0.10.0

A quality release: the deprecated names are gone, and a long list of bugs with them. No new elements; what was added is tooling for measuring and debugging on a device, and some defaults changed (listed under Changed defaults below). Most of what breaks was announced in 0.8 and 0.9; three things were not, and are listed below with the rest: new MersenneTwister() without a seed, WorldStore queries returning copies, and surface.openPanel's onClose. Migration.md has every pair and what to do.

Size: barrel (min, packages external) 970.3 KB raw / 327.2 KB gz, +2.30% on 0.9.0.

⚠️ Breaking

• onX callback options are removed. onChange, onClick, onSelect, onActivate, onClose, onKey, onAction, onCaretMove, onEnter and onFocus are no longer called; use handleX. An old name is not dropped silently: handlerOf warns once per name that it is ignored and names the replacement.
• <tosi-b3d-trigger>'s onEnter / onExit properties are removed; use whenEnter / whenExit. <tosi-b3d-manipulator>'s handleChange / handleCommit are removed; use whenChange / whenCommit.
• generateGalaxy is removed; use voxelGalaxy({ seed, brightBudget }).view(). galaxy.getStarSPS() is removed; use pickStar.
• new MersenneTwister() needs a seed. It used to seed from the clock when the argument was left out, which silently broke "same seed, same result". It now throws; pass Date.now() if that is what you want.
• WorldStore.getState(), getEntity() and query() return copies. A driver could change the simulation by writing to what a query returned. Code that relied on that (deliberately or not) must call a command instead. The engine's own systems read store.liveState. A copy is made on every call, so read once per frame, not once per entity.
• A store passed to the conformance kit must implement chooseOption(choiceId, optionId). The kit now covers the choice primitive and calls it.
• surface.openPanel({ onClose }) is handleClose. This one was never announced: it was the last onX option in the library and follows the same rule now (ignored, with a warning).

Changed defaults

Nothing here removes an API, but a scene can look or behave differently without a line of yours changing. Each has the setting that puts it back.

• Opaque meshes are drawn nearest first, the sky last (was material order). The picture is the same for ordinary opaque meshes; one with depth write off can differ. scene.setRenderingOrder(0) restores Babylon's order, and a camera-pinned sky of your own must set infiniteDistance.
• A headset session renders at a reduced scale with foveation on the low tier (xrRenderScale, xrFoveation; 1 and 0 for full).
• The decorator places a tier share of its budget (0.6 on the low tier) and plant rules thin with moisture (thirst), so the same seed gives a different layout from 0.9. budgetScale: 1 places the whole budget.
• Scatter rules gather more (clump 0.95, was 0.85).
• Water receives sun and cloud shadows.
• The low tier uses the lighter ground shader (groundDetail).
• PerlinNoise(0) is a seed, not "no seed".

Docs

• The doc site is reorganised. Sections now list the elements you put in a scene, with the pure models and helpers each is built from tucked beneath it (the flight model under the aircraft, the biome chart under the terrain). UI was one list of 39 pages and Environment one of 37; no section now has more than 12 direct children. Environment is split into Sky and light, Weather and clouds, Terrain and Water; Migration and the Changelog moved from second and third to a Reference section at the end. Page URLs are unchanged, except the section landing pages.
• Getting started: the install line with its peer dependencies, a first scene that needs no model file, and where to go next. The README has the same.
• Ten modules that other pages linked to had no page: the flight model, ballistics, guidance, warhead, destroyable, resource, Perlin noise, gradient filter, terrain grid and the primitives (box, sphere, ground). 17 links were dead. A test now fails on a link to a page that does not exist.
• Every section landing page says which page you want, where most were a bare list of links.

