@meterup/pascal-viewer@1.0.0-beta.6

Minor Changes

• d3aa350: Let a host bind the 3D camera's gestures.

  The bindings were fixed: left drag selects unless you're in preview mode or
  holding space, middle pans, right rotates, wheel zooms. A host whose scene is
  read-only, or whose users are on laptop trackpads with no middle button, had no
  say — panning meant the space bar.

  cameraInput on useViewer merges over the defaults rather than replacing
  them, so an unset entry keeps whatever the current mode asked for:

  useViewer.getState().setCameraInput({
    mouseButtons: { left: "pan" },
    touches: { one: "pan" },
  });

  Actions are named for what the user sees (pan, rotate, zoom, dolly,
  and zoomPan / dollyPan for the combined two-finger gestures) rather than
  mirroring the camera library's enum, so the binding API doesn't tie a host to
  that dependency. The wheel and the one-finger slots take narrower sets than the
  others, because a wheel has one axis and a pinch needs two fingers.

  Merging also covers the modifier path, which rebuilds the bindings on every key
  change, so holding space no longer drops a host's overrides.

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

• e1d3a5a: Add showSiteBoundary to hide the site's boundary geometry.

  The boundary was always drawn, with no way to turn it off. It is a useful
  reference while laying out a site and pure noise on a surface that only presents
  a finished building.

  showSiteBoundary joins useViewer next to showGrid / showZones and
  governs both views from one flag, because it is the same polygon in each:

  useViewer.getState().setShowSiteBoundary(false);

  In 2D that drops the dashed polygon, its edge labels and its handles. In 3D it
  drops the ground pad and the amber outline. Buildings, items and the horizon
  ground are unaffected, so the scene loses its lot outline rather than its floor
  — the horizon disc punches holes for slab footprints, not for the site polygon,
  so it still covers the ground the pad was covering.

  Two things stay deliberately outside the flag. The fitted viewport still sizes
  itself from the site polygon, so toggling this doesn't reframe the plan. And
  sculpted terrain keeps rendering, since that is modelled geometry rather than
  boundary chrome.

  Defaults to true and isn't persisted, so nothing changes until a host sets it.

Patch Changes

• 54e1b19: Let a scene theme set the ground grid's line colours and idle opacity.

  appearance hardcoded both colours, and in dark mode <Grid> also ignored its
  own cellColor / sectionColor props, so neither a theme nor a host could
  adjust how the grid read. That matters because the grid is drawn over the lit
  ground fill: its contrast against that surface is a property of the pair, and
  a light/dark flag can't settle it. ground and backgroundSky were already
  carved out of appearance for the same reason.

  const theme = SCENE_THEMES.find((t) => t.id === "night");
  theme.grid = { cell: "#2b2d3a", section: "#343747", idleOpacity: 0.18 };

  All three keys are optional and fall back to the previous appearance-derived
  values, so existing themes render identically. idleOpacity applies to the
  always-on ground reference only; an active placement patch keeps its own
  brightened treatment.

@meterup/pascal-nodes@1.0.0-beta.6

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

Patch Changes

• e1d3a5a: Add showSiteBoundary to hide the site's boundary geometry.

  The boundary was always drawn, with no way to turn it off. It is a useful
  reference while laying out a site and pure noise on a surface that only presents
  a finished building.

  showSiteBoundary joins useViewer next to showGrid / showZones and
  governs both views from one flag, because it is the same polygon in each:

  useViewer.getState().setShowSiteBoundary(false);

  In 2D that drops the dashed polygon, its edge labels and its handles. In 3D it
  drops the ground pad and the amber outline. Buildings, items and the horizon
  ground are unaffected, so the scene loses its lot outline rather than its floor
  — the horizon disc punches holes for slab footprints, not for the site polygon,
  so it still covers the ground the pad was covering.

  Two things stay deliberately outside the flag. The fitted viewport still sizes
  itself from the site polygon, so toggling this doesn't reframe the plan. And
  sculpted terrain keeps rendering, since that is modelled geometry rather than
  boundary chrome.

