dev/null@@ -1 +0,0 @@-- A film timeline frame that lands on a shot cut now shows the later shot even when shot lengths such as 0.1 s and 0.2 s sum with float error (#1393)diff --git a/changes/1396.md b/changes/1396.mddeleted file mode 100644index 5610451d..00000000--- a/changes/1396.md+++ /dev/null@@ -1 +0,0 @@-- A particle emitter from a scaled mesh now emits from vertices at the stated `normalSpeed` and orients instances with unit normals (#1396)diff --git a/changes/1396.size.md b/changes/1396.size.mddeleted file mode 100644index 51eb6f35..00000000--- a/changes/1396.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 12 bytes gzip for three correctness fixes (scaled-mesh emitter normals, pitch variance clamp, motionBlur.objects validation, #1396 to #1398); particle and effects code is bundled with the scene.diff --git a/changes/1397.md b/changes/1397.mddeleted file mode 100644index 4727da3f..00000000--- a/changes/1397.md+++ /dev/null@@ -1 +0,0 @@-- A sound with `pitchVariance` no longer throws a range error from `play` at low pitches (#1397)diff --git a/changes/1398.md b/changes/1398.mddeleted file mode 100644index 1e9f2955..00000000--- a/changes/1398.md+++ /dev/null@@ -1 +0,0 @@-- `motionBlur.objects` now rejects negative, NaN and text values instead of ignoring them (#1398)diff --git a/changes/1399.md b/changes/1399.mddeleted file mode 100644index 62402a33..00000000--- a/changes/1399.md+++ /dev/null@@ -1 +0,0 @@-- A curve through a repeated point, with a duplicated end control point or with repeated trailing points (poly and NURBS) now builds instead of throwing "Cannot read properties of null"; the direction there comes from the nearest distinct point (#1399)diff --git a/changes/1400.md b/changes/1400.mddeleted file mode 100644index 2c38ecac..00000000--- a/changes/1400.md+++ /dev/null@@ -1 +0,0 @@-- A glTF accessor or sparse block with a negative or fractional byteOffset now fails with an error naming byteOffset instead of reading the wrong bytes (#1400).diff --git a/changes/1400.size.md b/changes/1400.size.mddeleted file mode 100644index 8bd39a90..00000000--- a/changes/1400.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:gltf-box: 55 bytes gzip for a byteOffset check and a normalized guard in the accessor reader; they stop a hostile glTF from loading silently wrong bytes or NaN geometry, a functional fix (#1400, #1401).diff --git a/changes/1401.md b/changes/1401.mddeleted file mode 100644index 07402efc..00000000--- a/changes/1401.md+++ /dev/null@@ -1 +0,0 @@-- A glTF accessor that sets normalized on floats or 32-bit integers now keeps its values instead of decoding to NaN (#1401).diff --git a/changes/1405.md b/changes/1405.mddeleted file mode 100644index c964808d..00000000--- a/changes/1405.md+++ /dev/null@@ -1 +0,0 @@-- Library calls `bevelEdges`, `subdivideMesh`, `loopCut`, `bisectPlane`, `resampleCurve`, `resampleSamples` and `filletCurve` refuse NaN, missing and unbounded inputs with an error that names the parameter, the value and the allowed range, and share their limits with the recipe layer (#1405)diff --git a/changes/1405.size.md b/changes/1405.size.mddeleted file mode 100644index 6f3fe9e6..00000000--- a/changes/1405.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: one shared validator (modeling-limits.js) lets seven library calls (bevel, subdivide, loop cut, bisect, resample, fillet) throw an error naming the parameter and range instead of hanging or returning NaN points; the recipe layer already guards the same inputs (+575 bytes gzip)diff --git a/changes/1406.md b/changes/1406.mddeleted file mode 100644index e8d2d8e0..00000000--- a/changes/1406.md+++ /dev/null@@ -1 +0,0 @@-- Make `fitOrthographicHeight` throw for a view whose up is parallel to the view direction, whose eye equals its target or whose vectors are not finite, instead of returning 0 or a meaningless height (#1406)diff --git a/changes/1406.size.md b/changes/1406.size.mddeleted file mode 100644index 6fbb7ce5..00000000--- a/changes/1406.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:core: fitOrthographicHeight and createCameraSampler now validate their input, which turns a silent wrong height and TypeErrors on bad camera.json into errors that say what to pass.-- entry:core+lazy: same validation, shared by the lazy entry.diff --git a/changes/1407.md b/changes/1407.mddeleted file mode 100644index adf7fd0d..00000000--- a/changes/1407.md+++ /dev/null@@ -1 +0,0 @@-- Make `createCameraSampler` reject empty frames, a non-positive fps, malformed frames and a non-finite sample time with errors that say what to pass (#1407)diff --git a/changes/1409.md b/changes/1409.mddeleted file mode 100644index 14ece95a..00000000--- a/changes/1409.md+++ /dev/null@@ -1 +0,0 @@-- A glTF node with a `matrix`, `translation`, `rotation` or `scale` of the wrong shape, or a zero quaternion, is refused with the field named instead of reported as a singular node (#1409)diff --git a/changes/1409.size.md b/changes/1409.size.mddeleted file mode 100644index 83d00d8a..00000000--- a/changes/1409.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: names the bad field, base64 or primitive mode in loader errors an agent reads (372 bytes gzip for three clear refusals)-- scene:gltf-box+lazy: same code, loaded with the lazy decoder-- goal:gltf-box: the bytes are error text and validation of node transforms, data URIs and primitive modes, not new featuresdiff --git a/changes/1410.md b/changes/1410.mddeleted file mode 100644index baeca88e..00000000--- a/changes/1410.md+++ /dev/null@@ -1 +0,0 @@-- A glTF data URI accepts a capitalised `;BASE64` marker, and a bad payload says to use standard base64 instead of throwing a bare decoder message (#1410)diff --git a/changes/1411.md b/changes/1411.mddeleted file mode 100644index d9fa19e5..00000000--- a/changes/1411.md+++ /dev/null@@ -1 +0,0 @@-- A glTF TRIANGLE_STRIP, TRIANGLE_FAN or unknown primitive mode is refused with its name, mesh and primitive and the fix (#1411)diff --git a/changes/1416.md b/changes/1416.mddeleted file mode 100644index a8e43152..00000000--- a/changes/1416.md+++ /dev/null@@ -1 +0,0 @@-- Editing an imported mesh keeps the UVs of a face that has a real corner at UV 0,0 and rebuilds only faces with no UVs (#1416).diff --git a/changes/1417.md b/changes/1417.mddeleted file mode 100644index 7f66a0c9..00000000--- a/changes/1417.md+++ /dev/null@@ -1 +0,0 @@-- `scatterVegetation`, `createGrassClump` and `createFlowerPatch` check their options and throw a RangeError with a valid example instead of crashing, hanging or running out of memory (#1417).diff --git a/changes/1417.size.md b/changes/1417.size.mddeleted file mode 100644index 740aaee3..00000000--- a/changes/1417.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 301 B gzip of option checks that turn a hang, an out-of-memory death and an opaque TypeError into one-line errors.diff --git a/changes/1418.md b/changes/1418.mddeleted file mode 100644index 942b7574..00000000--- a/changes/1418.md+++ /dev/null@@ -1 +0,0 @@-- The MCP catalog decode refuses an external glTF buffer or image uri instead of fetching it, with an error that names the uri (#1418).diff --git a/changes/1418.size.md b/changes/1418.size.mddeleted file mode 100644index 05551585..00000000--- a/changes/1418.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:gltf-box: refusing external uris on the MCP decode path (#1418) costs 54 bytes gzip and keeps the SECURITY.md promise of no remote asset fetches.diff --git a/changes/1419.md b/changes/1419.mddeleted file mode 100644index 7942df7d..00000000--- a/changes/1419.md+++ /dev/null@@ -1 +0,0 @@-- Fix the material snapshot so a fast-path update with a changed `emissionFresnel` or `emissionMask` throws instead of being ignored (#1419).diff --git a/changes/1419.size.md b/changes/1419.size.mddeleted file mode 100644index 71270bb9..00000000--- a/changes/1419.size.md+++ /dev/null@@ -1,3 +0,0 @@-- goal:lit-box: 30 bytes gzip so a changed emission fresnel or mask is no longer dropped silently on the update path (#1419).-- goal:gltf-box: 24 bytes gzip for the same fix.-- goal:agent-scene: 33 bytes gzip for the same fix.diff --git a/changes/1420.md b/changes/1420.mddeleted file mode 100644index 2761e126..00000000--- a/changes/1420.md+++ /dev/null@@ -1 +0,0 @@-- Terrain ground normals on the outer border now use the real span between samples, `height` accepts plain and Float64 arrays and rejects other non-object values, and `update` rejects a non-finite eye position (#1420)diff --git a/changes/1420.size.md b/changes/1420.size.mddeleted file mode 100644index b193743a..00000000--- a/changes/1420.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 157 bytes gzip buy a working Float64Array and plain-array heightmap, a clear error for a bad height option and an eye-position check that stops NaN geometry.diff --git a/changes/1425.md b/changes/1425.mddeleted file mode 100644index 36e6e06b..00000000--- a/changes/1425.md+++ /dev/null@@ -1 +0,0 @@-- GLB export now rejects lightUvs of the wrong length and empty, non-finite or unsorted animation keyframes with errors that name the fix, and writes a Float32Array light map tint as a JSON array (#1425, #1426, #1428).diff --git a/changes/1425.size.md b/changes/1425.size.mddeleted file mode 100644index a2c16a39..00000000--- a/changes/1425.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:core: three guards that stop silently wrong GLB files (lightUvs length, animation keyframes) cost 50 bytes over the allowance.-- entry:core+lazy: same guards, same bytes.diff --git a/changes/1427.md b/changes/1427.mddeleted file mode 100644index d8ef0e38..00000000--- a/changes/1427.md+++ /dev/null@@ -1 +0,0 @@-- Navmesh and crowd bake refuse a reversed obstacle footprint, non-finite or out-of-range geometry (#1427, #1429) and a palette of more than 7200 frames (#1430), instead of building a silently wrong or oversized result.diff --git a/changes/1427.size.md b/changes/1427.size.mddeleted file mode 100644index 14d509e2..00000000--- a/changes/1427.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:navmesh: three input guards (reversed obstacle footprint, non-finite positions, indices past the positions) that stop a path through a wall and an empty grid with a misleading error; about 140 bytes gzip.-- scene:agent-scene+lazy: the same navmesh guards, about 230 bytes gzip.diff --git a/changes/1431.md b/changes/1431.mddeleted file mode 100644index 969c42b3..00000000--- a/changes/1431.md+++ /dev/null@@ -1 +0,0 @@-- Cloth `step` takes zero time as a no-op and rejects a negative, NaN or over one second `deltaSeconds`, `bakeCloth` rejects a bad `frameCount` or `framesPerSecond`, `sampleClothFrames` rejects a NaN time, and `wakeTrail` no longer returns NaN points for a zero lifetime or a missing speed (#1431)diff --git a/changes/1431.size.md b/changes/1431.size.mddeleted file mode 100644index 0c94c532..00000000--- a/changes/1431.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 153 bytes gzip of input guards that stop a cloth step with zero or runaway time, a bake with a bad frame count, a NaN timeline sample and NaN wake points from silently breaking a scene.diff --git a/changes/1436.md b/changes/1436.mddeleted file mode 100644index 18d3820b..00000000--- a/changes/1436.md+++ /dev/null@@ -1 +0,0 @@-- Pose solvers refuse a zero aim `forward` or two-bone `bendToward`, and a morph-only layer refuses a non-finite `reference`, instead of silently doing nothing or writing NaN weights (#1436, #1437, #1438)diff --git a/changes/1441.md b/changes/1441.mddeleted file mode 100644index 89623104..00000000--- a/changes/1441.md+++ /dev/null@@ -1 +0,0 @@-- GLB export no longer crashes on a skinned part without a matrix, folds an emission below 1 into emissiveFactor instead of an invalid emissiveStrength, and keeps the negative scale of an animated mirrored node (#1441, #1442, #1443).diff --git a/changes/1444.md b/changes/1444.mddeleted file mode 100644index def28114..00000000--- a/changes/1444.md+++ /dev/null@@ -1 +0,0 @@-- A scene query now hits a part after its world matrix moves in place, `raycast` throws on a NaN or non-number `maxDistance` instead of reporting a miss, and a rig with a zero-scale `rootMatrix` throws instead of filling the pose with NaN (#1444, #1445, #1446).diff --git a/changes/1450.md b/changes/1450.mddeleted file mode 100644index db52edb6..00000000--- a/changes/1450.md+++ /dev/null@@ -1 +0,0 @@-- Fix `trianglesIntersect` for a zero-area first triangle, `floatToHalf` rounding of subnormal halves, and make `RenderTargetLedger.record` throw on an unknown format or bad size instead of reporting NaN bytes (#1450, #1448, #1449).diff --git a/changes/1450.size.md b/changes/1450.size.mddeleted file mode 100644index a7e465ec..00000000--- a/changes/1450.size.md+++ /dev/null@@ -1,4 +0,0 @@-- goal:lit-box: about 50 bytes gzip for three guards that stop a wrong result (a false triangle hit, a half float rounded the wrong way, a NaN memory total).-- goal:gltf-box: same three guards, about 46 bytes gzip.-- goal:agent-scene: same three guards, about 52 bytes gzip.-- scene:agent-scene+lazy: same three guards, about 70 bytes gzip.diff --git a/changes/1453.md b/changes/1453.mddeleted file mode 100644index ac129766..00000000--- a/changes/1453.md+++ /dev/null@@ -1 +0,0 @@-- A light linked to a group node is now reported as an unknown name instead of lighting nothing (#1453)diff --git a/changes/1455.md b/changes/1455.mddeleted file mode 100644index fff012df..00000000--- a/changes/1455.md+++ /dev/null@@ -1 +0,0 @@-- Kernel offload no longer reuses another call's cached result for look-alike inputs, no longer hangs when a worker answer cannot be read, and runs a self-referencing input inline (#1455, #1456, #1457).diff --git a/changes/1463.md b/changes/1463.mddeleted file mode 100644index 99c72ad5..00000000--- a/changes/1463.md+++ /dev/null@@ -1 +0,0 @@-- Fix three silent faults in the fog and local volume code: a moved fog box leaks its old sky texture, the shadow ambient cache ignores the measure it is given, and packing accepts a grid volume that was never bound (#1463)diff --git a/changes/1463.size.md b/changes/1463.size.mddeleted file mode 100644index f9b569f6..00000000--- a/changes/1463.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 18 bytes gzip for three guards that stop a leaked GPU texture, a stale shadow ambient and a grid volume packed unbound (#1463 #1464 #1465).diff --git a/changes/1466.md b/changes/1466.mddeleted file mode 100644index f1e97ba7..00000000--- a/changes/1466.md+++ /dev/null@@ -1 +0,0 @@-- `scene_apply` fracture-node on a node that already has a physics body now names `node` and says to drop that body with set-physics first, instead of blaming `physics.bodies[n].node` after the node was replaced (#1466).diff --git a/changes/1467.md b/changes/1467.mddeleted file mode 100644index 5ef9ec77..00000000--- a/changes/1467.md+++ /dev/null@@ -1 +0,0 @@-- A path-traced scene with a negative, fractional or missing index, or a NaN in uvs or colors, and a real-time tracer with a bad `historyLimit`, `filterIterations`, `cacheCellSize` or `resolutionDivisor`, now stop with a message that names the part or option (#1467, #1461, #1462)diff --git a/changes/1468.md b/changes/1468.mddeleted file mode 100644index b84a0cac..00000000--- a/changes/1468.md+++ /dev/null@@ -1 +0,0 @@-- A pack JSON string that a browser reads as a remote address, such as a backslash form or a scheme with a tab inside, is now refused (#1468)diff --git a/changes/1469.md b/changes/1469.mddeleted file mode 100644index 1779ddc2..00000000--- a/changes/1469.md+++ /dev/null@@ -1 +0,0 @@-- An addon file in a folder whose name starts with two dots is no longer refused as outside its package (#1469)diff --git a/changes/1470.md b/changes/1470.mddeleted file mode 100644index 63fecf68..00000000--- a/changes/1470.md+++ /dev/null@@ -1 +0,0 @@-- The pack index build refuses an `--out` folder that overlaps `--packs` instead of deleting the packs (#1470)diff --git a/changes/1476.md b/changes/1476.mddeleted file mode 100644index 2eafcaef..00000000--- a/changes/1476.md+++ /dev/null@@ -1 +0,0 @@-- A `.cube` look with a bad DOMAIN line or a malformed row is refused, and a NaN or infinite channel gives a finite display colour instead of a NaN pixel (#1476, #1475).diff --git a/changes/1476.size.md b/changes/1476.size.mddeleted file mode 100644index 3ac56e4c..00000000--- a/changes/1476.