Fixed

• A headset is drawn at the device tier's resolution. It never was: the tier lowered the engine's hardware scaling for a session, and that does not reach the XR layer, so every tier drew both eyes at full size. The viewport is now scaled per frame (xrRenderScale, auto: 0.67 on the low tier, 0.83 on medium) and fixed foveation is set (xrFoveation, auto: 1 on low, 0.5 on medium). Measured on a Quest in Land and Sky at 90 ms a frame: half the pixels saved 19 ms and full foveation 12 ms. Both are attributes of <tosi-b3d>; set xrRenderScale="1" xrFoveation="0" for the old picture.
• Opaque meshes are drawn nearest first, and the sky last. Babylon draws them in creation order, so a sky made first was shaded across the whole view and then painted over by the land. The picture is the same; a covered pixel is now rejected before its shader runs. nearFirst(scene) is exported for a render target of your own (the water's two have it).
• A cheaper ground shader for a weak GPU (<tosi-b3d groundDetail>: auto, full, lite). The biome shader took eight noise samples a pixel; lite takes three and looks nearly the same from a distance. auto picks it on the low tier, and in a headset on the medium tier. Measured on a Quest in Land and Sky, the terrain was 40 ms of a 90 ms frame. The cost sweep measures the other form on your device.
• The ground shader reads its fine noise from a texture. It computed 3D noise four times a pixel, eight hashed corners each; it now reads a shared 64-texel tiling block (noiseVolume, exported) once per sample. Mip-mapped, so the detail fades in the distance where the computed noise shimmered. On every tier with WebGL 2; setBiomeComputedNoise(scene, true) is the way back, and the cost sweep measures it. The climate noise is still computed.
• The water's reflection and refraction are drawn once for both eyes, and every Nth frame in a headset too. Each is the scene drawn again, and a headset drew both for each eye on every frame: four extra scene passes. Each is now looked up through the view it was drawn from, so one drawing is in the right place from either eye and on later frames (reflectionRefresh, from the device tier: every 3rd frame on the low tier). This also fixes the refraction sliding against the seabed between redraws on a flat screen.
• Perf Stats reports what SVG texture repaints cost (the HUD, panels): repaints a second and main-thread milliseconds a second (svgTextureCost). It also shows the JS heap against its first reading and the total redraw count, so a leak is visible from inside a headset.
• Debug sink (debugSink.url, sendDebug): a page can post its readings to an endpoint, for a headset that has no console. With a URL set, <tosi-b3d> sends each cost sweep step, the Perf Stats lines while that popup is open in a session, and every logDebug event. Off by default.
• bun run xr … --counters [--again n]: documents, DOM nodes, listeners and JS heap before and after a script, after a collection.
• Cost sweep (b3d.costSweep(), and a button in Perf Stats): where a frame goes, measured on the device, by switching one thing off at a time. Runs are numbered and the last four are kept, each saying whether it finished or where it stopped, so a run that hangs does not replace the one before. Shader sweep is a separate button: it measures the ground shader's other forms, which recompiles shaders.
• A popup's move and close icons keep their size on a tall popup in a headset. The title bar was a fifth of the popup's height, so Perf Stats with a long list got icons several times too large, over its first rows.
• A popup is capped in height and scrolls past the cap (480 units, or the caller's maxHeight). Perf Stats with a sweep listed was taller than the screen flat and enormous in a headset, with no way to its lower rows.
• The decorator places a share of its budget set by the device tier (budgetScale, 0 = auto: 1, and 0.6 on the low tier, which is a Quest in a session). Measured on a Quest: decorations were 16 ms of a 42 ms frame. Set budgetScale: 1 to place the whole budget everywhere.
• A thinned scatter loses clumps, not density (ScatterOptions.share, clumpCentreFor): the weaker clumps go and the middle of the rest stays as full, in the same places as the whole scatter. The decorator's share and the moisture thinning both work this way. Rules also gather more by default (clump 0.95, was 0.85), so fewer things stand alone.
• The cost sweep's two headset rows measure what the resolution levers are saving (+ full resolution, + no foveation). The old rows set the values the tier already uses, so they measured nothing.
• Wetter country grows more. A scatter rule's moisture band only said where a kind may grow, and every plant's band reached the top, so anything damp enough for a tree grew the same forest as the wettest setting. Plant rules now have a thirst: the drier it is, the more of their places are left empty (moistureCover). In Land and Sky, trees go from about 120 at moisture 0.5 to about 240 at 1; rocks are unchanged.
• A headset session no longer freezes the second time your hands are tracked. Babylon's hand tracking (9.16) takes the topmost ancestor of its shared hand mesh as a root and parents it to the camera's parent every frame. The second time hands appear (a controller put down twice), that root is the camera's parent itself, which became its own parent; the next enable or dispose then recursed until the stack ran out and the picture stopped. Found on a Quest with Perf Stats open (#3135). The rig now refuses such a parent (refuseOwnAncestor, exported for anyone parenting an XR camera to their own node).