  Defaults to true and isn't persisted, so nothing changes until a host sets it.

@meterup/pascal-mcp@1.0.0-beta.7

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

All notable changes to @meterup/pascal-mcp will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

@meterup/pascal-ifc-converter@1.0.0-beta.6

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

Patch Changes

• Updated dependencies [2251db5]
• Updated dependencies [2f5c4e1]
• Updated dependencies [5ea6e96]
  • @meterup/pascal-core@1.0.0-beta.6

@meterup/pascal-editor@1.0.0-beta.6

Minor Changes

• d3aa350: Let a host bind the 3D camera's gestures.

  The bindings were fixed: left drag selects unless you're in preview mode or
  holding space, middle pans, right rotates, wheel zooms. A host whose scene is
  read-only, or whose users are on laptop trackpads with no middle button, had no
  say — panning meant the space bar.

  cameraInput on useViewer merges over the defaults rather than replacing
  them, so an unset entry keeps whatever the current mode asked for:

  useViewer.getState().setCameraInput({
    mouseButtons: { left: "pan" },
    touches: { one: "pan" },
  });

  Actions are named for what the user sees (pan, rotate, zoom, dolly,
  and zoomPan / dollyPan for the combined two-finger gestures) rather than
  mirroring the camera library's enum, so the binding API doesn't tie a host to
  that dependency. The wheel and the one-finger slots take narrower sets than the
  others, because a wheel has one axis and a pinch needs two fingers.

  Merging also covers the modifier path, which rebuilds the bindings on every key
  change, so holding space no longer drops a host's overrides.

• 068468e: Add onHeadingChange to the floor-plan compass slot.

  needleRef is the only way a host gets the per-frame heading, and it's typed
  RefObject<SVGSVGElement> for the element that visually rotates. That works for
  a host whose rotating element is an <svg> it owns, and not at all for one
  composing an existing icon component, which has nowhere to put the ref and
  falls back to re-rendering per commit.

  onHeadingChange carries the same values as a subscription:

  floorplanCompassSlot={({ northRotationDeg, alignToNorth, onHeadingChange }) => {
    const heading = useMotionValue(northRotationDeg)
    useEffect(() => onHeadingChange((deg) => heading.set(deg)), [onHeadingChange])
    return <motion.div style={{ rotate: heading }} onClick={alignToNorth}>{glyph}</motion.div>
  }}

  It fires immediately with the current heading, because each source only writes
  on its own updates and a still view sends nothing, so a subscriber that waited
  would show north until the first movement.

  Deliberately a callback rather than an animation library's value type, so the
  slot contract stays library-agnostic and a host can drive a plain style write
  instead if it prefers. needleRef is unchanged.

• 9a4878b: Add floorplanDividerSlot for host-owned split-view dividers.

  The divider's pill was a fixed neutral capsule, so a host whose own resize
  handles look nothing like it had no way to reconcile the two, and nowhere to
  hang a control that belongs on the boundary between the panes.

  The slot replaces the pill and receives the handler that starts the pane drag,
  so a host can put it on whatever the user actually grabs:

  <Editor
    floorplanDividerSlot={({ startResize }) => (
      <MyControlGroup>
        <MyToggle onClick={(event) => event.stopPropagation()} />
        <MyGrabHandle onPointerDown={startResize} />
      </MyControlGroup>
    )}
  />

  The divider itself still starts a drag on pointer down, so controls that aren't
  the grab handle need to stop propagation or clicking them resizes the panes.

  Only rendered in split view. Omitting the slot keeps the built-in pill, which
  also now carries data-pascal-floorplan-divider-thumb for hosts that only want
  to restyle it.

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

• 068468e: Add useFloorplanNavigationControls for host-driven 2D navigation.

  A host could already move the plan with publishNavigationSyncPose, but only by
  knowing rather a lot: that a pose is { target, azimuth, viewWidth }, that
  'host' is the one source that always reaches the plan, that azimuth is
  radians, that screen deltas have to be rotated into the plan by its own azimuth
  before they can move the target, and that animating means publishing a stream
  rather than a single pose. The demo in apps/editor did all of that by hand.