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:view-transform: refuses malformed .cube DOMAIN lines, rows and sizes and keeps NaN or infinite channels from becoming NaN pixels; about 140 bytes gzip for input checks that agents hit with bad files.diff --git a/changes/1477.md b/changes/1477.mddeleted file mode 100644index 1cc02495..00000000--- a/changes/1477.md+++ /dev/null@@ -1 +0,0 @@-- A zero-size fog box no longer gets NaN sky grid cells, tree sky matching no longer writes NaN when leaf coverage is zero, and an empty crown says what to pass (#1477, #1478, #1479).diff --git a/changes/1483.md b/changes/1483.mddeleted file mode 100644index 6e4a60dd..00000000--- a/changes/1483.md+++ /dev/null@@ -1 +0,0 @@-- Closed-mesh check for `softBody.pressure` now rejects a triangle wound the wrong way, and liquid optics keep the default for an undefined field and reject an unknown field (#1483)diff --git a/changes/1483.size.md b/changes/1483.size.mddeleted file mode 100644index 09863d0b..00000000--- a/changes/1483.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: two input guards with error messages that tell an agent what to pass, about 100 bytes gzip.-- goal:agent-scene: the same two guards; a wrongly wound closed mesh silently breaks softBody pressure and a mistyped liquid field is ignored.diff --git a/changes/1489.md b/changes/1489.mddeleted file mode 100644index 14409b26..00000000--- a/changes/1489.md+++ /dev/null@@ -1 +0,0 @@-- Compressed textures with a zero, negative or fractional base size or a mip chain whose levels do not halve, and texture sequences given a NaN or infinite time, now throw an error that says what to pass instead of failing later (#1489, #1490).diff --git a/changes/1489.size.md b/changes/1489.size.mddeleted file mode 100644index ac90ec10..00000000--- a/changes/1489.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: 133 bytes gzip of input guards that stop a compressed KTX2 texture with a zero or fractional size, or a mip chain whose levels do not halve, from failing in the GPU upload.-- scene:gltf-box+lazy: the same 133 bytes gzip.-- goal:gltf-box: the same 133 bytes gzip; a malformed compressed texture now throws a named error at load instead of a GL error at upload.diff --git a/changes/1492.md b/changes/1492.mddeleted file mode 100644index b6766ffb..00000000--- a/changes/1492.md+++ /dev/null@@ -1 +0,0 @@-- Name the track and keyframe in motion validation errors for a non-array value, repeated times and an empty or malformed keyframe list (#1492, #1493, #1494).diff --git a/changes/1492.size.md b/changes/1492.size.mddeleted file mode 100644index a2e9f6cd..00000000--- a/changes/1492.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: 82 bytes gzip so that a rejected keyframe names its track and keyframe, which agents read to fix the entry.-- goal:agent-scene: the scene is already over its goal; the 82 bytes name the track and keyframe in three motion errors (#1492, #1493, #1494).diff --git a/changes/1495.md b/changes/1495.mddeleted file mode 100644index 6100dbb0..00000000--- a/changes/1495.md+++ /dev/null@@ -1 +0,0 @@-- Object reflections are traced again when only `environment.sky` changes (#1495)diff --git a/changes/1496.md b/changes/1496.mddeleted file mode 100644index f1acd95c..00000000--- a/changes/1496.md+++ /dev/null@@ -1 +0,0 @@-- A `meshEdits` entry that is not an object is reported as that entry instead of a character index (#1496)diff --git a/changes/1497.md b/changes/1497.mddeleted file mode 100644index c3649870..00000000--- a/changes/1497.md+++ /dev/null@@ -1 +0,0 @@-- `lintMesh` with `geometry: true` no longer reports a false self-intersection on a prism over a concave outline, and the `occlusion` measure of `vertex_group` skips zero-area faces instead of throwing (#1497, #1498)diff --git a/changes/1497.size.md b/changes/1497.size.mddeleted file mode 100644index 5d239ba8..00000000--- a/changes/1497.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the occlusion and lint fixes add a degenerate-triangle filter and a shared triangulation, 169 bytes gzip, so valid meshes no longer fail or get false warningsdiff --git a/changes/1502.md b/changes/1502.mddeleted file mode 100644index 4a4d670b..00000000--- a/changes/1502.md+++ /dev/null@@ -1 +0,0 @@-- A drive wheel that gives both `node` and `nodes` is rejected with a field error instead of silently ignoring `node` (#1502)diff --git a/changes/1503.md b/changes/1503.mddeleted file mode 100644index bbf4c0d0..00000000--- a/changes/1503.md+++ /dev/null@@ -1 +0,0 @@-- `exportGlb` throws on texture sampler wrap or filter values that are not glTF enums and on a `boxMapAxis` outside 0, 1 and 2, and `exportObj` throws when `geometry.uvs` or `geometry.normals` does not match the vertex count, instead of writing an invalid file or dropping the data (#1503, #1505)diff --git a/changes/1503.size.md b/changes/1503.size.mddeleted file mode 100644index 6308a2a1..00000000--- a/changes/1503.size.md+++ /dev/null@@ -1,3 +0,0 @@-- entry:core: GLB export checks sampler enums and boxMapAxis through one shared requireOneOf (+150 B gzip, down from +193)-- entry:core+lazy: same code as entry:core (+150 B)-- entry:export-mesh: the OBJ uv and normal length check through the same helper (+97 B, down from +110)diff --git a/changes/1504.md b/changes/1504.mddeleted file mode 100644index 2e6aa08f..00000000--- a/changes/1504.md+++ /dev/null@@ -1 +0,0 @@-- A part that the flush pass ranks and that carries baked `lightUvs`, `skyVisibility` or `surfaceMasks` no longer reaches the path tracer with NaN lightmap coordinates (the demo room window-top had 507 of them), and the path tracer now refuses non-finite `lightUvs` (#1504)diff --git a/changes/1511.md b/changes/1511.mddeleted file mode 100644index b8e5b232..00000000--- a/changes/1511.md+++ /dev/null@@ -1 +0,0 @@-- A mesh particle emitter no longer emits from a triangle or edge outside its vertex group when the random draw is exactly 0 (#1511)diff --git a/changes/1512.md b/changes/1512.mddeleted file mode 100644index 2116e6a0..00000000--- a/changes/1512.md+++ /dev/null@@ -1 +0,0 @@-- `maximumInstances` and `maximumVoices` must be whole numbers; a fractional limit used to let one more voice play (#1512)diff --git a/changes/1513.md b/changes/1513.mddeleted file mode 100644index 7bb10889..00000000--- a/changes/1513.md+++ /dev/null@@ -1 +0,0 @@-- `markersCrossed` and `emitMarkers` throw for a NaN or missing clip time instead of firing no cues (#1513)diff --git a/changes/1514.md b/changes/1514.mddeleted file mode 100644index 11776153..00000000--- a/changes/1514.md+++ /dev/null@@ -1 +0,0 @@-- A looping clip or a repeating cycles modifier sampled on a cycle start at a fractional frame rate, such as frame 9 of 30 fps on a 0.1 s clip, now shows the first pose instead of the last (#1514, #1515)diff --git a/changes/1514.size.md b/changes/1514.size.mddeleted file mode 100644index 5bde11b4..00000000--- a/changes/1514.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 51 bytes gzip for one shared cycle-split helper that fixes a wrong first pose on loop starts at fractional fps (#1514, #1515).diff --git a/changes/1517.md b/changes/1517.mddeleted file mode 100644index d3cd8e7a..00000000--- a/changes/1517.md+++ /dev/null@@ -1 +0,0 @@-- An animated vehicle player now defaults to the speed drive and refuses an intentSpeed drive at set-game, instead of crashing on the first step (#1517); a physics listener that disposes the system no longer lets later listeners run on the freed world (#1518).diff --git a/changes/1517.size.md b/changes/1517.size.mddeleted file mode 100644index 2faed033..00000000--- a/changes/1517.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: stops an animated vehicle player from crashing on its first step, a game-breaking bug; 72 bytes gzip for a controller-aware default drive and one refusal message.-- goal:agent-scene: the same 72 bytes; the bundle was already over its goal and the fix is a crash fix, not a feature.diff --git a/changes/1519.md b/changes/1519.mddeleted file mode 100644index 653a5c54..00000000--- a/changes/1519.md+++ /dev/null@@ -1 +0,0 @@-- `createTerrain` rejects a non-positive or non-finite `lodDistance`, and `scatterVegetation` rejects a bad `scale` maximum or an infinite `mix` weight, instead of silently building empty terrain or NaN instances (#1519, #1520, #1521).diff --git a/changes/1519.size.md b/changes/1519.size.mddeleted file mode 100644index 7098ebec..00000000--- a/changes/1519.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: 91 B gzip of three option checks that turn silently empty terrain and NaN instances into one-line errors.-- goal:agent-scene: the checks are agent-facing input guards, the bundle is already past its goal and the guards share the existing helper.diff --git a/changes/1524.md b/changes/1524.mddeleted file mode 100644index 3822f89e..00000000--- a/changes/1524.md+++ /dev/null@@ -1 +0,0 @@-- `crucible-addon-check` now fails a `package.json` that is not a JSON object, refuses a repeated `--version`, and fails an entry that is a folder (#1524, #1525, #1526).diff --git a/changes/1530.md b/changes/1530.mddeleted file mode 100644index ceb6e759..00000000--- a/changes/1530.md+++ /dev/null@@ -1 +0,0 @@-- A shadow casting light with a range of 0.05 or less no longer writes a NaN or flipped shadow matrix, and cascaded shadows keep their splits ordered when `maximumDistance` is nearer than the camera near plane (#1530, #1531).diff --git a/changes/1532.md b/changes/1532.mddeleted file mode 100644index 38ce8993..00000000--- a/changes/1532.md+++ /dev/null@@ -1 +0,0 @@-- Calling `createNetcode(...).init` twice without `dispose` no longer leaves the first transport subscription live, so messages are not handled twice (#1532).diff --git a/changes/1533.md b/changes/1533.mddeleted file mode 100644index c319f0a2..00000000--- a/changes/1533.md+++ /dev/null@@ -1 +0,0 @@-- A `game.restart()` whose physics rebuild fails now sets `game.physicsError` and makes the physics step throw the cause, instead of leaving a silent dead world (#1533).diff --git a/changes/1536.md b/changes/1536.mddeleted file mode 100644index 9ad83392..00000000--- a/changes/1536.md+++ /dev/null@@ -1 +0,0 @@-- A camera with an `fStop` now focuses the raster depth of field at `camera.focusDistance` when it is set (the path tracer and Blender already did), and a non-positive `fStop` throws (#1536)diff --git a/changes/1538.md b/changes/1538.mddeleted file mode 100644index f0cb249a..00000000--- a/changes/1538.md+++ /dev/null@@ -1 +0,0 @@-- The real-time path tracer refuses a `width` or `height` that is not a whole number of pixels, and names `width` and `resolutionDivisor` when the traced size leaves 16 to 4096 (#1538)diff --git a/changes/1539.md b/changes/1539.mddeleted file mode 100644index 8f757b97..00000000--- a/changes/1539.md+++ /dev/null@@ -1 +0,0 @@-- `edgeAngle` reads 0 next to a zero-area face, mesh volumes measure the positions and indices a part holds now, and empty or NaN positions get a clear error (#1539, #1540, #1541)diff --git a/changes/1539.size.md b/changes/1539.size.mddeleted file mode 100644index 3f910c10..00000000--- a/changes/1539.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 11 bytes over the allowance buy a mesh volume that measures the positions it holds now and a clear error for empty or NaN positionsdiff --git a/changes/1545.md b/changes/1545.mddeleted file mode 100644index e9974a4d..00000000--- a/changes/1545.md+++ /dev/null@@ -1 +0,0 @@-- A cached modifier stack result is checked against the current vertex and face limits, and realized instance geometry keeps one geometry record per distinct content, so the same scene builds to the same recipe (#1545, #1546).diff --git a/changes/1545.size.md b/changes/1545.size.mddeleted file mode 100644index 60406530..00000000--- a/changes/1545.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the same realized instance spec now gets the same geometry record on every build instead of a new one each time, and a cached modifier stack result is checked against the current limits; 124 bytes gzip for two silently wrong results.diff --git a/changes/1547.md b/changes/1547.mddeleted file mode 100644index bc9ef296..00000000--- a/changes/1547.md+++ /dev/null@@ -1 +0,0 @@-- `knife_project` no longer fails on a line along existing edges, `grid_fill` handles long loops within a vertex budget, and `fill_holes` stops reporting an Infinity limit (#1547, #1548, #1549)diff --git a/changes/1547.size.md b/changes/1547.size.mddeleted file mode 100644index 03f4b45b..00000000--- a/changes/1547.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 80 bytes gzip for the knife edge guard, the grid_fill vertex budget and a clearer fill_holes error; without them a long loop overflows the stack and a knife line along an edge fails with an error that blames no fielddiff --git a/changes/1552.md b/changes/1552.mddeleted file mode 100644index 823e3b9e..00000000--- a/changes/1552.md+++ /dev/null@@ -1 +0,0 @@-- `meshFromGeometry` refuses uvs or colors of the wrong length, a face tag with values above 1 now combines correctly in `selectFaces`, and a sculpt edit with non-array strokes names the field (#1552, #1553, #1554)diff --git a/changes/1552.size.md b/changes/1552.size.mddeleted file mode 100644index 06ce8da6..00000000--- a/changes/1552.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: three small guards (137 bytes gzip) turn a silently dropped look, a wrong selection count and a raw JavaScript error into fixable messages.diff --git a/changes/1558.md b/changes/1558.mddeleted file mode 100644index 54bc7471..00000000--- a/changes/1558.md+++ /dev/null@@ -1 +0,0 @@-- The splat PLY loader drops a splat with any non-finite field (a scale that overflows included, which used to throw) and reports `skipped: { count, fields }`, throwing only when no splat is usable; createGaussianSplats refuses a covariance that overflows 32-bit floats, naming the splat (#1558)diff --git a/changes/1563.md b/changes/1563.mddeleted file mode 100644index 9431b13d..00000000--- a/changes/1563.md+++ /dev/null@@ -1 +0,0 @@-- A glTF sparse accessor with a fractional, negative, missing or too large count, or without indices or values, now fails with an error naming sparse.count, sparse.indices or sparse.values instead of a typed array error or a silent skip (#1563).diff --git a/changes/1563.size.md b/changes/1563.size.mddeleted file mode 100644index ad286822..00000000--- a/changes/1563.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: one guard of about 70 bytes gzip stops a fractional, negative or missing sparse count and missing sparse blocks from loading silently or failing with a typed array error.-- scene:gltf-box+lazy: the same guard, in the shared accessor reader.