[DrunkOnJava]

Summary

A room whose stored solid RE-101's exact walk does not close on its record box is read again from the faces its solid lists, up to the solid's own end, and labelled apart in SpaceOutlineSource: stored_solid_faces where the faces' edges close as the walk reads them, stored_solid_approximate where a column overlaps a wall (its area depends on whether Revit's export takes the column off) or the outline falls up to 0.02 ft short of the box (Griffin, 2026-11-03: close them, labelled). stored_solid stays the exact walk's. On the BIM4LCA office as built all 13 rooms that kept their box take an outline (9 stored_solid_faces, 4 stored_solid_approximate); at building permit 11 of 12 do (all stored_solid_approximate).

Motivation

Rooms first (#1260): a room written as its record box where it is not a rectangle is the failure that matters. RE-313 left 12 rooms per office file on their box with the cause not found; RE-316 finds it (a record longer than 0x200 bytes between two faces, walls on lines up to 0.0015 ft apart, a column overlapping a wall, a solid longer than the 64 KB window).

Round 3 of the review found room 1671995 at 323% over Revit's area (350% at permit) and asked for a guard. The 323% is the scorer's: it compares the faces of Revit's mesh that point up, and Revit's ceiling over that room is mostly not flat (1218.33 ft² in 9 patches). Revit's floor under it is 5177.37 ft² and ours 5154.12, 0.45% under; the same metric reads RE-313's 66% room (2611271) as 66% off where its floor agrees to 0.1%. See RE-316 section 3.

Changes

• src/research/partition_room_boundaries.rs: the span reads its edges two ways. As the walk reads them (stored_solid_faces: one outer loop, nested loops, no crossing, and a span bounded by the solid's list of objects and a gap of 1,024 bytes) or as the boundary of the region they wind around (stored_solid_approximate). Column-involved rooms are no longer labelled as the exact reading.
• tools/re/space_overlaps_vs_ifc.py (new scorer, registered in Measure's IFC scorers; its count lines are kept for held-out models by tools/ci/scrub-holdout-results.sh): each space's upward-face, floor and every-face plan areas against Revit's, and how much of a room lies inside the walls and rooms around it.
• Docs and labels: docs/rooms.md, docs/export-diagnostics.md (+ schema), docs/status.md, docs/supported-profile.md, docs/evidence.md, changelog.d/changed/1260-room-solids-resolved.md, reports/element-framing/RE-316-room-solids-resolved.md; the closure probe says why each resolved room's edges do or do not close.

Testing

Unit tests are banned in this repository: none added. End to end on Measure (commands in RE-316 section 6).

• tools/check-local.sh passes: not run on this Mac (no local compute); CI's fmt, clippy, doc, MSRV, real-file and IfcOpenShell jobs ran green on this branch
• Verified end to end on real Revit files, against Revit's own IFC4 exports: the BIM4LCA office (CC BY-SA 4.0) as built and at permit with space_profiles_vs_ifc and space_overlaps_vs_ifc, room by room in RE-316 section 3; Core Interior and RE1 Architecture unchanged (116 of 116, 11 of 11); a held-out Revit 2021 project, counts only
• Unfiltered Measure against main: 41728355227 (-f base=main -f nordic=true -f r2025=true -f samples=true -f holdout=true, head a00e705, base 7fd339e): scorer output differs on five files, the exports' hashes on the office, the held-out project and one rme room per file; nothing else
• N/A: not a performance change

Rooms per outline source (export diagnostics):

	stored_solid	stored_solid_faces	stored_solid_approximate	room record_box
Office as built	221	9	4	0
Office at permit	222	0	11	1 (3095831)

Area within 0.1% of Revit's export, the scorer's upward faces (before to after): office as built 221 to 230, at permit 222 to 232, Warmond (held out, counts only) 174 to 179 of 179; Core Interior and RE1 unchanged. By Revit's floor: stored_solid 221 of 221 and 222 of 222; stored_solid_faces 9 of 9 (and 6 of 6 on Warmond); stored_solid_approximate 0 of 4 as built (0.45% to 0.68% under) and 10 of 11 at permit (1671995, 0.45% under).

The review's four guards (wall overlap, shell closure, volume, containment) were measured on every room the span reads. None rejects 1671995 and keeps the rooms that agree: it covers 0.03% of itself with walls (the agreeing rooms 0.10% to 7.59%, the walk's rooms up to 23%) and 0.00% of other rooms (0.00% on every room); a closed shell and the volume need the floor and ceiling faces, which face_at reads on 3 of the 13 rooms as built; containment holds on all. No guard is applied, no room is refused for a guard's reason. The residual (a closed loop of another element's faces inside the box and within 1,024 bytes of the last listed face) is not closed by these either; a foreign face that crosses a wall makes the room stored_solid_approximate.

Where this differs from the round-3 instruction: the fallback (withhold the faces reading) would turn 1671995 and the 9 exact rooms back into boxes, which are 1.0% over Revit's floor for 1671995 and 3% to 67% over for the nine. Instead the exact label is narrowed to rooms whose edges close as the walk reads them, and every column-involved room is stored_solid_approximate with its measured error. Withholding it is a small change to resolved_outline_of_solids if the coordinator wants it.

Audit-honesty checklist

• README / docs reflect the current state after this PR
• A changelog fragment: changelog.d/changed/1260-room-solids-resolved.md
• Limitations resolved and introduced are in RE-316 sections 3 to 5 and docs/rooms.md: the approximate rooms, the scorer's upward-face limit, the unproven foreign-face bound, the six rac_advanced rooms with round columns, 49 of 50 rme rooms, Snowdon Towers' seven and room 3095831 keep their box
• Commit messages describe what and why
• No emojis
• No private paths, personal email addresses or internal jargon in committed files
• A reproducible probe or scorer under probes/examples/ and tools/re/, and a dated report

Legal and contribution hygiene

• Apache-2.0
• No Autodesk proprietary source, decompiled internals or NDA'd SDK content: byte observations on files Measure fetches
• No fixtures

Breaking changes

None to the API outside rvt::research. SpaceOutlineSource gains two values; an importer that matched stored_solid exactly now sees stored_solid_faces and stored_solid_approximate on rooms that were record_box before.

Screenshots or output samples

IFC output changes on 13 + 11 office rooms, the held-out project's 6 and one rme room per file: their IfcSpace body is the outline where it was the box. Room by room in RE-316 section 3.

Related issues

Refs #90 (rooms still drawn as their box: this resolves 13 office rooms as built and 12 at permit, and lists those that keep it: office 3095831 at permit, six rac_advanced rooms with round columns, 49 of 50 rme rooms, Snowdon's seven). #1260 (closed by #1288) is the rooms-first report that started this.

[DrunkOnJava]

Automated Claude Code review (the maintainer's review agent; it read the code, the scorers and the Measure runs, over three rounds). Round 1: the PR body held another PR's text, a solid's faces had no bound and no cap against hostile input, and the column-overlap slack had no label of its own; fixed. Round 2: no blocking item after a check that office room 1671995, which had read 323% over Revit's by its upward faces, is 0.45% under Revit's floor area (Revit's ceiling over it is not flat); the labels were narrowed so only rooms whose edges close as the walk reads them are stored_solid_faces. Round 3: no blocking items; the third cause of an approximate outline, the looser join, the room schedule note and a held-out scrub pattern were fixed in e55ff36.