TASK 31 — GT footprints ignored heading, and a flat-object box finding

Why

Two things surfaced while reviewing carpet node #0 in arabic_room against ground truth ([[TASK 29]], [[TASK 30]]).

1. The GT footprint in every debug/viewer view was drawn with the wrong orientation. dump_debug.py wrote each GT object's box as center ± size/2, dropping heading. arabic_room's carpets carry a heading: #58 (left) is -1.568 rad ≈ -90°, #32 (right) is ≈ -180°. So #58's long axis (size.x = 2.16) actually runs along world y, but the viewer drew it along x — the carpet appeared horizontal when it is really vertical. The lifted LiDAR cloud confirmed the truth: node #0's points span ≈1.4 m in x and ≈2.1 m in y.

2. For a flat floor object, the averaged box under-sizes the long axis. The ObjectMap box is a point-count-weighted mean of each observation's own box (_Node._recompute), which is correct for volumetric objects — pooling near-face LiDAR slabs across viewpoints inflates them ~6x. But a carpet is planar (all points z ≈ 0.04), so every view lifts points in the same plane with no depth scatter, and no single oblique view sees the full 2.17 m length. Averaging the per-view boxes therefore converges short.

What changed

• dump_debug.py now applies heading to GT boxes. New _gt_world_aabb(e) rotates the four footprint corners by e.heading (Rz) and takes their world AABB; z is left unrotated (floor objects). viz.json's gt[].bmin/bmax come from it, so merge_video.py and inspect_vp.py — which read those fields — now draw the correct orientation with no change of their own.
• Regenerated debug/arabic_room (frozen dets, 0.6, keyframes 2, image-lag 0) and the carpet merge video. Left carpet GT now renders vertical, right carpet horizontal; node #0 sits inside its GT footprint at the right orientation.

The box finding (measured, not yet fixed)

Node #0 (30 obs, GT #58 = 1.44 × 2.17 m):

box	dims	IoU vs GT
current (weighted mean of per-view boxes)	1.38 × 1.52	0.670
pooled cloud, raw min/max	1.41 × 2.09	0.948
pooled cloud, 2–98 percentile	1.31 × 1.58	0.662

Raw min/max over the accumulated cloud nearly nails the carpet (0.67 → 0.95), because the flat object has no near-face scatter to inflate it. The percentile trim fails too — for a large flat object the "2–98% tail" it clips is real extent. This is exactly the option that is unsafe for volumetric furniture, so the natural fix is shape-conditional: size flat floor objects (carpet/rug/mat, box-z ≈ 0, large footprint) from the pooled raw cloud, keep the averaged box for volumetric objects. Prototype + re-score is the follow-up.

Scope note

The scorer (perception/eval.py) is unaffected by item 1: its primary metric is center-distance, and it deliberately ignores heading (axis-aligned approximation). This was purely a visualization defect. The box under-sizing (item 2) does touch IoU-based secondary metrics but not the center-distance primary.