-- goal:gltf-box: the guard turns a silently wrong or misleading read into a named error and removes no feature; the bundle was already over its goal.diff --git a/changes/1566.md b/changes/1566.mddeleted file mode 100644index f04b2e4f..00000000--- a/changes/1566.md+++ /dev/null@@ -1 +0,0 @@-- Light profiles with a near zero axis value, IES headers with a fractional angle count and GPU bakes of spot cones narrower than cosine precision no longer give Infinity, a wrong error or a NaN edge (#1566, #1567, #1568).diff --git a/changes/1569.md b/changes/1569.mddeleted file mode 100644index fb77079c..00000000--- a/changes/1569.md+++ /dev/null@@ -1 +0,0 @@-- An animation graph state with `loop: false` over a looping clip now holds the clip's last frame instead of snapping back to its first frame (#1569)diff --git a/changes/1573.md b/changes/1573.mddeleted file mode 100644index 01fe5f75..00000000--- a/changes/1573.md+++ /dev/null@@ -1 +0,0 @@-- glTF import normalizes the direction of a scaled spot or directional light, skips one scaled to a zero direction with a warning that names the node, and ignores a non-numeric light range with a warning (#1573)diff --git a/changes/1573.size.md b/changes/1573.size.mddeleted file mode 100644index 065da133..00000000--- a/changes/1573.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: stops a scaled light from shining along a stretched direction and a zero-scale light from failing later with an environment error.-- scene:gltf-box+lazy: same bytes as scene:gltf-box.-- goal:gltf-box: the 142 bytes gzip are the unit-vector helper and two warnings that stop a silently wrong light direction and a NaN range.diff --git a/changes/1574.md b/changes/1574.mddeleted file mode 100644index 3921e30b..00000000--- a/changes/1574.md+++ /dev/null@@ -1 +0,0 @@-- createCameraSampler rejects a frame whose eye equals its target and an fps that makes the duration infinite (#1574)diff --git a/changes/1575.md b/changes/1575.mddeleted file mode 100644index 3c04cda2..00000000--- a/changes/1575.md+++ /dev/null@@ -1 +0,0 @@-- A kernel worker request with an unregistered name (including `toString` or `constructor`) now answers an unknown kernel error, `encodeVolumeGrid` refuses channel names and sizes that `decodeVolumeGrid` rejects, and a grid header that is not a JSON object says so (#1575, #1576, #1577).diff --git a/changes/1579.md b/changes/1579.mddeleted file mode 100644index 5d415e25..00000000--- a/changes/1579.md+++ /dev/null@@ -1 +0,0 @@-- Vegetation scatter errors and clearings are exact: an area off the ground names `vegetation.scatter.area`, a node exclusion follows the node's world position, and vertex group errors in particles, cloth and exclusions name their list index (#1579, #1580, #1581)diff --git a/changes/1579.size.md b/changes/1579.size.mddeleted file mode 100644index 47d1cf15..00000000--- a/changes/1579.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 142 bytes gzip buy a named error for an off-ground scatter area, a clearing that follows a node's world position (a wrong scene before) and list indexes in four vertex group and exclusion errors.diff --git a/changes/1585.md b/changes/1585.mddeleted file mode 100644index 09a0ad3a..00000000--- a/changes/1585.md+++ /dev/null@@ -1 +0,0 @@-- A motion track on the environment kind with keyIntensity, keyColor or keyDirection now says the key light is the sun kind and names the sun track to use (#1585)diff --git a/changes/1585.size.md b/changes/1585.size.mddeleted file mode 100644index 258487ab..00000000--- a/changes/1585.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the error hint sends an agent that wrote an environment key light track to the sun kind in one call instead of a failed retry loop (112 bytes gzip).-- goal:agent-scene: the same hint, 16 bytes gzip, saves agents a round trip on a common mistake.diff --git a/changes/1586.md b/changes/1586.mddeleted file mode 100644index 4bc2b370..00000000--- a/changes/1586.md+++ /dev/null@@ -1 +0,0 @@-- Refuse mismatched OIDN buffers, release the OrbitControls listeners on dispose and refuse short or out-of-range geometry in toThreeObject (#1586, #1587, #1588).diff --git a/changes/1589.md b/changes/1589.mddeleted file mode 100644index 93a2d8fd..00000000--- a/changes/1589.md+++ /dev/null@@ -1 +0,0 @@-- The Blender add-on keeps a Roughness, Metallic or Alpha constant when its image has no pixels instead of exporting an empty texture, and notes an anamorphic aperture ratio it cannot render (#1589, #1591).diff --git a/changes/1592.md b/changes/1592.mddeleted file mode 100644index 6d0605c9..00000000--- a/changes/1592.md+++ /dev/null@@ -1 +0,0 @@-- Terrain hole errors and `scene_inspect` path query errors now name the exact field (`terrain.holes[1].radius`, `path.from`, `path.avoid[1]`) instead of a bare message or a generic `id` (#1592)diff --git a/changes/1592.size.md b/changes/1592.size.mddeleted file mode 100644index ed3a8fc0..00000000--- a/changes/1592.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: about 180 bytes gzip of field-path checks for terrain holes and the path query, so an agent reads which field to fix instead of a bare navmesh message-- goal:agent-scene: the same 180 bytes; the scene is already over its goal and these checks turn two unnamed errors into named onesdiff --git a/changes/1595.md b/changes/1595.mddeleted file mode 100644index ca73ffea..00000000--- a/changes/1595.md+++ /dev/null@@ -1 +0,0 @@-- glTF import clamps negative skin weights to 0 and renormalises each vertex (an all-negative vertex binds to joint 0), with one warning per primitive that counts them, instead of refusing the file (#1595)diff --git a/changes/1595.size.md b/changes/1595.size.mddeleted file mode 100644index e9233bba..00000000--- a/changes/1595.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: files with negative skin weights now load (as in three.js and Blender) instead of failing, 204 bytes gzip-- scene:gltf-box+lazy: same clamp-and-renormalise repair in the shared glTF decoder, 204 bytes gzip-- goal:gltf-box: negative skin weights now load instead of failing the file, as in three.js and Blender; 204 bytes gzip is the clamp and the warningdiff --git a/changes/1596.md b/changes/1596.mddeleted file mode 100644index c3878419..00000000--- a/changes/1596.md+++ /dev/null@@ -1 +0,0 @@-- Grass and clutter fields refuse an area that needs over 262144 tiles, `filletCurve` rejects an arc segment count of 0, a fraction or over 64, and field `div` and `mod` give 0 for a divisor so small the result overflows (#1596, #1597, #1598)diff --git a/changes/1599.md b/changes/1599.mddeleted file mode 100644index d973c01d..00000000--- a/changes/1599.md+++ /dev/null@@ -1 +0,0 @@-- three-compat sums several DirectionalLights, honours every rotation.order and rejects a torus radius of 0; three-bridge clamps an InstancedMesh count to its matrix array with a warning (#1599)diff --git a/changes/1603.md b/changes/1603.mddeleted file mode 100644index e659b545..00000000--- a/changes/1603.md+++ /dev/null@@ -1 +0,0 @@-- A vegetation exclusion that mixes min/max with center, radius or node, a hair `swayRate` without `sway`, and a `radius` on a box particle emitter are now refused with an error that names the fix instead of being ignored (#1603, #1604, #1605).diff --git a/changes/1603.size.md b/changes/1603.size.mddeleted file mode 100644index 871407d7..00000000--- a/changes/1603.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: three recipe checks that refuse conflicting forms (box emitter radius, swayRate without sway, box and circle exclusion) an agent would otherwise get silently ignoreddiff --git a/changes/1608.md b/changes/1608.mddeleted file mode 100644index a2142318..00000000--- a/changes/1608.md+++ /dev/null@@ -1 +0,0 @@-- The Blender add-on exports the camera's own clip end, says so when every mesh has no triangles, expands `//` paths for the tracer folder, and refuses a NaN or infinite environment cap (#1608, #1609, #1610, #1611).diff --git a/changes/1617.md b/changes/1617.mddeleted file mode 100644index 35d294f2..00000000--- a/changes/1617.md+++ /dev/null@@ -1 +0,0 @@-- `triangulate`, `createStreaming`, `createInput` and the crowd `setTimes` refuse bad input with a named error, `crowdFrameBlend` shows the first frame for a non-finite time, and input `dispose` finishes every source before rethrowing (#1617, #1618, #1619, #1620)diff --git a/changes/1617.size.md b/changes/1617.size.mddeleted file mode 100644index fbc05477..00000000--- a/changes/1617.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: 129 bytes over the allowance buy named errors for an empty contour, NaN points, a fractional load count and a NaN crowd clip time at setTimes, and a dispose that reaches every input source-- goal:agent-scene: the same 193 bytes are guards that stop NaN offsets, a fractional load count and a leaked input source; the scene was already over its goaldiff --git a/changes/1622.md b/changes/1622.mddeleted file mode 100644index 64fd4812..00000000--- a/changes/1622.md+++ /dev/null@@ -1 +0,0 @@-- Physics collider limit errors name the node by its list index, `nodes[1].geometry`, instead of putting the node ID in the brackets (#1622)diff --git a/changes/1623.md b/changes/1623.mddeleted file mode 100644index c80f2151..00000000--- a/changes/1623.md+++ /dev/null@@ -1 +0,0 @@-- A crowd `farDistance` is measured in the space of the crowd instances, so a crowd under a moved parent switches to the far rate by its real distance from the camera (#1623)diff --git a/changes/1623.size.md b/changes/1623.size.mddeleted file mode 100644index 8d0c9945..00000000--- a/changes/1623.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 109 bytes over the allowance buy a crowd far distance measured in the space of its instances, so a crowd under a moved parent switches to the far rate by its real distance from the cameradiff --git a/changes/1625.md b/changes/1625.mddeleted file mode 100644index 3f80e3b7..00000000--- a/changes/1625.md+++ /dev/null@@ -1 +0,0 @@-- The live runtime retries a system whose init threw, keeps time finite after one NaN clock reading, and an arc shape refuses a span wider than half a turn (#1625, #1626, #1627).diff --git a/changes/1628.md b/changes/1628.mddeleted file mode 100644index 25e01d37..00000000--- a/changes/1628.md+++ /dev/null@@ -1 +0,0 @@-- Look up a part's matrix by the part before its name in the closed-room, sky openness, ground bounce and sun bounce tests, so glTF parts that share a name keep their own matrix (#1628)diff --git a/changes/1628.size.md b/changes/1628.size.mddeleted file mode 100644index b950d630..00000000--- a/changes/1628.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: the matrix lookup by part (about 30 bytes gzip) fixes a silently wrong closed-room test for glTF parts that share a name.diff --git a/changes/1629.md b/changes/1629.mddeleted file mode 100644index eae55693..00000000--- a/changes/1629.md+++ /dev/null@@ -1 +0,0 @@-- Crowd `farDistance` follows a parent group that has a motion track: the camera distance is measured from where the parent is at that time, not where it was when the scene was built (#1629)diff --git a/changes/1631.md b/changes/1631.mddeleted file mode 100644index b12c0a2d..00000000--- a/changes/1631.md+++ /dev/null@@ -1 +0,0 @@-- Mesh normals: weighted_normal falls back to the face normal, then +Z, for a zero-area face instead of writing a zero normal (#1631), and cornerNormals refuses a NaN, negative or oversized sharpAngle (#1632).diff --git a/changes/1631.size.md b/changes/1631.size.mddeleted file mode 100644index 020970c5..00000000--- a/changes/1631.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 145 bytes gzip stop a NaN or degrees-as-radians sharpAngle from silently shading a hard-edged mesh smooth, and a zero-area face from rendering with a zero normal.diff --git a/changes/1633.md b/changes/1633.mddeleted file mode 100644index 0f6c4943..00000000--- a/changes/1633.md+++ /dev/null@@ -1 +0,0 @@-- A ragdoll rig with a later parent, a repeated bone id or a bone that is its own parent is refused when built, camera rigs ignore a time step that is not above zero, and `buildSprayGeometry` refuses a bad `maxPoints` or `radius` (#1633, #1634, #1636)diff --git a/changes/1633.size.md b/changes/1633.size.mddeleted file mode 100644index 8cd730ee..00000000--- a/changes/1633.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 128 bytes gzip buy three guards that stop a ragdoll rig crashing with no bone named, a camera rig turning NaN for good, and spray built from a NaN or negative radiusdiff --git a/changes/1638.md b/changes/1638.mddeleted file mode 100644index 33e473ff..00000000--- a/changes/1638.md+++ /dev/null@@ -1 +0,0 @@-- `projectUv`, `packUvIslands`, `unwrapUv` and `bisectPlane` refuse an unknown mode or axis, a margin outside 0 to 0.2 and a bad distance with an error that names the field (#1638, #1640, #1641).diff --git a/changes/1638.size.md b/changes/1638.size.mddeleted file mode 100644index 444f389d..00000000--- a/changes/1638.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: three argument guards (UV margin, projection mode and axis, bisect distance) turn silent wrong UVs and a misleading error into messages that name the field, 144 bytes gzip.diff --git a/changes/1642.md b/changes/1642.mddeleted file mode 100644index 27c3fa6e..00000000--- a/changes/1642.md+++ /dev/null@@ -1 +0,0 @@-- Fix three edge cases from a bug hunt: the render target ledger refuses negative sizes, assembly layouts ignore inherited object keys, and a clutter mask returning NaN places nothing (#1642, #1643, #1644).diff --git a/changes/1642.size.md b/changes/1642.size.mddeleted file mode 100644index c16e6696..00000000--- a/changes/1642.size.md+++ /dev/null@@ -1,3 +0,0 @@-- goal:lit-box: 9 bytes gzip for the render target ledger refusing negative sizes, which recorded -4x-4 as a real 64 byte target.-- goal:gltf-box: same 9 bytes as lit-box, the ledger check in the shared renderer.-- goal:agent-scene: same 9 bytes as lit-box, the ledger check in the shared renderer.diff --git a/changes/1645.md b/changes/1645.mddeleted file mode 100644index d436c850..00000000--- a/changes/1645.md+++ /dev/null@@ -1 +0,0 @@-- A sealed room whose glTF parts share one name is now recognised as a closed room by the floor sky trace, and object motion blur gives each part of a shared name its own motion (#1645, #1646)diff --git a/changes/1650.md b/changes/1650.mddeleted file mode 100644index ae92e0fa..00000000--- a/changes/1650.md+++ /dev/null@@ -1 +0,0 @@-- Radiance .hdr files in the original format, where a 1,1,1 pixel repeats the previous pixel, now load instead of reading wrong or failing as truncated (#1650)diff --git a/changes/1652-clutter-seed.md b/changes/1652-clutter-seed.mddeleted file mode 100644index 0eef9f59..00000000--- a/changes/1652-clutter-seed.md+++ /dev/null@@ -1 +0,0 @@-- `createClutter` refuses a seed that is not a finite number, as grass and terrain already do, instead of scattering one fixed layout or joining a string seed into its hashes (#1652).diff --git a/changes/1652.md b/changes/1652.mddeleted file mode 100644index 18d39bad..00000000--- a/changes/1652.md+++ /dev/null@@ -1 +0,0 @@-- Refuse a NaN, Infinity or non-number seed in `createGrass` and `createTerrain` (noise height), and a non-finite budget in `createSlicer`, with errors that name the field; fractional seeds still truncate (#1652)diff --git a/changes/1652.size.md b/changes/1652.size.mddeleted file mode 100644index c7849f5c..00000000--- a/changes/1652.