• Underwater light rays go away when you leave the water. They were still fading on the frame you surfaced, the update read the water that was no longer above you and threw, and it threw again every frame after, so the rays stayed for good.
• A popup moved in a headset stays with its panel. Dragging an owned popup used to tear it off, so moving the panel afterwards left it behind. It now moves relative to its opener and stays owned; tearOff() is still how a caller lets one go.
• No second, dead copy of a popup after a panel rebuild in a headset. The rebuild dropped its popups' handles and relied on each dying with the panel plane. One that was torn off, or not yet mounted, stayed in the scene: a Perf Stats that never updated and could not be closed. They are now closed before the plane goes, and a popup whose opener has been disposed closes itself.
• Perf Stats refreshes every two seconds in a headset (was 0.4 s). A Quest hung within minutes with it open and ran on with it closed; the cause is not known, this redraws its texture less often.
• Water, and the ice on it, receives shadows. The surface never did: it was not marked as a shadow receiver, so nothing cast a shadow on the sea, which on a white ice sheet is obvious. <tosi-b3d-water receiveShadows="off"> restores the old behaviour and saves a shadow lookup per water pixel (the cost on a headset has not been measured).
• Cloud shadows fall on water and ice. They never did, on any device: cloud shadows are a material plugin, and Babylon's water material runs no plugins. The lookup is now in the water's own shader, fed from the scene's cloud shadow map (cloudShadowMapOf(scene), new). receiveShadows="off" turns it off with the sun's shadows. Every water now uses the patched shader, not only shore="on".
• PerlinNoise(0) uses seed 0. It was treated as "no seed" and gave a different field every time.
• world-store.ts compiles under noUncheckedIndexedAccess (the steer loop read r.portals[0] unguarded), so a project that vendors it with strict settings no longer has to pin an old copy.
• The world contract now says what it meant: a driver must supply spec.id to spawn (the return value cannot cross a worker boundary), and setIntent is legacy (steer is what the sim honours).
• The conformance kit covers the choice primitive: a pick emits exactly one choiceMade, an option that is not on the menu emits nothing, a choice is reported once, and the sim resolves nothing.
• A bomb that goes off on terrain now hurts what is above it. The blast's line of sight started on (or just under) the surface it hit, so the terrain was cover for everything and a ground burst damaged nothing. A burst near a surface now looks from a metre above it; a hill or wall in between is still cover. (Only a mesh named ground was exempt before.)
• Decorator: naming the default library explicitly no longer empties the scatter. Default rules were chosen by whether url was set, so url="…/quaternius/libraries/nature.glb" got the other library's rules, which matched no model. The choice is now made from the library in use.
• Decorator: NATURE_ROLES applies only to the default library. Its keys are generic material names (White, Wood), so a third-party model with a material of the same name had it replaced by flat biome shading. Set roles to opt a custom library in.
• The Perf Stats "Occlusion" button says whether occlusion is running, not what was asked for (on, not running here when the method cannot run in the current mode). Land and Sky's own occlusion select, which drifted out of step with it, is gone.
• Lightning bolts are solid, not see-through. The channel was faded by transparency between re-strokes, so up close it was a pale translucent tube. It is opaque while the stroke is live, far brighter than white (so it blooms), and the glow around it fades to nothing at its edge. The flicker is carried by the glow, the cloud and the ground.
• Thunder stops when its storm does. Thunder is queued as far ahead as sound takes to arrive, so changing world left the old storm's thunder playing over the new one.
• In VR, the pointer no longer lands on the vehicle you are sitting in. In the aircraft's cockpit the controller ray hit the canopy first, and on a Quest the panel behind it could be hovered but not pressed or dragged, so there was no way out of the session from inside. Anything beneath the node the XR rig rides on is no longer a pointer target; your own panels are. (Reproduced and checked with the emulated headset; not yet on a device.)
• Sea ice: the plate pattern no longer depends on the GPU's sin. The plates were laid out with the classic fract(sin(x) * 43758.5) hash, which mobile GPUs evaluate differently from desktop ones; in a headset the ice showed severe mismatches that a Mac never did. The hash is arithmetic only now. Plate edges are also a little softer. (The layout of plates for a given place changes. Not yet confirmed on a headset.)
• No holes in solid sea ice near its edge. The invisible mesh a car or a shell lands on kept a triangle only if all three corners were solid, so a band one cell wide (16 m and more far from the viewer) was drawn solid and had nothing to stand on. A triangle is kept when it is solid on average, which is where the drawn ice crosses the same line.
• The HUD page no longer throws 30 times a second. Its demo still called setTraces(traces, viewer, options), a signature that went away when the HUD started projecting through the camera. The demo passes the camera, and setTraces warns once (and draws nothing) if it is handed anything else.
• Three demo controls were dead because they still used an old name the compiler cannot see in a doc comment: the aircraft page's "chase pitch follow" and "follow lag" sliders, and the particles page's Burst button. (They worked, with a warning, until the fallback was removed here.)