Wall load fixes: treatment fan-out, one CSG cut per wall, capture draw leak, converter duplicates

[wass08]

Metrics (charter backlog rows 8, 9, 10)

• IFC castle (Ifc2x3_SampleCastle.ifc, 932 walls / 464 openings / 6 levels — the production import that froze the editor): load-to-settled never (514 s) → 10.1 s, all 932 walls batched, frame 42 → 21 ms.
• Ladder (private run-scaling-matrix.ts, vs the Wall batching stays live in cutaway #761 build, same machine): load settled −13…−29 %; wall-move settle Maxi-8× −27 %, rich-2× −23 %; undo −18…−29 % (rich 2×/4×, tower); item-move −25…−28 %; Maxi-8× end-of-run dirt 162 → 0.
• Draw leak (charter finding 1): rich-4× draws stay 4 494 across the thumbnail timer (were 5 662 after it, permanently).

What changes (one commit each)

1. Wall treatments — the level treatment miters were recomputed every frame while any wall on the level was dirty, and every wall renderer on the level re-rendered (65 k WallTreatments renders in 12 s on the castle; the same fan-out on every wall drag). Now: skip when the level's effective walls (identity, live overrides cached per override) and trim offsets are unchanged; per-(level, proud) miter cache; each wall subscribes to a slice holding only its own endpoint junction entries, compared structurally; no treatment work mounts for walls with no trim enabled.
2. Openings CSG — cutouts were subtracted one by one from a growing result. Now grouped by bounding-box overlap: disjoint cutouts merge into one cutter, overlapping groups of ≤ 4 are unioned, larger groups fall back to the bounded chained path; one subtraction per wall. A wall with ≥ 6 cutouts is deferred to its own frame when the frame already rebuilt something (the budget was only checked between walls).
3. Coincident cutouts — a real import carried the same door eight times on one wall; box cutouts contained in another box are dropped before grouping (the union chain over coincident boxes took 148 ms and froze the page).
4. Capture draw leak — thumbnail:before-capture (fired by the thumbnail autosave 30–40 s after load) made the node batch's sweep reveal every mesh held under the shared 'batched' reason, including the wall sources the merged wall batch already drew; nothing re-hid them. Wall holds now use 'wall-batched' and the wall batch handles capture itself (reveal before, re-hide after — exports still prune off-layer meshes).
5. Placeholder sweep — a degenerate rebuilt wall (zero-length / fully cut) has as few vertices as the mount placeholder and was re-marked every 30 frames; rebuilt geometry is now stamped userData.built.
6. IFC converter — each fill (door/window) is emitted once; spatial children use sets. Existing conversions keep their duplicates until reconverted (the editor now tolerates them).

Gates

• Unit: viewer + nodes + ifc-converter 2 103 pass (58 new: frame skipping / override identity / level membership / selector neighbours / byte-identical trim geometry; disjoint, overlapping, coincident and nested cutouts by signed volume + material groups; budget deferral; capture holds; sweep stamp; converter dedupe on the Open House fixture).
• E2e against the branch build: perf-regression pack 5/5, click-batched-wall, click-hidden-wall, opening-placement-preview green; placement-preview-follows-cursor fails identically on main (pre-existing).
• Canvas parity vs the Wall batching stays live in cutaway #761 build: rich 2×/4× ≤ 0.4 % on all four poses (load / after orbit / full height / cutaway again); Maxi-4× post-orbit poses differ only by orbit end-pose drift.
• Bakes (Maxi 1×, rich 2×): identical meshes, primitives, nodes and materials; 2 and 46 fewer triangles (one cut instead of chained cuts).

Notes

• The castle stays a top-of-ladder fixture in the private repo (project_IfcCastle2x3Smpl); its remaining cost is the steady-state frame (21 ms at 5 600 draws), not load.
• Found on the way, not addressed here: select mode mounts three transparent corner-bracket meshes per ceiling corner for every ceiling on the level (1 236 meshes on rich-4×) — charter row 11.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Jmmz2AMwTzcnKsSHHPEMhN

---

Note

Medium Risk
Touches core wall CSG, treatment subscriptions, and scene-layer holds; behavior is heavily tested but geometry and render paths are easy to regress subtly.

Overview
Targets wall load and steady-state cost after large IFC imports: fewer redundant treatment updates, cheaper opening booleans, no permanent double-draw after thumbnails, and unique doors/windows from conversion.

Wall treatments skip level miter rebuilds when effective walls (with cached live overrides) and trim proud keys are unchanged; a per-level miter cache and per-wall zustand selectors (endpoint junction slices only) stop every wall on a level from re-rendering when one wall is dirty. Trim mounts only when hasWallTreatments is true.

Wall geometry groups opening cutouts by overlap, drops contained box duplicates, unions small groups (≤4) via mergeWallCutoutBrushes, then does one CSG subtraction per wall (with chained fallback for large groups). Progressive rebuilds use shouldDeferWallRebuild so walls with ≥6 cutouts wait for a fresh frame. Rebuilt meshes get userData.built so degenerate walls are not swept as placeholders.

Batching / visibility uses a distinct 'wall-batched' hold reason; WallBatchSystem reveals source meshes on thumbnail:before-capture and re-hides after capture so exports see geometry without leaving walls drawn twice.

IFC converter dedupes spatial children with Sets and emits each door/window fill once (expressIdToNodeId guard + register after successful node creation), with fixture tests for repeated relationships.

^{Reviewed by Cursor Bugbot for commit 5e2243b. 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: 5e32fcc2-5966-4b9c-a3ab-43f3e2cc60ee