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the seed and slice-budget refusals stop a NaN or Infinity from silently giving the seed-0 layout or freezing the page, through one shared helper.-- goal:gltf-box: the createSlicer budget check is 28 bytes and stops Infinity from never yielding to the page.diff --git a/changes/1653.md b/changes/1653.mddeleted file mode 100644index 2accc79c..00000000--- a/changes/1653.md+++ /dev/null@@ -1 +0,0 @@-- Fix unsubdivide losing every face tag, and boolean cutterFaces counting faces of earlier booleans when a cutTag is reused (#1653)diff --git a/changes/1653.size.md b/changes/1653.size.mddeleted file mode 100644index 5a8af7c3..00000000--- a/changes/1653.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: unsubdivide now carries tags and face attributes, and the boolean counts only its own cutter faces; 104 bytes gzip for two silently wrong results.diff --git a/changes/1655.md b/changes/1655.mddeleted file mode 100644index c5f48e8e..00000000--- a/changes/1655.md+++ /dev/null@@ -1 +0,0 @@-- Spume whitewater skips points with NaN or infinite values and `partsForFrame` clamps an out-of-range frame instead of throwing (#1655).diff --git a/changes/1656.md b/changes/1656.mddeleted file mode 100644index 821226a0..00000000--- a/changes/1656.md+++ /dev/null@@ -1 +0,0 @@-- Hair roots stay finite when a surface vertex has a NaN or infinite position or normal (#1656).diff --git a/changes/1660.md b/changes/1660.mddeleted file mode 100644index e6c2b711..00000000--- a/changes/1660.md+++ /dev/null@@ -1 +0,0 @@-- Game setup no longer hangs on scene nodes whose parents form a cycle, the NPC path follower system skips the loop's zero-length first frame, and `createLookAt` and `createFootPlacement` ignore a non-finite pose time instead of staying NaN (#1660, #1661, #1662)diff --git a/changes/1664.md b/changes/1664.mddeleted file mode 100644index a1755314..00000000--- a/changes/1664.md+++ /dev/null@@ -1 +0,0 @@-- `uploadTextureInSlices` refuses short RGBA data and a zero size before any GL call, and `effectsWithFStop` accepts a camera without a target like the renderer does (#1664, #1663)diff --git a/changes/1664.size.md b/changes/1664.size.mddeleted file mode 100644index de7fa4d4..00000000--- a/changes/1664.size.md+++ /dev/null@@ -1,4 +0,0 @@-- entry:core+lazy: two guards that stop a GL error with no JS error and a camera TypeError on a frame the renderer draws-- scene:lit-box+lazy: same two guards-- scene:gltf-box+lazy: same two guards-- scene:agent-scene+lazy: same two guardsdiff --git a/changes/1666.md b/changes/1666.mddeleted file mode 100644index 782af275..00000000--- a/changes/1666.md+++ /dev/null@@ -1 +0,0 @@-- Refuse a NaN or out-of-range `minVertices` and `angle` in `triangulateFaces` and `splitSharpEdges` when called directly, instead of returning an unchanged mesh (#1666)diff --git a/changes/1669.md b/changes/1669.mddeleted file mode 100644index a6adab7a..00000000--- a/changes/1669.md+++ /dev/null@@ -1 +0,0 @@-- Fixed three agent-facing bugs: a branch snapshot now copies the forked recipe instead of sharing it, a tail light with a non-numeric `off` or `on` emission is refused instead of driving NaN, and an image sequence time that overflows frame arithmetic is refused instead of picking frame NaN (#1669)diff --git a/changes/1669.size.md b/changes/1669.size.mddeleted file mode 100644index 72b90d7f..00000000--- a/changes/1669.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 88 bytes gzip for three guards that stop a NaN tail light emission, a NaN sequence frame and a branch that shares its recipe with the caller.diff --git a/changes/1672.md b/changes/1672.mddeleted file mode 100644index 68241292..00000000--- a/changes/1672.md+++ /dev/null@@ -1 +0,0 @@-- Refuse an irradiance grid resolution under 2 or samples under 1, an odd probe map width and an unreachable adaptive `raiseBelow`, instead of writing NaN grids, dropping probe rows or pinning the resolution at its minimum (#1672)diff --git a/changes/1672.size.md b/changes/1672.size.mddeleted file mode 100644index 96a071d5..00000000--- a/changes/1672.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: three small guards (grid resolution and samples, even probe width, adaptive raiseBelow) that stop NaN grids and silently dropped probe rows.diff --git a/changes/1675.md b/changes/1675.mddeleted file mode 100644index 9bcf001e..00000000--- a/changes/1675.md+++ /dev/null@@ -1 +0,0 @@-- A non-finite look-at target or foot groundOffset no longer throws each frame: it keeps the last good value and warns once per controller (#1675)diff --git a/changes/1679.md b/changes/1679.mddeleted file mode 100644index 65cccbc1..00000000--- a/changes/1679.md+++ /dev/null@@ -1 +0,0 @@-- `flatten` without a normal now refuses a closed mesh instead of squashing it onto a random plane (#1679)diff --git a/changes/1681.md b/changes/1681.mddeleted file mode 100644index 5d2107a6..00000000--- a/changes/1681.md+++ /dev/null@@ -1 +0,0 @@-- `buildWithKernels` no longer hangs when its kernel worker goes silent: a new `kernelTimeoutMilliseconds` option (default 60000, above 0, at most 3600000) ends the worker and finishes the build inline (#1681).diff --git a/changes/1682.md b/changes/1682.mddeleted file mode 100644index 241e80b0..00000000--- a/changes/1682.md+++ /dev/null@@ -1 +0,0 @@-- Fix three LOD input bugs: `generateMeshLods` gives `relativeError` 0 instead of NaN for a zero-extent mesh (#1682), `addLodsToDenseParts` checks `ratios` on a sparse scene too (#1683), and `generateAssetLods` names a part that has no geometry (#1685).diff --git a/changes/1684.md b/changes/1684.mddeleted file mode 100644index 74fbff6f..00000000--- a/changes/1684.md+++ /dev/null@@ -1 +0,0 @@-- split_edges and delete_loose (and the other edits that rebuild from polygons) now keep vertex groups and other point attributes: a split vertex copies its source vertex's values and removed vertices drop theirs (#1684)diff --git a/changes/1684.size.md b/changes/1684.size.mddeleted file mode 100644index 82ba0646..00000000--- a/changes/1684.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: vertex groups and other point attributes survive split_edges, delete_loose and every polygon rebuild instead of silently vanishing (155 bytes gzip)diff --git a/changes/1686.md b/changes/1686.mddeleted file mode 100644index df50d82e..00000000--- a/changes/1686.md+++ /dev/null@@ -1 +0,0 @@-- `box()` now refuses a bevel too small to move its core instead of returning NaN normals, and `text()` draws a glyph whose overlapping contours start inside each other instead of throwing a TypeError (#1686, #1687)diff --git a/changes/1686.size.md b/changes/1686.size.mddeleted file mode 100644index 111f0021..00000000--- a/changes/1686.size.md+++ /dev/null@@ -1,2 +0,0 @@-- goal:agent-scene: about 10 bytes gzip for the one-line check that stops box() returning NaN normals for a bevel below 5e-17-- goal:lit-box: the same box() check, a few bytes gzipdiff --git a/changes/1695.md b/changes/1695.mddeleted file mode 100644index 25025c96..00000000--- a/changes/1695.md+++ /dev/null@@ -1 +0,0 @@-- A world renderer with no `shadowBounds` target or extent fits the sun shadow view to its shadow casters, so the shadow bias stays thinner than 0.2 m walls and a closed room no longer shows sun lines along its seams (#1695).diff --git a/changes/1695.size.md b/changes/1695.size.mddeleted file mode 100644index 2960ae93..00000000--- a/changes/1695.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- entry:core+lazy: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- scene:lit-box: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- scene:lit-box+lazy: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- scene:gltf-box: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- scene:gltf-box+lazy: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- scene:agent-scene: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- scene:agent-scene+lazy: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- goal:lit-box: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- goal:gltf-box: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)-- goal:agent-scene: fits the world sun shadow view to the casters so the shadow bias (about 22 cm before) stays thinner than 0.2 m walls; a closed room no longer shows sun lines along its seams (#1695)diff --git a/changes/1698.md b/changes/1698.mddeleted file mode 100644index d9771db0..00000000--- a/changes/1698.md+++ /dev/null@@ -1 +0,0 @@-- The MCP server decodes Draco-compressed GLBs in the asset catalog and packs with the optional `draco3d` package, so an agent can place a Sketchfab or NASA download as it is (#1698)diff --git a/changes/1704.md b/changes/1704.mddeleted file mode 100644index dc77a95a..00000000--- a/changes/1704.md+++ /dev/null@@ -1 +0,0 @@-- `subdivide` and `loop_cut` keep point attributes such as vertex groups: new points blend their sources, and integer values take the first one (#1704)diff --git a/changes/1704.size.md b/changes/1704.size.mddeleted file mode 100644index 7ee7d922..00000000--- a/changes/1704.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: vertex groups and other point attributes survive subdivide and loop_cut instead of silently vanishing, the last two mesh operators that dropped them (306 bytes)diff --git a/changes/1709.md b/changes/1709.mddeleted file mode 100644index 953fc679..00000000--- a/changes/1709.md+++ /dev/null@@ -1 +0,0 @@-- The MCP server transcodes KTX2 textures of catalog GLBs to PNG with the optional `three` package, so an agent can place an optimised download as it is (#1709)diff --git a/changes/1711.md b/changes/1711.mddeleted file mode 100644index f56ab86b..00000000--- a/changes/1711.md+++ /dev/null@@ -1 +0,0 @@-- Add focusDistance (metres, with fStop) to the recipe camera so an agent can focus on a subject without moving the camera target (#1711)diff --git a/changes/1711.size.md b/changes/1711.size.mddeleted file mode 100644index 6ee07345..00000000--- a/changes/1711.size.md+++ /dev/null@@ -1,8 +0,0 @@-- tool-total: agents can focus a camera on a subject without moving its target; the one field costs 88 bytes in each of four tools, detail stays in scene_capabilities-- tool:scene_apply: focusDistance sits next to fStop in the shared camera fields (88 bytes)-- tool:scene_create: focusDistance sits next to fStop in the shared camera fields (88 bytes)-- tool:scene_create_asset: focusDistance sits next to fStop in the shared camera fields (88 bytes)-- tool:scene_review: focusDistance sits next to fStop in the shared camera fields (88 bytes)-- scene:agent-scene: the focusDistance check with a message that names the field and a valid value (168 bytes gzip)-- scene:agent-scene+lazy: the focusDistance check with a message that names the field and a valid value (201 bytes gzip)-- goal:agent-scene: one validated lens field, the smallest check that names the field and a valid valuediff --git a/changes/1714-signed.md b/changes/1714-signed.mddeleted file mode 100644index f1699313..00000000--- a/changes/1714-signed.md+++ /dev/null@@ -1 +0,0 @@-- Dispersive glass keeps signed spectral weights as Cycles does, so prism fringes match its saturation, and negative pixels are clamped before OIDN and the display transform (#1714)diff --git a/changes/1714.md b/changes/1714.mddeleted file mode 100644index a1ed1022..00000000--- a/changes/1714.md+++ /dev/null@@ -1 +0,0 @@-- Glass can fringe into colour: a material `dispersion` from 0 to 10 reaches the native path tracer, and glTF `KHR_materials_dispersion` is read instead of ignored (#1714)diff --git a/changes/1714.size.md b/changes/1714.size.mddeleted file mode 100644index 1cc9f2a0..00000000--- a/changes/1714.size.md+++ /dev/null@@ -1,5 +0,0 @@-- scene:agent-scene: 65 bytes gzip for the `dispersion` material layer (its range, default and capabilities line) and the glTF `KHR_materials_dispersion` row, which the path tracer needs in the core to read the field.-- scene:agent-scene+lazy: 75 bytes gzip for the same layer field, counted again in the lazy build.-- goal:lit-box: 35 bytes gzip for the same layer field in the shared material code.-- goal:gltf-box: 56 bytes gzip for the layer field and the glTF extension row that stops `KHR_materials_dispersion` from being ignored.-- goal:agent-scene: 65 bytes gzip for the same layer field and capabilities line.diff --git a/changes/1722.md b/changes/1722.mddeleted file mode 100644index 6fce6b21..00000000--- a/changes/1722.md+++ /dev/null@@ -1 +0,0 @@-- Two soft volumes that overlap are marched together, so a soft edge no longer shows a hard slab where the other volume starts (#1722)diff --git a/changes/1722.size.md b/changes/1722.size.mddeleted file mode 100644index 0ed3e17f..00000000--- a/changes/1722.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 309 bytes gzip for marching overlapping box and sphere volumes together, in the rasterizer shader (a partner lane and a segment end) and in the CPU path tracer (#1722).diff --git a/changes/1723.md b/changes/1723.mddeleted file mode 100644index f19a0730..00000000--- a/changes/1723.md+++ /dev/null@@ -1 +0,0 @@-- `scene_look` no longer fails on a mesh local volume that names a node added by `scene_apply`: the renderer now keeps its part list after a topology change, which also keeps volume shadow ambient and irradiance grid fog current after an edit (#1723).diff --git a/changes/1725.md b/changes/1725.mddeleted file mode 100644index 107c2e0d..00000000--- a/changes/1725.md+++ /dev/null@@ -1 +0,0 @@-- Local volumes take `noiseOctaves` (1 to 5, default 1): finer octaves of noise on top of the large lumps give smoke and fire a ragged edge, in the rasterizer and the CPU path tracer (#1725)diff --git a/changes/1725.size.md b/changes/1725.size.mddeleted file mode 100644index 3548ff17..00000000--- a/changes/1725.size.md+++ /dev/null@@ -1,5 +0,0 @@-- scene:agent-scene: noiseOctaves adds the fractal noise to the CPU tracer, the shader variant and the capability text, 210 bytes gzip (449 with the lazy parts), and a scene without the field compiles the shader it had.-- scene:agent-scene+lazy: same change, the fractal noise function in the rasterizer's post shader variant and the CPU tracer's noise.-- goal:lit-box: 41 bytes gzip for the noiseOctaves field name and its check in the local volume resolver, shared by every bundle that carries local volumes.-- goal:gltf-box: 30 bytes gzip, the same field check.-- goal:agent-scene: 210 bytes gzip for the capability text, the fractal noise and the field check; a scene without the field keeps its shader text.diff --git a/changes/1729.md b/changes/1729.mddeleted file mode 100644index 94810010..00000000--- a/changes/1729.md+++ /dev/null@@ -1 +0,0 @@-- Native video jobs that blend frames over time trace natively again: the temporal blend accepted only a 320-byte kernel parameter block, but Cook sends 336 since its anamorphic aperture ratio, so every such job fell back to another backend on its first frame (#1729).diff --git a/changes/1733.md b/changes/1733.mddeleted file mode 100644index 40102089..00000000--- a/changes/1733.md+++ /dev/null@@ -1 +0,0 @@-- Breaking: a grid volume's `blackbodyIntensity` now gives the physical brightness and Planck colour of Blender's Principled Volume, so existing grid fires render 3 to 4 times dimmer; raise `blackbodyIntensity` or the exposure to get the old look back (#1733)diff --git a/changes/1733.size.md b/changes/1733.size.mddeleted file mode 100644index d2274f8f..00000000--- a/changes/1733.size.md+++ /dev/null@@ -1,4 +0,0 @@-- entry:light-bake-gpu: the grid blackbody now reads a hue table baked from the Planck integral of the emission ramp (about 1.1 KB gzip); that is what makes a grid fire match Cycles in colour as well as brightness, replacing the Tanner Helland fit.