Measure collision fixes from convex pieces, and gate the glb on its own

[ZheningHuang]

Stacked on #35 (the publish move). Review that first; this PR's diff is against it.

Three changes to collision QC, in the order the problems were found.

---

1. The fix distance was wrong, so correct fixes were thrown away

FCL decides contact from the raw triangles — the only model that gets a tucked chair right. But the furniture meshes are not watertight (Table0 alone has 846 connected components), so FCL's penetration depth on them is garbage — a desk flush against a wall once reported 1.16 m. The fix distance therefore came from the two objects' world AABBs, which over-states it by 1.6–2.8× and can only push along a world axis. Over-stated moves exceed MAX_NUDGE, get reverted, and the clash ships.

Convex pieces are closed by construction, so their arithmetic is meaningful. Detection is untouched — the decomposition is asked only how far to move a pair the raw triangles have already confirmed is touching, so its ~9 mm over-approximation can never invent a clash, only size a real one.

A single convex hull would be far too coarse to detect with — measured, it swallows the entire 80 mm gap under a tucked chair and reports contact:

pair	true gap	one hull	decomposition
Chair2/Table0	80.2 mm	0.0 (contact)	17.2 mm error

Effect across the seven built rooms

17 clashes
  total requested displacement   AABB 8.15 m  ->  convex 3.49 m
  fixes exceeding MAX_NUDGE=0.3  AABB 13      ->  convex 2

11 of 17 fixes flip from reverted to applied. The clearest case is a room where all six chair fixes asked for 0.625 m (the chair's own AABB, not a penetration depth) against a 0.30 m cap; the real moves are 0.127–0.250 m. The two that still fail are genuine — objects overlapping by 0.3–0.8 m, which is a placement error, not a mismeasured nudge.

Cost

Accuracy is governed by CoACD's preprocess_resolution, not its threshold — 24 → 128 pieces moves the error 17.2 → 17.4 mm, while the voxel cell size tracks it ~1:1:

pre_res    cell     error
     50   16.2mm   17.2mm   <- CoACD default
    100    8.1mm    8.7mm   <- chosen
    400    2.0mm    2.1mm

Decomposition runs only for bodies already in contact, and is cached beside the glb keyed on canonical (un-placed) geometry, so eight chairs cut from one asset decompose once. Cold 112 s on the worst room (was 465 s before restricting to bodies in contact), warm ~2 s, and a clean room pays nothing.

---

2. The check needed two files that could disagree

check_all wanted a glb and whichever Room.py was newest — and checks.py picks its layout by mtime from up to five candidates, so it could grade a room that no longer exists.

Wall centrelines, floor_z/ceiling_z and per-object category are not recoverable from geometry, so they are now written into the glb as standard glTF extras at build time. check_glb(path) is one deterministic function of one file.

Verified on all eight built rooms: findings identical to the Room.py path. Before, running from the glb alone found 1 of 12 on Elliott-Studio. Unstamped glbs return {} and fall back, so older builds still work.

room_qc/collision.py is the gate — reads only, exit 0 clean / 1 clashes. Grounding and containment are reported but do not fail it: a floating object means a missing support, and outside_room means the scan's wall loop is wrong — neither is something the collision resolver can repair.

---

3. Nothing ever re-checked after fixing

correct.py printed "rebuild the room, then re-run to verify 0 clashes" and nothing did. publish runs it once, so a room ships with whatever the first pass left — Panda-2 went 8 moves, then 6, then still 6 across runs.

resolve.py closes it. A fix only translates objects in the floor plane, so the same translation applied to the in-memory meshes is what a rebuild would produce — the loop runs detect → nudge → detect entirely in memory and writes Room.py once.

Panda-2            6 → 6 → 4 → 3 → 3 → 3 → 1 → 0    converged in 7 rounds
Meeting Room copy  6 → 3 → 2 → 1 → 1 → 0            converged in 5 rounds

That second room is the one whose six fixes were all previously reverted.

Verified end to end — resolve --apply → rebuild in Blender → gate the rebuilt glb: 0 clashes, exit 0. So the in-memory shortcut is checked against the real artifact, not trusted.

Why it stopped is part of the report, because a clash count cannot tell "give me more rounds" from "this needs a human": converged / capped / nothing_movable / max_rounds.

---

Layering

extract_shell moves to room_ops/shell.py (re-exported from geometry.py, no caller changes) so room_ops can read a SHELL without importing agent. tests/test_architecture.py caught that on the first attempt.

Verification

• 567 tests pass; 4 new test files (convex MTV, glb metadata, the gate, the loop)
• ruff check src tests clean
• layering test passes

Not in this PR

publish still runs correct → checks; wiring it to resolve → build → gate is left for a follow-up so this stays reviewable.