  The hook does it instead:

  const plan = useFloorplanNavigationControls()
  
  useGesture({
    onDrag: ({ delta: [dx, dy] }) => plan.panByPixels(-dx, -dy),
    onPinch: ({ offset: [scale] }) => plan.zoomTo(scale),
  })
  
  <button onClick={() => plan.rotateByDegrees(90, { durationMs: 260 })}>Rotate</button>

  Rotation tweens take the short way round, so turning from 350° to 10° travels
  20° forwards rather than 340° back, and an in-flight tween is cancelled by the
  next command or by unmount.

  Whether a command moves the 3D camera too is the link's business, not the
  hook's: these publish as 'host', which always drives the plan, and the camera
  follows only while floorplanNavigationLink is on.

  The hook reads the shared pose at call time through getPose() and
  getMetersPerPixel() rather than subscribing to it. navigationSyncPose
  changes every frame while anything is moving, so a subscription would re-render
  the caller at that rate, which for a component that renders the editor is the
  whole cost the imperative viewport exists to avoid. The returned object has a
  stable identity, so it can go straight into an effect's dependencies.

  Supporting this, the floor-plan panel now publishes its measured size in CSS
  pixels. A pose carries only its width in metres, so until now nobody but the
  panel could convert one back to screen space, which is why the playground had
  to stand in window.innerWidth and why it was wrong in split view. The panel
  also republishes navigationSyncPose on every committed viewport change rather
  than seeding it once: a single seed goes stale as soon as the plan moves any
  other way, most obviously the initial fit, and a host reading that stale pose
  and publishing it back dragged the plan to wherever the seed was taken.

• 5fbeb64: Add hideBuiltInOverlays, and let hosts style the split-view divider.

  Hosts that bring their own toolbar had one way to get rid of the editor's
  floating overlays: hold capture mode on. That works, but capture mode is a
  snapshot mode. It also unmounts SelectionManager, which owns the highlight
  material and the outliner as well as click handling, so the canvas ends up
  unselectable and nothing highlights even when something is selected from the
  2D plan. Permanent capture mode is the wrong tool and this was the cost.

  hideBuiltInOverlays hides exactly the chrome and nothing else:

  <Editor hideBuiltInOverlays layoutVersion="v2" />

  That drops the level selector, the action menu, the panel manager, the helper
  manager, the built-in Plugins rail entry and the camera-controls hint, in both
  layout versions. The Plugins entry matters because useHostPanels appends it
  in the edit workspace, so a host passing sidebarTabs={[]} still got a
  non-empty tab bar and the v2 layout rendered a left column for a panel it never
  as...

@meterup/pascal-core@1.0.0-beta.6

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

Patch Changes

• 2251db5: Add an alignmentGuideStroke slot to FloorplanPalette and draw the 2D snap
  guides from it.

  FloorplanAlignmentGuideLayer hardcoded #ef4444, so the one piece of 2D
  chrome most visible during a drag was the one piece a theme couldn't reach.
  It already consumes the render context, so it now reads the slot and keeps the
  red as its no-provider fallback.

• 5ea6e96: Let plugins type their node kinds into AnyNode by declaration merging.

  Plugin kinds were absent from AnyNode, AnyNodeType and AnyNodeId, so
  plugin code and anything consuming it had to widen to string or cast. Ids
  made it worse: objectId produces `${kind}_${string}`, so a plugin node's
  id was never assignable to AnyNodeId and generic scene code reached for
  as AnyNodeId.

  Declare kinds in PluginNodes and they join all three unions, which restores
  exhaustive switches and drops the casts:

  declare module "@meterup/pascal-core/plugin-nodes" {
    interface PluginNodes {
      tree: z.infer<typeof TreeNode>;
    }
  }

  Purely type-level and additive. The interface ships empty, PluginNodeType
  resolves to never, and AnyNode stays exactly equal to the new BuiltinNode
  alias, so a host with no plugins sees no change at all. There is no runtime
  component and no API signature churn; nodeRegistry.get still takes string,
  which it has to, since setPluginDiscovery allows kinds that were never
  compiled against.

