* fix(core): elect wall slab support from the carrying profile, not face coverage
An elevated deck drawn against a house wall covers the wall's outer face
line end-to-end (boundary contact counts), so the max-across-polylines
election handed the wall origin to the deck: every wall-hosted window/door
rode along whenever the deck height changed, and placement local-Y was
clamped above the deck top. Elect from the carrying profile instead (per
arc segment: highest support per face, min across supported faces), with
the pointer cap applied inside the profile so a capped-away deck still
falls back to the floor that carries the wall. Also pass curveOffset/
thickness/supportSlabId at the window/door tool query sites so their
cursor agrees with the rendered wall frame.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(core): honor live node overrides in spatial-grid support queries
Group drags publish translated slab polygons and wall endpoints to
useLiveNodeOverrides only; the committed spatial index made floor items
and walls re-elect support against the pre-drag slab footprint, so
multi-selection moves and room-preset placement jumped vertically until
the validating click committed the batch. Support queries now read
live-effective slab/wall records and bypass the rendered-polygon cache
while a slab or wall on the level has an override; the committed cached
path stays the fast path otherwise.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(editor): give the room-preset stamp tool a polygon snap context
The host app's room stamp drives placement with tool='room', which has no
registry entry, so snapContextOf resolved null: Shift never cycled the
snapping mode and the HUD chip stayed hidden during room preset placement.
A tool-level context map hands the stamp the no-angle polygon set.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(editor): only synthesize walkthrough fallback floors for ground levels
Every slab-less visible level got a >=30x30m opening-free fallback floor
box at its elevation, so a walkthrough spawned on an upper level (the
no-spawn-node fallback ray picks the highest surface) stood on a phantom
plane it could never descend from. Match the baked-GLB viewer policy:
only the lowest level of each building (derived baseY === 0) gets the
fallback; upper levels rely on their real slabs, whose stair openings
are cut into the geometry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(core): fold useLiveTransforms slab deltas into support queries
The slab move tool and the room-preset stamp publish a translation DELTA
to useLiveTransforms (no polygon override), so the spatial index still
elected support against the slab's committed footprint: furniture riding
a room-preset preview (or sitting on a dragged deck) dropped to ground
under the visually-moved deck until the validating click. Effective slab
records now apply the live delta to polygon/holes/elevation, mapped
exactly once at each query's loop entry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(viewer): reapply floor lift every frame for nodes with live previews
FloorElevationSystem only wrote mesh Y for dirty nodes, but the React
commit that rebinds a dragged node's base-Y group position can land
between frames, after priority-2 systems consumed the dirty mark — the
lift then vanished until the next pointer tick re-dirtied the node,
blinking the Y of items dragged over elevated slabs. Nodes holding a
live override or transform now get the lift reapplied every frame.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(editor): stop the entry camera swap from insta-cancelling walkthrough
Entering walkthrough with a persisted orthographic camera swaps it to
perspective, which recreates the interaction callbacks and re-ran the
pointer-lock effect: its cleanup called exitPointerLock, and the unlock
handler read that as "user left walkthrough" — instantly cancelling the
fresh entry and arming the browser's ~1.25s re-lock cooldown (hence
needing multiple button presses). The effect is now mount-stable (the
changing callback rides a ref, deps down to [gl]), and the entry lock
request swallows async cooldown rejections like the P-resume path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(core): don't block build-JSON import over plugin-node children
Exports from projects with plugins carry nodes like trees:tree whose ids
sit in level.children; the import validator parsed parents against the
static children id union, so one tree id hard-failed the level schema and
blocked the whole import — while the same data loads fine from the DB
(setScene never runs this gate). Parents are now validated against a copy
with non-static-schema child ids filtered out; those nodes keep surfacing
through the unknown-types warning and the imported payload is unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* feat(core): treat registered plugin kinds as first-class in import validation
Nodes of runtime-registered plugin kinds (trees:tree, trees:grass) were
lumped into the unknown-types warning even when the plugin is loaded.
The validator now consults the node registry: registered kinds validate
against their own registered schema (corrupt plugin nodes still block),
count under stats.pluginTypes, and raise no warning — only genuinely
unregistered types do, which stays correct for hosts without the plugin.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* style: appease biome (format + forEach block body)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