-- entry:path-tracer-progressive: same hue table, shared with the grid volume march of the CPU path tracer.-- entry:path-tracer-realtime: same hue table, shared with the grid volume march of the CPU path tracer.-- goal:agent-scene: the capabilities line says the grid blackbody is physical and gives its brightness at 1500 K and 3000 K (39 bytes gzip), so an agent can set emissionStrength without a trial render.diff --git a/changes/1735-detail.md b/changes/1735-detail.mddeleted file mode 100644index 47b5ea4b..00000000--- a/changes/1735-detail.md+++ /dev/null@@ -1 +0,0 @@-- `bakeGasSim` takes `detail` (0 to 1, default 0): a noise upres of each frame's smoke density, so a coarse sim breaks into puffs and gaps below its cell size; 0 keeps the bytes of every existing bake (#1735)diff --git a/changes/1735-detail.size.md b/changes/1735-detail.size.mddeleted file mode 100644index 28a1ddc7..00000000--- a/changes/1735-detail.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the gas sim's detail option is about 160 bytes gzip of noise code, the first noise upres for a baked explosion's smoke (#1735)diff --git a/changes/1735-mass.md b/changes/1735-mass.mddeleted file mode 100644index 7ec6f444..00000000--- a/changes/1735-mass.md+++ /dev/null@@ -1 +0,0 @@-- The gas-sim bake keeps smoke and fuel mass constant between sources and sinks: advection no longer swells the plume (a 3 s blast used to roughly double its smoke) (#1735).diff --git a/changes/1735.md b/changes/1735.mddeleted file mode 100644index 9fbf8795..00000000--- a/changes/1735.md+++ /dev/null@@ -1 +0,0 @@-- `bakeGasSim` bakes an explosion from a grid gas simulation (advection, pressure projection, vorticity, fuel burning into flame and heat) into grid volume frames, in `src/gas-sim.js` (#1735)diff --git a/changes/1736-emission.md b/changes/1736-emission.mddeleted file mode 100644index 4bbc7f63..00000000--- a/changes/1736-emission.md+++ /dev/null@@ -1 +0,0 @@-- add-explosion fireball emission is 9.5 instead of 14 at the reference size, so the core no longer clips to white and the edge falloff reaches 0.79 of the Cycles frame on a flat background, with the RMSE against it down from 0.030 to 0.020 (#1736)diff --git a/changes/1736.md b/changes/1736.mddeleted file mode 100644index 6a074a91..00000000--- a/changes/1736.md+++ /dev/null@@ -1 +0,0 @@-- add-explosion bakes with turbulence 1 and vorticity 40, so the fireball outline breaks up more (silhouette roughness 5.4 to 6.1 and edge falloff 0.74 to 0.82 against a Cycles reference 10.1 and 1.06); `scripts/measure-explosion-breakup.mjs` scores a blast against it (#1736)diff --git a/changes/1742.md b/changes/1742.mddeleted file mode 100644index a40a788c..00000000--- a/changes/1742.md+++ /dev/null@@ -1 +0,0 @@-- `vdb-to-grid.py` reads Blosc compressed VDBs, openvdb's default, when run with `uv run --no-project --with blosc` (#1742)diff --git a/changes/1744-palette.md b/changes/1744-palette.mddeleted file mode 100644index 50fbfccb..00000000--- a/changes/1744-palette.md+++ /dev/null@@ -1 +0,0 @@-- `createGaussianSplats` takes `palette` (2 to 256 colours) and `dither` (`bayer4` or `bayer8`) to snap `alphaCut` splats to a palette for a pixel-art look (#1744)diff --git a/changes/1744-palette.size.md b/changes/1744-palette.size.mddeleted file mode 100644index 18c1b3e8..00000000--- a/changes/1744-palette.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:gaussian-splats: 1230 bytes gzip for the palette and Bayer dither option (option checks, the palette fragment shader and its program cache), the pixel-art snapping splat users otherwise do in a separate image pass after every frame (#1744). The lead accepted this growth: a visible opt-in capability, with the default shader pinned.diff --git a/changes/1744.md b/changes/1744.mddeleted file mode 100644index dcb45209..00000000--- a/changes/1744.md+++ /dev/null@@ -1 +0,0 @@-- `createGaussianSplats` takes `alphaCut` to draw opaque, unblended discs and `parseGaussianSplatPly` takes `upAxis: 'z'` to load Z-up files such as Blender splats upright (#1744)diff --git a/changes/1744.size.md b/changes/1744.size.mddeleted file mode 100644index 49f5f3e3..00000000--- a/changes/1744.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:gaussian-splats: 571 bytes gzip for the opaque `alphaCut` fragment shader and its draw state, and the `upAxis: 'z'` turn in the PLY reader: the Blender-style hard-disc look and Z-up loading that splat users had to work around by hand (#1744).diff --git a/changes/1748.md b/changes/1748.mddeleted file mode 100644index 893ec986..00000000--- a/changes/1748.md+++ /dev/null@@ -1 +0,0 @@-- A grid volume with `castShadow: true` now shades the surfaces under it from the sun, so an explosion no longer floats over a lit floor (#1748)diff --git a/changes/1748.size.md b/changes/1748.size.mddeleted file mode 100644index aaa36d0f..00000000--- a/changes/1748.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the grid branch of the surface shadow pass (16 taps of the simulated density along the sun) and its CPU path tracer twin, +133 bytes gzip, so a grid explosion shades the floor under it as Cycles does.diff --git a/changes/1749.md b/changes/1749.mddeleted file mode 100644index 6fd15f48..00000000--- a/changes/1749.md+++ /dev/null@@ -1 +0,0 @@-- `render()` takes `linearOutput: true` with a float `outputFramebuffer` to write linear scene radiance after fog and volumes, with no view transform or clamp, and `compare-volume-render.mjs` reads it as a `.f32` candidate (#1749)diff --git a/changes/1749.size.md b/changes/1749.size.mddeleted file mode 100644index 74dc469d..00000000--- a/changes/1749.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: the GLSL and check moved to the lazy addon linear-output-shader (433 bytes gzip); core keeps only the three-line hook, 71 bytes over the base.-- entry:core+lazy: same hook as entry:core, 70 bytes.-- scene:lit-box: same hook, 67 bytes.-- scene:lit-box+lazy: same hook, 67 bytes.-- scene:gltf-box: same hook, 74 bytes.-- scene:gltf-box+lazy: same hook, 73 bytes.-- scene:agent-scene: same hook, 69 bytes.-- scene:agent-scene+lazy: same hook, 65 bytes.-- goal:lit-box: the 67-byte hook is the only core cost of linear float output, an agent-facing capability.-- goal:gltf-box: the 74-byte hook is the only core cost of linear float output, an agent-facing capability.-- goal:agent-scene: the 69-byte hook is the only core cost of linear float output, an agent-facing capability.diff --git a/changes/1752.md b/changes/1752.mddeleted file mode 100644index 05d2ef1d..00000000--- a/changes/1752.md+++ /dev/null@@ -1 +0,0 @@-- Grid fires no longer show vertical streaks, cross lines between voxels or a hard flat bottom edge in Glaze and the CPU path tracer (#1752)diff --git a/changes/1752.size.md b/changes/1752.size.mddeleted file mode 100644index e15ec388..00000000--- a/changes/1752.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:fog-shader: 391 bytes gzip for the baked blackbody texture of a grid (one linear tap instead of the per-step blackbody) and the padded texture, which remove the streaks, cross lines and hard bottom edge on fires with no extra cost per frame (#1752).-- scene:agent-scene+lazy: 371 bytes gzip, the same fog shader code carried inside the agent scene bundle.diff --git a/changes/1755.md b/changes/1755.mddeleted file mode 100644index 8ae654ce..00000000--- a/changes/1755.md+++ /dev/null@@ -1 +0,0 @@-- An agent can place a part by its top, bottom or centre face with a new `level` node relation (`{ "kind": "level", "at": "top", "y": 0.5 }`), so it no longer adds half the thickness by hand. Existing recipes are unchanged. (#1755)diff --git a/changes/1755.size.md b/changes/1755.size.mddeleted file mode 100644index ce5cc45d..00000000--- a/changes/1755.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the level relation adds about 360 bytes gzip of validation and solving to the relations module and one line to its capabilities topic, and removes a half-thickness sum agents get wrong about one run in four.-- goal:agent-scene: the core grows 38 bytes gzip for the level entry in the relations capabilities topic; the solver itself loads on demand.diff --git a/changes/1758.md b/changes/1758.mddeleted file mode 100644index f30cc878..00000000--- a/changes/1758.md+++ /dev/null@@ -1 +0,0 @@-- Particle systems of `type: "sparks"` throw emissive streaks that cool along the blackbody curve with gravity and drag, and `preset: "explosion"` makes a blast (#1758)diff --git a/changes/1758.size.md b/changes/1758.size.mddeleted file mode 100644index d180973f..00000000--- a/changes/1758.size.md+++ /dev/null@@ -1,2 +0,0 @@-- scene:agent-scene+lazy: the spark system (`src/particle-sparks.js`, the `sparks` type of the particles module) loads only in a scene that has particles; its bytes are the streak mesh, the cooling colour and the bounded, explained input errors, and it reuses the blackbody table of #1751 instead of a second copy.-- goal:agent-scene: 42 bytes gzip for the one `sparks` line in the particles capability text, so agents can find the new system type; the existing particle wording (force-field types, mesh emitter detail, that raising count keeps earlier particles) stays as it was rather than being trimmed to make room.diff --git a/changes/1759.md b/changes/1759.mddeleted file mode 100644index 225cfc5a..00000000--- a/changes/1759.md+++ /dev/null@@ -1 +0,0 @@-- A grid volume blurs along the gas velocity when the scene's motion blur shutter is on: bake with `velocity: true` and a fireball or smoke smears along its expansion in Glaze and the CPU path tracer, as in Cycles (#1759)diff --git a/changes/1759.size.md b/changes/1759.size.mddeleted file mode 100644index 0e4d32c2..00000000--- a/changes/1759.size.md+++ /dev/null@@ -1,12 +0,0 @@-- entry:fog-shader: the motion blur of a grid volume adds a velocity texture upload and a second program variant of the march (781 bytes gzip); it is the capability the issue asks for, gas that smears along its velocity as Cycles renders it, and a scene without a shutter compiles exactly the shader it compiled before and pays nothing at run time-- scene:lit-box: 79 bytes gzip for the velocity channel names, the shutter-time constants and the gridMotion program step in the shared modules the scene bundle carries-- scene:lit-box+lazy: the same 79 bytes gzip-- scene:gltf-box: 71 bytes gzip, the same shared module additions-- scene:gltf-box+lazy: the same 71 bytes gzip-- entry:core: the same velocity channel names, constants and program step in the core bundle-- entry:core+lazy: the same additions-- scene:agent-scene+lazy: the same fog shader code plus the binding for the shutter in the volume code-- goal:lit-box: 79 bytes gzip for the velocity channel names, the shutter-time constants and the gridMotion program step in the shared modules; the scene is over its goal and this is the capability the issue asks for-- goal:gltf-box: 71 bytes gzip, the same shared module additions-- goal:agent-scene: 55 bytes gzip, the same shared module additions-- scene:agent-scene: 55 bytes gzip, the same shared module additionsdiff --git a/changes/1763.md b/changes/1763.mddeleted file mode 100644index 4f262d88..00000000--- a/changes/1763.md+++ /dev/null@@ -1 +0,0 @@-- The shadow a grid volume casts on surfaces blurs along the gas velocity when the motion blur shutter is on, so a blurred blast no longer casts a sharp shadow (#1763)diff --git a/changes/1767.md b/changes/1767.mddeleted file mode 100644index 26e69633..00000000--- a/changes/1767.md+++ /dev/null@@ -1 +0,0 @@-- A render job or still no longer draws its first frame with the fallback post program while the driver builds the real one, so a cold first frame matches a warm one (#1767)diff --git a/changes/1767.size.md b/changes/1767.size.mddeleted file mode 100644index 2a47e794..00000000--- a/changes/1767.size.md+++ /dev/null@@ -1,3 +0,0 @@-- goal:lit-box: 23 bytes gzip for the exactFrames switch that makes a cold first frame equal a warm one-- goal:gltf-box: the same 31 bytes-- goal:agent-scene: the same 23 bytesdiff --git a/changes/1768-bake.md b/changes/1768-bake.mddeleted file mode 100644index 745c80a2..00000000--- a/changes/1768-bake.md+++ /dev/null@@ -1 +0,0 @@-- The fireball bake that every explosion scene sends with `scene_create` runs about a third to a half faster (72 frames at 32 cubed, same bytes): the gas solver locates each traced position once per step instead of once per field (#1768)diff --git a/changes/1768-bake.size.md b/changes/1768-bake.size.mddeleted file mode 100644index 981d4644..00000000--- a/changes/1768-bake.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 160 bytes gzip buy a gas solver that locates each traced position once per step, so the 72-frame blast bake every explosion scene_create waits for runs about a third to a half faster with the same bytes.diff --git a/changes/1768-cache.md b/changes/1768-cache.mddeleted file mode 100644index fb7a6af9..00000000--- a/changes/1768-cache.md+++ /dev/null@@ -1 +0,0 @@-- The MCP server keeps baked add-explosion frames in a cache folder, so a second process that needs the same blast loads it instead of simulating for about 5 seconds (#1768)diff --git a/changes/1768-cache.size.md b/changes/1768-cache.size.mddeleted file mode 100644index ad00075a..00000000--- a/changes/1768-cache.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the store hook and bake descriptor in explosion-grids.js (71 bytes gzip) let the MCP server reuse baked blasts across processes, saving about 5 s per first grid scenediff --git a/changes/1768-create-timeout.md b/changes/1768-create-timeout.mddeleted file mode 100644index 2133aa27..00000000--- a/changes/1768-create-timeout.md+++ /dev/null@@ -1 +0,0 @@-- A `scene_create` whose recipe has a grid volume may run 60 s instead of 30 s, the same limit `add-explosion` already had, because the first draw compiles its shader in the graphics driver (#1768)diff --git a/changes/1768-shader-cache.md b/changes/1768-shader-cache.mddeleted file mode 100644index 98181041..00000000--- a/changes/1768-shader-cache.md+++ /dev/null@@ -1 +0,0 @@-- The MCP server keeps Chrome's compiled GPU shaders between runs, so the first volume `scene_create` of a new server process takes about 0.8 s instead of 4.9 s; `CRUCIBLE_SHADER_CACHE=0` turns it off (#1768)diff --git a/changes/1768-variants.md b/changes/1768-variants.mddeleted file mode 100644index fed7ecd8..00000000--- a/changes/1768-variants.md+++ /dev/null@@ -1 +0,0 @@-- A scene with only grid volumes, or only box and sphere volumes, builds the shader of that kind alone, about half the shader build of the combined program, with the same pixels (#1768)diff --git a/changes/1768-variants.size.md b/changes/1768-variants.size.mddeleted file mode 100644index 5c26392b..00000000--- a/changes/1768-variants.size.md+++ /dev/null@@ -1,12 +0,0 @@-- entry:fog-shader: the volume march is split into a grid part and a box/sphere part that a frame with one kind of volume compiles alone, so the driver builds about half the code-- entry:core: the grid-only and shape-only program markers and the volume kind that picks them, about 100 bytes gzip-- entry:core+lazy: the same markers-- scene:lit-box: the same markers, in every scene that bundles the renderer-- scene:lit-box+lazy: the same markers-- scene:gltf-box: the same markers-- scene:gltf-box+lazy: the same markers-- scene:agent-scene: the same markers-- scene:agent-scene+lazy: the same markers and the split march, for a build that is about half as long on one kind of volume-- goal:lit-box: the program markers add about 110 bytes gzip so a scene with one kind of volume builds half the shader, which is what keeps a grid volume under the preview limit; approved by the lead: cold grid compile ~0.55x-- goal:gltf-box: the same markers, about 100 bytes gzip; approved by the lead: cold grid compile ~0.55x-- goal:agent-scene: the same markers, about 155 bytes gzip; approved by the lead: cold grid compile ~0.55xdiff --git a/changes/1768.md b/changes/1768.mddeleted file mode 100644index 72e780d5..00000000--- a/changes/1768.md+++ /dev/null@@ -1 +0,0 @@-- A glowing or shadow-casting grid volume no longer stalls the first draw: its shader builds in about 8 s instead of 45 s on ANGLE d3d11, and a cold `scene_create` fits well inside the 30 s preview limit (#1768)diff --git a/changes/1768.size.md b/changes/1768.size.mddeleted file mode 100644index f002f8b6..00000000--- a/changes/1768.