  Two design points worth knowing:

  The interface lives in its own module and is reachable only at
  @meterup/pascal-core/plugin-nodes. declare module merges only with the module
  that declares the interface, and the package entry is a pure barrel, so
  augmenting @meterup/pascal-core would quietly declare an unrelated second
  interface and widen nothing. Nothing else is exported from that module, so an
  augmentation has nothing to shadow.

  PluginNode adds the common BaseNode fields and takes type from the key
  rather than trusting the declared shape. AnyNode is a union, and a union only
  exposes properties present on every member, so a declaration that
  under-described itself would silently strip parentId and friends from
  AnyNode across the entire host, with the errors landing in host code far from
  the plugin responsible.

• Updated dependencies [2f5c4e1]

  • @meterup/pascal-capture-protocol@1.0.0-beta.5

@meterup/pascal-cli@1.0.0-beta.2

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

@meterup/pascal-capture-viewer@1.0.0-beta.5

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

@meterup/pascal-capture-protocol@1.0.0-beta.5

Minor Changes

• 2f5c4e1: Rebase the fork onto upstream pascalorg/editor@4c34032c (2026-08-24) — 478
  commits, and a move off the fork's own 0.9.x line onto upstream's 1.0.0-beta
  versioning. The fork now carries nothing but its release machinery and two
  editor patches; everything else is upstream.

  Nothing we previously exported was removed. The public API grew substantially
  (core 172 → 388 exports, viewer 106 → 193, editor 156 → 470) without breaking
  underneath consumers.

  What matters most for consumers:

  • Near-collinear wall junctions no longer explode.
    calculateJunctionIntersections now rejects a miter on its outcome, capped
    at MITER_LIMIT (10) half-thicknesses, rather than only rejecting walls that
    are exactly parallel. A junction where two walls run straight through each
    other but miss collinearity by a fraction of a degree used to put a plan
    vertex halfThickness / sin(θ) away — kilometres out, collapsing the 2D
    fit-to-bounds into one flat rectangle and firing beams across the 3D scene. It
    now butts. Imported floor plans with off-axis wall runs render correctly.
  • Level assembly is dramatically faster. findJunctions buckets walls into
    a spatial grid by AABB instead of testing every wall against every junction
    point. On a 1081-wall imported floor that is 584 ms → 11 ms per call, and
    WallSystem repeats the call every frame while progressively rebuilding.
  • Vertical building model. Scenes now carry explicit building and level
    structure (buildingId, schema/nodes/building.ts), with load-time
    migrations for saves that predate it — vertical, site-child, wall-slot,
    elevator, retired-floorplan and construction-dimension. Anything synthesizing
    a SceneGraph rather than loading a saved one needs to match the new shape;
    this is the single change most likely to require work on the consumer side.
  • New packages: capture-protocol, capture-viewer, cli. New core
    systems alongside wall/ and stair/: fence/, slab/, elevator/.
  • healSceneNodes repairs already-saved scenes on load, dropping
    degenerate nodes such as zero-length walls.

  Three fork patches are dropped here because upstream made them redundant:

  • The floorplan-registry-layer geometry cache. Upstream landed its own
    geometryCacheRef/levelDataCacheRef memoisation plus the findJunctions
    prefilter above, which together address the same stall from further up.
  • The ThreeElements/lineBasicNodeMaterial narrowing in viewer that let
    tsgo check the package. Upstream adopted the TypeScript 7 native compiler
    and landed the identical narrowing.
  • crossOrigin on the 2D guide image. It forced the 2D <image> to fetch in
    the same CORS mode as the 3D texture load so the two shared one browser cache
    entry instead of downloading the image twice. Consumers that hand Pascal
    same-origin blob URLs never hit the split in the first place, and the flag is
    actively wrong for guide images served without CORS headers.

@meterup/pascal-viewer@0.9.1

Published to GitHub Packages.