perf(nodes): batch ceiling undersides and slab bodies (charter row 16)

[wass08]

Metric

rich-4× idle frame CPU (charter backlog row 16): ceilings and slabs — 696 opaque, level-parented, one-geometry-per-node meshes on that fixture — draw through the per-level BatchedMesh containers of shared/node-batch/ like items, columns, doors and windows already do.

Same-load PR-vs-main pairs (the machine ran at load average ~7 all afternoon, so absolute numbers are inflated and only back-to-back pairs are quoted):

fixture	idle frame CPU	idle draws	hover CPU
rich 4× (the metric)	9.1 → 8.3 ms (−9 %; −15 % in a second pair)	4 343 → 3 219	10.1 → 8.8
rich 2×	9.3 → 4.9 ms (−47 %)	2 580 → 2 018	8.4 → 5.0
Maxi 8×	15.2 → 10.9 ms (−28 %)	5 472 → 1 376	15.9 → 11.8
tower 2×18	6.0 → 5.4 ms (−10 %)	4 219 → 1 147	6.6 → 5.9
Maxi 1×/2×/4×	within the noise band at 2–4 ms	999 → 487 / 1 587 → 563 / 2 877 → 833	—

Like-gesture edit receipts (wall/door body moves, undo) are flat or better; every fixture ends at dirty 0 with zero phantom marks. One receipt regressed deterministically (rich-2×, moving the wall behind a ceiling-hosted light: 32 frames vs 6–9) and turned out to be a real bug this batching exposed: the grid raycaster never enabled BATCHED_LAYER, so the ground ray under a dragged wall missed the now-batched slab, the wall jumped 8.5 m and created a bridge wall (135 walls re-marked). Fixed in dd745a0 (291 ms / 8 frames after, main 336 / 8), and the same guard was applied to every non-hover raycaster that can hit batched sources first-person door/window targeting, item/column support caps, accessory placement, door/window paint, wall paint — 7a64290; the support-cap and door/window ones were already exposed on main for the kinds batched since #760. The audit table of all 35 raycasters lives in the private charter folder.

What it took (beyond adding two kinds)

• Ceiling dirty marks were consumed before the batch could see them. viewer/src/systems/ceiling/ceiling-system.tsx consumed at useFrame priority 0; the batch snapshots pending marks at 1. It now consumes at 2 like GeometrySystem (definition system.priority orders components, not frames — node-definitions.md said otherwise and is corrected).
• The coloured ceiling underside was transparent: true with opacity 1 (introducing commit 967a905 calls it opaque; no fade). Now opaque, so it batches in the default textures-on appearance. Blend order against glass changes by a few hundred pixels (parity below).
• Slab default side materials were created fresh per slab, so sides could never share a batch or a GLB material. They now use the viewer default cache; the legacy slab cache marks its materials as cached so rebuild disposal skips shared instances, and a cache clear rebuilds slabs and releases batches before disposing (after the next frame).
• Surfaces move through useLiveTransforms, and paint previews swap materials on every fan-out target, without dirty marks. Both release now and defer joins; a paint commit ends its holds through a single owner (paint-preview-owner.ts) — the first cut leaked the hold forever on commit and reached columns/doors/windows too; an apply() that throws after the scene write ends holds without restoring stale materials.
• Shadow flags are part of the container identity (ceilings cast none, slabs cast and receive); externalSelectedIds release like local selection; hidden-level candidates are re-offered when the level comes back.
• Rebuilt surface geometry replaces instead of appending. release() used to keep obsolete packed geometry in used until an overflow rebuilt the whole bucket. Surfaces now own a stable slot with 25 % headroom and rejoin via setGeometryAt when the new geometry fits; overflow rebuilds size from live geometry only; multi-node releases flush once per frame; an emptied container survives the quiet window; released surfaces on a level wait while any wall there carries a live override and until the wall queue drains, then join in one wave. Counters under __pascalPerf.batchStats() (releases, joins, replacements, overflow rebuilds, bytes copied).
• Found by the bake gate, pre-existing: the merged wall batch never carried pascalExport = 'strip', so a capture that landed while it was sewn exported 80 k duplicate wall vertices. Fixed (separate commit).
• Perf receipts for drag:move now carry the moved node id (scope nodeId), so the matrix runner can refuse to diff unlike gestures.

Gates

• Editing: private e2e gate 9/9 (perf-regression pack ×5, batched-wall click, hidden-wall click, floor placement, slab room paint), twice.
• Visual: outline-parity.ts poses H/S/SH/SI on rich 4× + Maxi 1×: 0.36–0.46 % pixels differ, localized to window glass, glass-room seams and two small floor patches; selection census identical on the stable poses (the 3a/3b batching row accepted 0.4 % of the same kind).
• Bake: rich-2× raw GLB node/mesh/texture/animation sets identical to main's minus the leaked wall-batch meshes; materials 91 → 60 (deduplicated slab defaults), −1.1 MB.
• Unit: bun test packages/nodes --randomize --seed=1 2 127 pass; discriminating tests for every mechanism above (the new node-batch tests were written to fail without each fix).
• Four independent review rounds (read-only recon → audit → perf diagnosis → audit of the fix round) closed one blocker and eight should-fixes; the two diagnoses are recorded in the charter.

Follow-ups filed (idea inbox, not changed here)

• The editor's standalone undo marks every node dirty on top of core's changed-node diff (history.ts refreshSceneAfterHistoryJump), releasing every batched member on every level per undo.
• Window move on rich-4× spends 12–17 s in long tasks on main and on this branch alike.

Charter: plans/performance/editor-scalable-scene-runtime.md row 16 (private PR follows with the evidence row, dashboard refresh and the gesture-aware runner).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Jmmz2AMwTzcnKsSHHPEMhN

---

Note

Medium Risk
Large changes to batch membership, material cache lifecycle, and many editor raycast paths; regressions would show up as wrong placement, paint, or support surfaces rather than auth/data issues.

Overview
Extends level BatchedMesh batching to ceilings and slabs (alongside items, columns, doors/windows), with a reworked batch store: per-shadow containers, reserved geometry slots for mutable surfaces, coalesced releases, and deferred empty-container disposal. Sources stay mounted but draw-hidden; live transforms, slot paint previews, external selection, and material cache clears now release and re-offer batch members on the same rules as hover/selection.

Raycasting for tools and walkthrough no longer misses batched floors/slabs: non-hover raycasters call setSurfaceRaycastLayers so hits still use source meshes when batch draws take over. Ceiling rebuild timing moves to frame priority 2 so the batch’s priority-1 dirty snapshot sees geometry changes; ceiling underside materials are opaque so they can batch in the default appearance.

Paint gets a shared paint-preview-owner (fan-out cleanup/commit) and slot-preview holds that keep batched nodes on their sources until previews settle. clearMaterialCache registers cleanups and disposes after consumers rebuild; selection highlights and slab caches hook in so previews never restore disposed materials. Wall merged batches and node batches are tagged pascalExport: 'strip' for GLB capture.

^{Reviewed by Cursor Bugbot for commit 7a64290. Bugbot is set up for automated code reviews on this repo. Configure here.}

[pascal]

I hit an error while handling your request (Model unavailable on AI Gateway free tier: Free tier users do not have access to this model. Upgrade to paid credits at https://vercel.com/d?to=%2F%5Bteam%5D%2F%7E%2Fai%3Fmodal%3Dtop-up for unrestricted…).

Please try again, rephrase, or reach out if it keeps failing.

Error id: d4012a5d-4673-411a-98cb-33f3158b6e6c