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the grid volume shader builds in about 8 s instead of 45 s, for 200 bytes gzip of split in-scatter functionsdiff --git a/changes/1770.md b/changes/1770.mddeleted file mode 100644index 3a46621f..00000000--- a/changes/1770.md+++ /dev/null@@ -1 +0,0 @@-- Environment variables now take the `CRUCIBLE_` name first (`CRUCIBLE_NO_UPDATE_CHECK`, `CRUCIBLE_MCP_LOG`); the `SIXTYSEVENJS_` names keep working as aliases (#1770)diff --git a/changes/1771.md b/changes/1771.mddeleted file mode 100644index 9fb789b0..00000000--- a/changes/1771.md+++ /dev/null@@ -1 +0,0 @@-- `scene_apply` refuses an add-node that would take the scene's unique assets past the 64 MiB combined limit, with a clear message, instead of letting the later render job crash the page ('Target page closed') (#1771)diff --git a/changes/1775.md b/changes/1775.mddeleted file mode 100644index cb05600e..00000000--- a/changes/1775.md+++ /dev/null@@ -1 +0,0 @@-- `add-explosion` takes `sparks` and `motionBlur` options, on by default: spark streaks and a velocity motion blur of the fireball (#1775)diff --git a/changes/1775.size.md b/changes/1775.size.mddeleted file mode 100644index 7f587e5e..00000000--- a/changes/1775.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: the capability text for the two new add-explosion options (sparks, motionBlur), 74 bytes gzip, so an agent can turn the approved sparks and blur on or off.-- scene:agent-scene+lazy: the option checks and the spark streak specification of add-explosion, 819 bytes gzip; the velocity bake itself is server side and adds nothing to the bundle.-- goal:agent-scene: same 74 bytes of capability text as above; the options are an agent-facing capability the issue asks for.diff --git a/changes/1776.md b/changes/1776.mddeleted file mode 100644index c0706906..00000000--- a/changes/1776.md+++ /dev/null@@ -1 +0,0 @@-- `crucible-mcp --memory` goes up to half of the machine RAM (at most 4096 MiB) and the render job memory guard and the native tracer allowance follow it, so 1080p and native path trace jobs of heavier scenes finish (#1776, #1696)diff --git a/changes/1777.md b/changes/1777.mddeleted file mode 100644index 65c72c51..00000000--- a/changes/1777.md+++ /dev/null@@ -1 +0,0 @@-- Particle emitter systems take `start` (the same field as `delay`) and `stop` in seconds, so a system with a `rate` can begin partway through a film and stop before it ends; nothing is born at or after `stop`, and systems without them are unchanged (#1777)diff --git a/changes/1777.size.md b/changes/1777.size.mddeleted file mode 100644index bd546954..00000000--- a/changes/1777.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene+lazy: the start and stop emission window of particle systems (about 230 bytes gzip of validation and the birth check).-- scene:agent-scene: the same window code; agent-scene is over its goal and this is a requested film feature (#1777), 16 bytes gzip.-- goal:agent-scene: the same 16 bytes gzip, the emission window of particle systems.diff --git a/changes/1785-ramp-octaves.md b/changes/1785-ramp-octaves.mddeleted file mode 100644index 39a4bf0f..00000000--- a/changes/1785-ramp-octaves.md+++ /dev/null@@ -1 +0,0 @@-- The WebGPU path tracer now draws local volumes with `emissionRamp` and `noiseOctaves`, matching the CPU tracer; before, backend "webgpu" threw for both (#1785).diff --git a/changes/1785-ramp-octaves.size.md b/changes/1785-ramp-octaves.size.mddeleted file mode 100644index 3f6961ae..00000000--- a/changes/1785-ramp-octaves.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:light-bake-gpu+lazy: the ramp light, fractal noise and ramp packer add about 1 KB gzip to the same lazy WebGPU chunk (#1785); the core entries do not grow.-- entry:path-tracer+lazy: the same lazy WebGPU chunk (grid march, ramp light, fractal noise and packers, #1785); loads only for backend "webgpu", and the non-lazy entry grows by 55 bytes.diff --git a/changes/1785.md b/changes/1785.mddeleted file mode 100644index 3ae08b67..00000000--- a/changes/1785.md+++ /dev/null@@ -1 +0,0 @@-- The WebGPU path tracer now draws grid volumes (the #1735 blast, smoke and fire baked to a grid): density, flame, baked blackbody and, with `volumeShutter`, the motion blur along the velocity, matching the CPU tracer; before, backend "webgpu" threw and named "cpu" (#1785).diff --git a/changes/1785.size.md b/changes/1785.size.mddeleted file mode 100644index 321952f2..00000000--- a/changes/1785.size.md+++ /dev/null@@ -1,3 +0,0 @@-- entry:light-bake-gpu+lazy: about 2 KB gzip for the WebGPU grid volume march and the grid packer (#1785), in the lazy WebGPU path tracer chunk that only loads for backend "webgpu"; the core entries do not grow.-- entry:path-tracer-progressive+lazy: the same lazy WebGPU chunk (grid march and packer, #1785); loads only for backend "webgpu".-- entry:path-tracer-realtime+lazy: the same lazy WebGPU chunk (grid march and packer, #1785); loads only for backend "webgpu".diff --git a/changes/1786.md b/changes/1786.mddeleted file mode 100644index 0ac35721..00000000--- a/changes/1786.md+++ /dev/null@@ -1 +0,0 @@-- Native sequence jobs keep one Cook session through moving geometry such as particles: new triangles rebuild only the changed meshes, and each frame's route reports `changedMeshes` (#1786)diff --git a/changes/1789.md b/changes/1789.mddeleted file mode 100644index 882c9ae7..00000000--- a/changes/1789.md+++ /dev/null@@ -1 +0,0 @@-- `backend: "native"` path traces a grid volume with Cook, and the Cook pin has grid volumes (#1789)diff --git a/changes/1793.md b/changes/1793.mddeleted file mode 100644index 318f9403..00000000--- a/changes/1793.md+++ /dev/null@@ -1 +0,0 @@-- The CPU tracer no longer traces a hidden part as a wall in a scene with local volumes, so a soft volume next to an idle hidden part is not cut by a hard edge any more (#1793)diff --git a/changes/1806.md b/changes/1806.mddeleted file mode 100644index 793f8105..00000000--- a/changes/1806.md+++ /dev/null@@ -1 +0,0 @@-- The CPU path tracer never traces a part marked `visible: false` any more, so a hidden box casts no shadow and is not drawn, as in Glaze, while a mesh volume that names a hidden part still takes its shape from it (#1806)diff --git a/changes/1811.md b/changes/1811.mddeleted file mode 100644index 11445599..00000000--- a/changes/1811.md+++ /dev/null@@ -1 +0,0 @@-- `scripts/render-file-native.mjs` serves a GLB above 32 MiB over a local HTTP server instead of a Playwright route, so a single-mesh GLB of 4.8 million triangles (about 82 MiB) no longer closes the loader tab and reaches the native trace's triangle-count error (#1811).diff --git a/changes/1812.md b/changes/1812.mddeleted file mode 100644index 4c922afa..00000000--- a/changes/1812.md+++ /dev/null@@ -1 +0,0 @@-- A light list over its cap now throws an error that names the cap, says it comes from the 32 bit light-linking mask (1 sun, 16 point, 8 spot, 7 area) and points to merging lights, an HDRI or emissive parts, or `backend: native` (#1812)diff --git a/changes/1812.size.md b/changes/1812.size.mddeleted file mode 100644index bda40337..00000000--- a/changes/1812.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:agent-scene: the error for a light list over its cap now says why the cap exists (the 32 bit light-linking mask) and what to do instead; agents read these errors, and the text lives only in the agent scene, not in the core.-- scene:agent-scene+lazy: same error text.-- goal:agent-scene: same error text, the only growth of this change (about 290 bytes gzip).diff --git a/changes/1817.md b/changes/1817.mddeleted file mode 100644index 4a68f507..00000000--- a/changes/1817.md+++ /dev/null@@ -1 +0,0 @@-- Materials take `retro: { vertexSnap: [width, height], affineUV: true }` for the PS1 look in the raster renderer: vertices snap to a screen grid and UVs interpolate without perspective correction. Materials without it render as before.diff --git a/changes/1817.size.md b/changes/1817.size.mddeleted file mode 100644index 11823859..00000000--- a/changes/1817.size.md+++ /dev/null@@ -1,11 +0,0 @@-- entry:core: 109 bytes gzip for the retro hooks in the core (one vertex and one fragment slot, a feature key, the batch key and the updateParts check that rejects a retro change after creation with an error that says what to do); the retro GLSL, validation and binding are in the lazy `retro-shader` module. The restored check was the lead's decision (a silent no-op is what agents get stuck on) and costs about 50 of these bytes.-- entry:core+lazy: 109 bytes gzip, the same hooks.-- scene:lit-box: 99 bytes gzip, the same hooks.-- scene:lit-box+lazy: 98 bytes gzip, the same hooks.-- scene:gltf-box: 109 bytes gzip, the same hooks; the lead accepts it.-- scene:gltf-box+lazy: 108 bytes gzip, the same hooks.-- scene:agent-scene: 325 bytes gzip for the same hooks, the `retro` recipe field, its lazy module entry and its scene_capabilities topic; the lead accepts it.-- scene:agent-scene+lazy: 1288 bytes gzip, the lazy `retro-shader` module (validation, GLSL, binding) that loads only for a recipe with `retro`, plus the same core hooks.-- goal:lit-box: 99 bytes gzip, the feature hooks the PS1 look needs in the core; the maintainer decides.-- goal:gltf-box: 109 bytes gzip, the same hooks; the maintainer decides.-- goal:agent-scene: 325 bytes gzip, the same hooks and the capabilities line; the maintainer decides.diff --git a/changes/1825.md b/changes/1825.mddeleted file mode 100644index 2af78dde..00000000--- a/changes/1825.md+++ /dev/null@@ -1 +0,0 @@-- Blender export reports a light's max bounces below the scene's as not reproduced (Cook has one global bounce limit, no per-light field) (#1825)diff --git a/changes/1833.md b/changes/1833.mddeleted file mode 100644index fa095cd7..00000000--- a/changes/1833.md+++ /dev/null@@ -1 +0,0 @@-- The path tracer lights a scene with `environment.keyIntensity` (a scene with no sun, sky or lamp traces black) and applies `effects.exposure` to the traced picture, as the rasterizer does (#1833)diff --git a/changes/1834.md b/changes/1834.mddeleted file mode 100644index da6e3298..00000000--- a/changes/1834.md+++ /dev/null@@ -1 +0,0 @@-- A post pass or material fragment is measured at 4096 bytes of code: comments and extra white space no longer count, and the error gives both sizes (#1834).diff --git a/changes/1834.size.md b/changes/1834.size.mddeleted file mode 100644index 918bbe2b..00000000--- a/changes/1834.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 81 bytes gzip for a code-only size measure and a two-number error, which let one commented effect shader stay one post pass.diff --git a/changes/1835.md b/changes/1835.mddeleted file mode 100644index 9804e0b4..00000000--- a/changes/1835.md+++ /dev/null@@ -1 +0,0 @@-- A post pass reads its `time` from the motion time when the frame has one (a still at `motion.start: 6.5` gets 6.5, as the video frame does), instead of the renderer time, which stayed 0 in a still (#1835)diff --git a/changes/1843.md b/changes/1843.mddeleted file mode 100644index 6778b31b..00000000--- a/changes/1843.md+++ /dev/null@@ -1 +0,0 @@-- A recipe places a catalog GLB by file name, `asset: { "name": "sheep_body.glb" }`, in `scene_create`, `scene_apply` and (as `assetName`) `scene_create_asset`, so a re-exported file no longer breaks it; the scene keeps the resolved SHA-256 (#1843)diff --git a/changes/1843.size.md b/changes/1843.size.mddeleted file mode 100644index b52d5ed8..00000000--- a/changes/1843.size.md+++ /dev/null@@ -1 +0,0 @@-- goal:agent-scene: 27 bytes gzip, the `name` field in the placedAssets capability note and a shared placement walker; naming a catalog file lets a recipe survive a re-export, and the name-resolving code lives outside the core bundlediff --git a/changes/1844.md b/changes/1844.mddeleted file mode 100644index 70f5fc34..00000000--- a/changes/1844.md+++ /dev/null@@ -1 +0,0 @@-- Native "cycles" exports keep a part with a base or coat normal map and a non-uniform matrix as an instance, so Cook tilts the normal map from the matrix like Cycles instead of from baked unit normals (#1844).diff --git a/changes/1848.md b/changes/1848.mddeleted file mode 100644index 480e2cdf..00000000--- a/changes/1848.md+++ /dev/null@@ -1 +0,0 @@-- `scene_apply`, `scene_render_submit` and `scene_review` count the scene's grid (.67vg) files toward the 64 MiB combined cap, and a scene already over it fails with the cap message instead of crashing the page (#1848)diff --git a/changes/1852-sequence-folder.md b/changes/1852-sequence-folder.mddeleted file mode 100644index f933be45..00000000--- a/changes/1852-sequence-folder.md+++ /dev/null@@ -1 +0,0 @@-- A folder of more than 256 consecutively numbered images is now one `sequences` entry in `scene_asset_catalog` (pattern, first, last, frames, size) instead of one entry per frame, and a material `map.sequence` can name it with `catalog` instead of listing the frame IDs; the frames are hashed when named, not at startup (#1852).diff --git a/changes/1852-sequence-folder.size.md b/changes/1852-sequence-folder.size.mddeleted file mode 100644index 814b3bc5..00000000--- a/changes/1852-sequence-folder.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 209 bytes gzip for the `catalog` line of the texture sequence topic, the `catalog` field name and its advice error, so one folder name replaces thousands of frame IDs in a recipe (415,855 bytes down to 173 for the catalog listing of 2,700 frames).diff --git a/changes/1852.md b/changes/1852.mddeleted file mode 100644index c8ffad9e..00000000--- a/changes/1852.md+++ /dev/null@@ -1 +0,0 @@-- A material `map.sequence` may list up to 4096 frames of 2048 pixels: a sequence over 64 frames streams, decoding the frame of each time and keeping two, so a canvas that changes over a whole film can be textured (#1852)diff --git a/changes/1852.size.md b/changes/1852.size.mddeleted file mode 100644index 857c7c92..00000000--- a/changes/1852.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: the streamed texture sequence window (first frame only, frame of the current time and the one before it, byte-bounded), which lets a canvas that changes over a whole film be textured instead of the 64 frame cap.diff --git a/changes/1853.md b/changes/1853.mddeleted file mode 100644index baea3e81..00000000--- a/changes/1853.md+++ /dev/null@@ -1 +0,0 @@-- A `stills` render job takes `stillCount` (2 to 300) and writes that many numbered PNGs, one per 1/fps s of the motion clock, so a transparent overlay comes from one job; `scene_capabilities` jobs also states the largest square frame (#1853).diff --git a/changes/1853.size.md b/changes/1853.size.mddeleted file mode 100644index b4566d42..00000000--- a/changes/1853.size.md+++ /dev/null@@ -1,3 +0,0 @@-- tool-total: `stillCount` lets one stills job write a numbered frame range, so a transparent overlay needs one job instead of dozens (about 90 bytes, no description; the detail is in scene_capabilities jobs).-- tool:scene_render_submit: same field, the only way to ask for more than 8 stills.-- goal:agent-scene: 26 bytes gzip for the stills frame cap and the square-frame limits scene_capabilities reports; they replace guessing the size from a failed submit.diff --git a/changes/1855.md b/changes/1855.mddeleted file mode 100644index da5233ab..00000000--- a/changes/1855.md+++ /dev/null@@ -1 +0,0 @@-- A local volume with `noiseOctaves` above 1 takes its march steps from its finest noise cell (up to 96), so its edge no longer comes out wider than the real density field (#1855)diff --git a/changes/1855.size.md b/changes/1855.size.mddeleted file mode 100644index ecc345c7..00000000--- a/changes/1855.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 66 bytes gzip (limit 64) for the octave step rule in the post shader and the CPU tracer, which removes a 22 to 42 percent wider edge on fractal volumes (#1855)diff --git a/changes/1861.md b/changes/1861.mddeleted file mode 100644index fc8ce2e1..00000000--- a/changes/1861.md+++ /dev/null@@ -1 +0,0 @@-- A native render job and `npm run render:file-cook` read the scene tables from the page as one Blob over the DevTools stream, which cuts the per-frame export of a dense scene with 1,000 particles from about 125 ms to about 64 ms (#1861)diff --git a/changes/1872.md b/changes/1872.mddeleted file mode 100644index 5627bd0c..00000000--- a/changes/1872.md+++ /dev/null@@ -1 +0,0 @@-- `scene_capabilities` takes `topic` as a list of up to 8 names and answers each one in one call, keyed by name (#1872)diff --git a/changes/1892.md b/changes/1892.mddeleted file mode 100644index fd4ad92e..00000000--- a/changes/1892.md+++ /dev/null@@ -1 +0,0 @@-- A native render that falls back because the installed Cook addon is stale now says so, with Cook's own fix (`npx cook fetch-addon`), instead of blaming a missing Vulkan ray query adapter (#1892)diff --git a/changes/1894.md b/changes/1894.mddeleted file mode 100644index 6a9d630a..00000000--- a/changes/1894.md+++ /dev/null@@ -1 +0,0 @@-- Shape and grid volumes weight each march step's glow by the exact integral over the step, `(1 - exp(-tau)) / tau` per channel, in the rasterizer, the CPU path tracer and the WebGPU path tracer, so a dense glowing volume is dimmer and its edge narrower: the 1-octave sphere edge goes from 0.2926 to 0.2295 sphere radii against Cycles 0.2330 (#1894)diff --git a/changes/1894.size.md b/changes/1894.size.mddeleted file mode 100644index 49202298..00000000--- a/changes/1894.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 193 bytes gzip for the exact step weight `(1 - exp(-tau)) / tau` in the rasterizer's shape and grid marches, the CPU path tracer's shape, grid and mesh marches and the WebGPU kernel's two marches. It brings the 1-octave glowing sphere edge from 0.2926 to 0.2295 sphere radii against Cycles' 0.2330, a visible fix for dense glowing volumes (#1894).diff --git a/changes/1897.md b/changes/1897.mddeleted file mode 100644index 58dbe52a..00000000--- a/changes/1897.md+++ /dev/null@@ -1 +0,0 @@-- The shipped scene skill now matches the tools: the scene contract says up to seven area lights, `light.md` says an area light casts a shadow with `castShadow: true`, and `look.md` gives 8 MiB as the `.hdr` file limit (#1897)diff --git a/changes/1899.md b/changes/1899.mddeleted file mode 100644index 59600fdf..00000000--- a/changes/1899.md+++ /dev/null@@ -1 +0,0 @@-- `createGaussianSplats` and `createGaussianSplatsAsync` now refuse a fractional `count`, a missing or short `positions`, `scales`, `rotations` or `colors` array, and `matrix: null` or `options: null`, each with an error that names the field, the expected length and what to pass (#1899)diff --git a/changes/1899.size.md b/changes/1899.size.mddeleted file mode 100644index f3238ad1..00000000--- a/changes/1899.size.md+++ /dev/null@@ -1 +0,0 @@-- entry:gaussian-splats: 561 bytes gzip buy named, actionable errors for short or missing splat arrays, a fractional count and null options or matrix; hand-built splats (make-pixel-splats) used to draw garbage with no errordiff --git a/changes/1900.md b/changes/1900.mddeleted file mode 100644index 811e54a6..00000000--- a/changes/1900.md+++ /dev/null@@ -1 +0,0 @@-- `simplifyGeometry` throws a `RangeError` that names the bad index, its place in `indices` and the vertex count (or the positions length when it is not a multiple of three) instead of crashing in the collapser, and treats `options: null` as the defaults (#1900)diff --git a/changes/1900.size.md b/changes/1900.size.mddeleted file mode 100644index 525a810a..00000000--- a/changes/1900.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 134 bytes gzip buy agent-readable errors (bad index, positions length) in place of a crash inside the collapser (#1900)diff --git a/changes/1903.md b/changes/1903.mddeleted file mode 100644index 3af8eedd..00000000--- a/changes/1903.md+++ /dev/null@@ -1 +0,0 @@-- A `scene_apply` or `scene_create` call with several distinct `add-explosion` operations is now refused at once when its grids pass the 64 MiB combined limit: the check counts each blast at its exact size before baking, instead of after minutes of synchronous baking (#1903).diff --git a/changes/1904.md b/changes/1904.mddeleted file mode 100644index b12e6506..00000000--- a/changes/1904.md+++ /dev/null@@ -1 +0,0 @@-- `scene_capabilities` no longer answers names inherited from Object.prototype (`environment.constructor`, `parts.toString`) as fields or entries; a dotted suffix that is not an entry of a plain topic now fails with the list of entries (#1904)diff --git a/changes/1910.md b/changes/1910.mddeleted file mode 100644index 6b0457e7..00000000--- a/changes/1910.md+++ /dev/null@@ -1 +0,0 @@-- `EXT_meshopt_compression` views are checked before decoding: a bad `byteStride`, `byteOffset`, `byteLength`, `count`, `mode` or `filter` now throws an error naming the buffer view, the field, the value and what to pass, instead of `Malformed buffer data: -1`.diff --git a/changes/1910.size.md b/changes/1910.size.mddeleted file mode 100644index adb3a332..00000000--- a/changes/1910.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:gltf-box+lazy: the meshopt decoder chunk now validates every extension field and names the buffer view, field, value and allowed values, so an agent can fix a bad file instead of reading "Malformed buffer data: -1" (353 bytes gzip, error text included).diff --git a/changes/1913.md b/changes/1913.mddeleted file mode 100644index 8f2bfbd4..00000000--- a/changes/1913.md+++ /dev/null@@ -1 +0,0 @@-- A Draco GLB whose point or face count disagrees with its accessor `count`, or whose decoded buffer would pass the asset byte limit, is refused with an error naming the primitive, the attribute and both numbers before the decoder heap allocates it (#1913)diff --git a/changes/1919.md b/changes/1919.mddeleted file mode 100644index c9626cfb..00000000--- a/changes/1919.md+++ /dev/null@@ -1 +0,0 @@-- `readGlb` in the catalog scripts now refuses a GLB with a wrong header version or total length, an unaligned or past-the-end chunk length, a zero or second BIN chunk, and a bad JSON chunk, each with an error that names the chunk and the lengths (#1919).diff --git a/changes/1921.md b/changes/1921.mddeleted file mode 100644index e49d4c9a..00000000--- a/changes/1921.md+++ /dev/null@@ -1 +0,0 @@-- A glTF primitive compressed with KHR_draco_mesh_compression now applies a sparse block on its accessors on top of the decoded data instead of ignoring it (#1921).diff --git a/changes/1921.size.md b/changes/1921.size.mddeleted file mode 100644index f0861f71..00000000--- a/changes/1921.size.md+++ /dev/null@@ -1,3 +0,0 @@-- scene:gltf-box: 147 bytes gzip buys the sparse substitution on Draco-compressed accessors, which were silently ignored (wrong geometry, no error).-- scene:gltf-box+lazy: same code as gltf-box.-- goal:gltf-box: the scene is over its goal already; this fixes silently wrong geometry for a rare but real file shape, and the 147 bytes are the fix plus its type guard.diff --git a/changes/1927.md b/changes/1927.mddeleted file mode 100644index 3f5bb922..00000000--- a/changes/1927.md+++ /dev/null@@ -1 +0,0 @@-- `scene_render_status` of a finished job now returns its record while a newer job runs, instead of failing with "The renderer is leased".diff --git a/changes/1928.md b/changes/1928.mddeleted file mode 100644index fe079c3c..00000000--- a/changes/1928.md+++ /dev/null@@ -1 +0,0 @@-- A render job with `animation` on a scene that has no clips now fails with an error that says to drop `animation` or add a clip, and a missing clips list counts as empty instead of a TypeError (#1928)diff --git a/changes/1929.md b/changes/1929.mddeleted file mode 100644index 792c7bfb..00000000--- a/changes/1929.md+++ /dev/null@@ -1 +0,0 @@-- `scene_capabilities` with a topic list answers the valid topics and reports unknown names (at most 4) with suggestions in `unknown`, and an unknown op name suggests the dotted `meshEdit.<op>` form.diff --git a/changes/1938.md b/changes/1938.mddeleted file mode 100644index 2e249b61..00000000--- a/changes/1938.md+++ /dev/null@@ -1 +0,0 @@-- The shipped skills now cover review links with `reviewFile`, placing a GLB by file name, catalog image sequences, topic lists with `unknown`, `stillCount` stills jobs and the no-clips `animation` error (#1938)diff --git a/changes/1939.md b/changes/1939.mddeleted file mode 100644index 0ae39a89..00000000--- a/changes/1939.md+++ /dev/null@@ -1 +0,0 @@-- The agent bench passes the skill prompt to `claude` as a file, so `--skill-mode references` runs on Windows instead of failing with `spawn ENAMETOOLONG` (#1939)diff --git a/changes/1971.md b/changes/1971.mddeleted file mode 100644index afdeef78..00000000--- a/changes/1971.md+++ /dev/null@@ -1 +0,0 @@-- `detachedPartFlight` now settles a part with a `size` on its most stable face (the largest it can reach by falling or tipping with its landing energy) instead of the nearest face, so a bumper no longer ends standing on its narrow edge (#1971).diff --git a/changes/1971.size.md b/changes/1971.size.mddeleted file mode 100644index a54f962a..00000000--- a/changes/1971.size.md+++ /dev/null@@ -1 +0,0 @@-- scene:agent-scene+lazy: 247 bytes gzip for choosing the most stable resting face of a detached part in the lazy `vehicle-damage` module, which loads only for a recipe with a `dented` geometry. It fixes a bumper ending on its narrow edge, a visible glitch. A fixed glitch for 247 bytes. Approved by lead.diff --git a/changes/905-report-noise.md b/changes/905-report-noise.mddeleted file mode 100644index 4a649f20..00000000--- a/changes/905-report-noise.md+++ /dev/null@@ -1 +0,0 @@-- A finished path traced job reports `noise` (relative noise before denoising and a one-line verdict) in `scene_render_result` (#905).diff --git a/changes/905.md b/changes/905.mddeleted file mode 100644index e492108f..00000000--- a/changes/905.md+++ /dev/null@@ -1 +0,0 @@-- `scene_look` and render jobs name the strongest lights, with the field to lower, in an `exposure-clipped` finding, so an over-exposed light can be fixed without opening the image (#905).diff --git a/changes/923.md b/changes/923.mddeleted file mode 100644index 5389686b..00000000--- a/changes/923.md+++ /dev/null@@ -1 +0,0 @@-- `scene_ask_user` can show its annotate view inline in chat apps that support MCP Apps, with the same marks as the browser flow (#923)diff --git a/changes/924.md b/changes/924.mddeleted file mode 100644index 012511c5..00000000--- a/changes/924.md+++ /dev/null@@ -1 +0,0 @@-- A scene can be exported as a phone review page (`scene_export_web` with `review: true`) whose saved marks `scene_ask_user` reads back with `reviewFile` (#924)diff --git a/changes/924.size.md b/changes/924.size.mddeleted file mode 100644index 14be26cb..00000000--- a/changes/924.size.md+++ /dev/null@@ -1,3 +0,0 @@-- tool:scene_ask_user: 86 bytes for the reviewFile input that reads a saved review page file back, instead of a new tool (#924).-- tool:scene_export_web: 75 bytes for the review flag that writes a phone review page, instead of a new tool (#924).-- tool-total: 161 bytes (74504 to 74665) for the two inputs above.diff --git a/changes/983-doors.md b/changes/983-doors.mddeleted file mode 100644index 6e034944..00000000--- a/changes/983-doors.md+++ /dev/null@@ -1 +0,0 @@-- `hingedDoorSwing` in `crucible/addons/vehicle` swings a damaged door open on its hinge: a torque-driven, damped swing that stops at `maxAngle` with a small bounce and rests (#983)diff --git a/changes/983-doors.size.md b/changes/983-doors.size.mddeleted file mode 100644index 44377aa2..00000000--- a/changes/983-doors.size.md+++ /dev/null@@ -1,2 +0,0 @@-- entry:vehicle: 976 bytes gzip for `hingedDoorSwing`, the torque-driven hinge swing of a damaged door (damping, a stop at maxAngle with a bounce, a rest, and the door transform about the hinge line), which the addon did not have (#983). It is opt-in and only ships in the vehicle addon entry. Approved by lead: a visible opt-in capability.-- scene:agent-scene+lazy: 1052 bytes gzip for the same function inside the lazy `vehicle-damage` module, which loads only for a recipe with a `dented` geometry. It is the same vehicle code counted again in the lazy agent chunk, not a second cost. Approved by lead.diff --git a/changes/983-flight.md b/changes/983-flight.mddeleted file mode 100644index b1b288e7..00000000--- a/changes/983-flight.md+++ /dev/null@@ -1 +0,0 @@-- `detachedPartFlight` in `crucible/addons/vehicle` gives a loose part torn off by a crash a seeded flight: launch from the impulse above its threshold, tumble, bounce and rest on the ground (#983)diff --git a/changes/983-flight.size.md b/changes/983-flight.size.mddeleted file mode 100644index 986ac766..00000000--- a/changes/983-flight.size.md+++ /dev/null@@ -1,3 +0,0 @@-- entry:vehicle: 735 bytes gzip for `detachedPartFlight`, the seeded launch, tumble and ground bounce of a part a crash tears off, which the addon did not have (#983).-- goal:agent-scene: 772 bytes gzip for the same function; the lead accepted the +735 B gzip growth because it is a visible new capability.-- scene:agent-scene+lazy: 772 bytes gzip for the same function inside the lazy `vehicle-damage` module, which loads only for a recipe with a `dented` geometry.diff --git a/changes/983.md b/changes/983.mddeleted file mode 100644index f9d7ecdb..00000000--- a/changes/983.md+++ /dev/null@@ -1 +0,0 @@-- `crucible/addons/vehicle` dents vehicle parts on impact, finds the moment a drive meets a wall and lists the parts that tear off, and recipes get a `dented` geometry (#983)diff --git a/changes/983.size.md b/changes/983.size.mddeleted file mode 100644index c05cf5b3..00000000--- a/changes/983.size.md+++ /dev/null@@ -1,5 +0,0 @@-- entry:vehicle: 1987 bytes gzip for dents, the crash impulse, the drive-into-wall search and loose-part thresholds, a first slice of vehicle crash damage that the addon did not have (#983).-- scene:agent-scene: 250 bytes gzip for the `dented` geometry type in the core (its field list, two hooks and a capabilities line); the code itself is a lazy module.-- scene:agent-scene+lazy: 3197 bytes gzip for the lazy `vehicle-damage` module, which loads only for a recipe with a `dented` geometry.-- goal:agent-scene: 250 bytes gzip for the same `dented` hooks, since a recipe needs them in the core to load the lazy module.-- tool-total: 18 bytes (73510 to 73528), tool:scene_apply 9 bytes, tool:scene_create 9 bytes and result:scene_capabilities 16 bytes, all within their allowance, for the `dented` geometry type in the recipe schema and the vehicleDamage capabilities line.diff --git a/docs/game.md b/docs/game.mdindex c3a79558..453e389e 100644--- a/docs/game.md+++ b/docs/game.md@@ -34,7 +34,7 @@ A scene recipe can carry a `game` section. It names the node the player controls## Fields-A field newer than the last npm release carries a tag such as `{since 0.21.0}`; an older server refuses it with `is not a field`. Untagged fields work in the last release.+A field newer than the last npm release carries a tag such as `{since 0.22.0}`; an older server refuses it with `is not a field`. Untagged fields work in the last release.| Field | Default | Meaning || --- | --- | --- |diff --git a/docs/mcp.md b/docs/mcp.mdindex 9c445eb9..b5e10a73 100644--- a/docs/mcp.md+++ b/docs/mcp.md@@ -10,7 +10,7 @@ Release `0.7.0` adds [procedural motion](agent-motion.md), transient preview seeRelease `0.8.0` adds [imported character motion](agent-character-motion.md) with source joints and independent clip timing.Release `0.9.0` adds [visual review](agent-review.md) for several labeled views or poses from one isolated snapshot.-This guide follows `main`, and the npm release can be older. A field, tool or behavior that the last release does not have carries a tag such as `{since 0.21.0}`: it needs that release or a later one, and an older server refuses it by name. `npx -y -p @codepawlhq/crucible crucible-mcp --version` prints the installed version. `npm run check:guide-fields` fails when this guide or the [game guide](game.md) names a field that is newer than the last release without the tag.+This guide follows `main`, and the npm release can be older. A field, tool or behavior that the last release does not have carries a tag such as `{since 0.22.0}`: it needs that release or a later one, and an older server refuses it by name. `npx -y -p @codepawlhq/crucible crucible-mcp --version` prints the installed version. `npm run check:guide-fields` fails when this guide or the [game guide](game.md) names a field that is newer than the last release without the tag.To see what changed, including changes not yet released, read `https://crucible.codepawl.com/changelog.json`: `released` sections, `unreleased` entries and each entry's issue number and `breaking` flag.@@ -70,7 +70,7 @@ A Draco-compressed GLB (`KHR_draco_mesh_compression`, common in Sketchfab and NAA GLB with KTX2 textures (`KHR_texture_basisu`, what gltf-transform `optimize` and Sketchfab's optimised exports write) loads as it is too: the server transcodes each KTX2 image to a PNG (base mip level only, at most 4096 x 4096 pixels per image and 64 million pixels per file) with the Basis transcoder that the optional `three` package ships, and the preview page receives a plain GLB. Install it where the MCP server runs (`npm install three`). Without it, `scene_asset_catalog` lists the file under `skipped` with that advice, or re-export the textures as PNG or JPEG. The 64 MiB limit applies to the transcoded file.A GLB that points at image files (a Kenney kit ships `boat.glb` plus `Textures/colormap.png`) loads as downloaded: the catalog reads each relative PNG, JPEG or WebP path from that folder or below it, counts it against the 64 MiB file limit and stores it inside the asset, so the asset ID covers the texture bytes. A remote or `data:` address, a path that leaves the folder (`..`, an absolute path, a link) and a missing or non-image file are refused with an error that names the GLB and the path. Buffers must still be embedded. Poly Haven and similar sites ship a glTF folder (`.gltf`, `.bin` and textures): bundle it first with one command (the `.bin` must be embedded; image files may stay beside the GLB), then put the result in the catalog directory: `npx @gltf-transform/cli copy tree_01_1k.gltf tree_01.glb`. A recipe places a catalog asset with an `asset` node field (see [the scene guide](agent-scene.md#placed-catalog-assets)).-The asset ID is the SHA-256 of the file, so it changes on every re-export. A placement can name the file instead: `asset: { "name": "sheep_body.glb" }` in `scene_create` and in a `scene_apply` add-node, or `assetName` {since 0.21.0} beside `format` in the `scene_create_asset` recipe. Pass exactly one of `id` and `name`. The server resolves the name against the catalog when the call arrives and keeps only the ID, so the scene, its checkpoints and its render jobs stay pinned to the exact file. The name is the one `scene_asset_catalog` lists (`pack/file.glb` for a pack file) and must match exactly. An unknown name is refused with the nearest names, and two GLBs under one name are refused with a request to pass the ID.+The asset ID is the SHA-256 of the file, so it changes on every re-export. A placement can name the file instead: `asset: { "name": "sheep_body.glb" }` in `scene_create` and in a `scene_apply` add-node, or `assetName` beside `format` in the `scene_create_asset` recipe. Pass exactly one of `id` and `name`. The server resolves the name against the catalog when the call arrives and keeps only the ID, so the scene, its checkpoints and its render jobs stay pinned to the exact file. The name is the one `scene_asset_catalog` lists (`pack/file.glb` for a pack file) and must match exactly. An unknown name is refused with the nearest names, and two GLBs under one name are refused with a request to pass the ID.`--jobs` selects the absolute folder that holds render jobs and exports (`scene_export_glb`, `scene_export_web`). The server checks it at startup. A directory link, such as a Windows junction made with `mklink /J` to put outputs on another drive, is followed once at startup, and the job store then works on and stays inside the real folder it points at; a path that is a file stops the server with a message naming `--jobs`. Links inside the folder are still never followed, and no tool call can name a path. `scene_capabilities` works without a job store: if the store cannot open it lists no records and reports `jobs.unavailable` with the reason, and a later call tries the store again.@@ -156,14 +156,14 @@ Inspect or export the recipe before retrying. Context loss still permits recipe| `mesh_edit` | `sessionId`, `expectedRevision`, `geometry`, `edits`, optional `replace` | Updated revision, render status and the mesh report with per-edit results || `mesh_stack` | `sessionId`, `expectedRevision`, `geometry`, optional `stack`, `apply` | Updated revision, render status and the mesh report with per-node results || `scene_look` | `sessionId`, optional `width`, `height`, `motionTime`, `labels` | PNG image and metadata |-| `scene_ask_user` | `sessionId`, optional `prompt`, `maxMarks`, `timeoutSeconds`, `reviewFile` {since 0.21.0} | Opens the view in the browser, waits for Send, returns numbered marks with `partId`, `point`, `normal`, `note`, the camera and the marked PNG. With `reviewFile` it reads the marks a [review page](#review-link) saved instead (no browser, no image) |+| `scene_ask_user` | `sessionId`, optional `prompt`, `maxMarks`, `timeoutSeconds`, `reviewFile` | Opens the view in the browser, waits for Send, returns numbered marks with `partId`, `point`, `normal`, `note`, the camera and the marked PNG. With `reviewFile` it reads the marks a [review page](#review-link) saved instead (no browser, no image) || `scene_review` | `sessionId`, `expectedRevision`, `shots` (each with an optional `branch`), optional `width`, `height`, `labels`, `response_format` | Labeled PNG grid and metadata from an isolated snapshot || `scene_checkpoint` | `sessionId` | Checkpoint ID, session ID, and revision || `scene_restore` | `sessionId`, `expectedRevision`, and a `checkpointId`, a `revision` or a `branch` | Updated revision and render status || `scene_release_checkpoint` | `sessionId`, `checkpointId` | Release confirmation || `scene_export_recipe` | `sessionId` | Canonical recipe and revision || `scene_export_glb` | `sessionId`, optional `clip`, `motionTime`, `maxTextureSize` | A GLB file under `<jobs>/exports/` with a server-chosen name: path, bytes, SHA-256. See [GLB export](export-glb.md) |-| `scene_export_web` | `sessionId`, optional `title`, `review` {since 0.21.0} | A static web build folder under `<jobs>/exports/`: path of `index.html`, file count, bytes. Needs a `game` set with `set-game`, unless `review` is true. See [Game](game.md) and [Review link](#review-link) |+| `scene_export_web` | `sessionId`, optional `title`, `review` | A static web build folder under `<jobs>/exports/`: path of `index.html`, file count, bytes. Needs a `game` set with `set-game`, unless `review` is true. See [Game](game.md) and [Review link](#review-link) || `scene_dispose` | `sessionId` | Disposal confirmation || `scene_render_submit` | `sessionId`, `expectedRevision`, `settings` | Snapshot job ID and queued record || `scene_render_status` | `jobId`, optional `response_format` | State and durable frame progress |@@ -181,7 +181,7 @@ Each scene keeps up to 16 checkpoints. `scene_checkpoint` returns `checkpointId`A person who is not at the agent's keyboard can mark a scene on a phone. `scene_export_web` with `review: true` needs no game: it renders the current view and writes `index.html` and `frame.png` under `<jobs>/exports/`, in the same folder and with the same newest-four retention as game builds. Host the folder on any static server or send it as is. The reviewer taps the picture to drop a pin (up to 16), writes a note for each, and presses Save file or Copy. Two fingers zoom and move the picture and never drop a pin. The page is a still, so it does not orbit.-Pass the text of the saved file to `scene_ask_user` as `reviewFile` {since 0.21.0} (at most 16 KiB). The answer has the same fields as a local `scene_ask_user` call (`marks` with `partId`, `point`, `normal`, `note`, and `view`), without the marked PNG. The server keeps the camera of the last 8 review exports in memory, so the file carries only the marks and cannot choose the camera the server casts their rays from. Each export is read once: a second read of the same file, a file from an earlier server run, from an export dropped after 8 newer ones or disposed with its scene, or for another scene is refused with a message that says to export again. A review file proves nothing about who made it: the review ID is written into the exported page, so anyone who can read that folder, a model with file access included, can write a valid file. Unlike the inline view's nonce, which the model never sees, treat a review file's marks as notes from whoever had the page.+Pass the text of the saved file to `scene_ask_user` as `reviewFile` (at most 16 KiB). The answer has the same fields as a local `scene_ask_user` call (`marks` with `partId`, `point`, `normal`, `note`, and `view`), without the marked PNG. The server keeps the camera of the last 8 review exports in memory, so the file carries only the marks and cannot choose the camera the server casts their rays from. Each export is read once: a second read of the same file, a file from an earlier server run, from an export dropped after 8 newer ones or disposed with its scene, or for another scene is refused with a message that says to export again. A review file proves nothing about who made it: the review ID is written into the exported page, so anyone who can read that folder, a model with file access included, can write a valid file. Unlike the inline view's nonce, which the model never sees, treat a review file's marks as notes from whoever had the page.### Tool errors@@ -193,7 +193,7 @@ Every tool schema is read on every turn, so `node scripts/measure-tool-cost.mjs``mesh_inspect` lint adds warnings for `intersecting-parts`, `self-intersection`, `thin-wall`, `floating-part` and `asymmetric`. They never set `ok` to false, run within face budgets, and list any check they skipped in `skippedChecks`. `scene_inspect` returns a `diagnostics` list only when there is something to say: ignored material fields, blended transmission, unused materials or geometries, and for imported assets the decoder's not-drawn features, rigs with no joints or no skin, and assets with no clips. Each item has `kind`, `severity`, `target`, `message` and `fix`.-`scene_look` keeps the PNG under the 1 MiB response budget by itself {since 0.21.0}: when a busy picture (planks, bricks, emissive text) encodes over it, the preview renders again at 0.75 of the size, up to four times with the aspect kept, and the metadata names the size that came back (`returnedWidth`, `returnedHeight`, `note`). It fails only when the smallest step still does not fit.+`scene_look` keeps the PNG under the 1 MiB response budget by itself: when a busy picture (planks, bricks, emissive text) encodes over it, the preview renders again at 0.75 of the size, up to four times with the aspect kept, and the metadata names the size that came back (`returnedWidth`, `returnedHeight`, `note`). It fails only when the smallest step still does not fit.`scene_look` returns the same kind of `diagnostics` list next to the picture, so an agent can check a frame without reading it. The findings carry the part names and the numbers in fields, and each list is capped (8 parts per finding, 8 findings per kind, with `nodeCount` or `droppedCount` saying how many were left out):@@ -363,7 +363,7 @@ The script prints calls, errors, error rate and median duration per tool, most c| Unique geometry buffers | 32 MiB || Recipe or command JSON | 1 MiB || Incoming stdio buffer | 2 MiB |-| Preview dimensions | 64 to 1024 pixels per dimension (image jobs: up to 3840x2160, stills and video up to 1920x1080; `scene_capabilities` jobs lists both as `maximumWidth`, `maximumHeight`, `maximumPixels` and `maximumSequenceWidth`, `maximumSequenceHeight`, `maximumSequencePixels`; `maximumSquareSide` and `maximumSquareSequenceSide` give the largest square frame; a stills job with `stillCount` {since 0.21.0} writes up to `maximumStillCount` numbered PNGs, see [agent-render-jobs](agent-render-jobs.md)) |+| Preview dimensions | 64 to 1024 pixels per dimension (image jobs: up to 3840x2160, stills and video up to 1920x1080; `scene_capabilities` jobs lists both as `maximumWidth`, `maximumHeight`, `maximumPixels` and `maximumSequenceWidth`, `maximumSequenceHeight`, `maximumSequencePixels`; `maximumSquareSide` and `maximumSquareSequenceSide` give the largest square frame; a stills job with `stillCount` writes up to `maximumStillCount` numbered PNGs, see [agent-render-jobs](agent-render-jobs.md)) || Default preview dimensions | 768 by 512 pixels || PNG payload | 1 MiB |diff --git a/package.json b/package.jsonindex 7e846570..b9ee49bd 100644--- a/package.json+++ b/package.json@@ -1,6 +1,6 @@{"name": "@codepawlhq/crucible",- "version": "0.20.0",+ "version": "0.21.0","private": false,"type": "module","description": "3D engine that AI agents drive: a WebGL2 rasterizer, a path tracer on CPU and WebGL2 (Vulkan with an optional package, WebGPU opt-in), an MCP server, and PNG, MP4 and GLB export",diff --git a/src/version.js b/src/version.jsindex a90187ea..d35fd482 100644--- a/src/version.js+++ b/src/version.js@@ -1,2 +1,2 @@// Kept as a literal so bundles do not carry all of package.json. tests/version.test.js keeps it in step.-export const VERSION = '0.20.0';+export const VERSION = '0.21.0';