All done: 11 renders used of 25, no warnings on the final check, and the GLB re-imports at the same size. Final report follows.

Treehouse — dogfood report (treehouse)

1. The model

dogfood/treehouse.aix is a treehouse scene about 10 units tall (surface extent 7.61 × 10.21 × 7.87), output with scene tree, house, ladder, swing so the GLB has four nodes. Beauty render: out/treehouse/beauty.png (1024); close-up of the hut: out/treehouse_hut/sheet.png.

How it is built (44 shape steps, all in the output; the numbers deck_y, hut_x, sw_top, … are printed by check too):

• tree — trunk is a cone(0.75, 0.4, 7). Three branch(az, y0) calls (a def) are tube(0.3, curve([...]), taper=0.4) leaving the trunk at y≈3.6, curving out through the deck and up to r≈2.4 at y≈6.6, each turned by rotate(y=az) to azimuths 300°, 60°, 150° so none passes through the hut; swing_branch is a fourth, fatter tube that leaves the trunk at y=5.2 (above the deck) and runs out to x=4 at y=6.8, clearing the +x railing. wood = union(trunk, branches, swing_branch, k=0.25) | displace(0.05, 0.35) | paint(bark). canopy is seven spheres in a union(..., k=0.7), displaced and painted; tree = union(wood, leaves, k=0.3) - below, where below is a box under y=0 giving the trunk a flat base exactly on the floor.
• house — deck: ten box(4.8, 0.16, 0.4) planks in an array, minus clearance = branches | offset(0.2) so the planks are cut clear of the branches that pass through them; a collar disc round the trunk closes the gaps there; three joists sunk a cell into the planks. railing: 20 posts in a for over five positions per edge and two rails at deck+0.45 and deck+0.85, with the front rail split to leave the ladder gap. hut: wall_box - inner (walls 0.1), minus door_cut on the front (+z) and window_cut on +x, a mullion cross, and roof = extrude(polygon(triangle), 2.0, "x") — a gable whose ridge runs along x with 0.2 eaves. lantern: an iron hook tube into the joist, a brass ring_top torus, lan_cap cone, lan_glass amber cylinder with a glow=2 sphere inside, lan_base.
• ladder — two tube(0.06, ...) rails from (±0.35, 0, 4.5) to (±0.35, 4.8, 2.3), ending in the deck's front edge, and ten cap="flat" rungs placed with lerp in a for.
• swing — seat box at y=1.5; two rope tubes from inside the seat up to sw_top = height(wood, 3.4, 0.1) - 0.13, i.e. measured into the branch, painted with a 0.05 stripes material so they read as twisted rope.

Materials: bark (noise), leaf (noise), plank (wood, axis x), siding (0.2 stripes, clapboard), roofwood, rope (stripes), brass, amber, iron, a glowing #ffd070. Lighting: light_azimuth 40, light_elevation 50, ambient 1.6, zoom 1.3, grid 200 (cell 0.055).

2. Print-readiness (final out/treehouse/report.md) and re-import

Row	Value
Size	8.685 × 11.045 × 8.48 (box); surface extent x −3.457..4.154, y −0.004..10.213, z −3.309..4.566
Triangles	214412
Watertight	no: 24 edges — "mostly 2 at (−2.114, 7.839, −2.485) in 'blob', 'canopy'; 2 at (1.142, 8.953, −1.494) in 'blob', 'canopy'; 2 at (−0.651, 0.006, 0.006) in 'trunk', 'below'". These are scattered pairs on the displaced canopy surface and on the trunk's flat base cut — sub-cell folds of displace, not a construction fault. (Down from 123 → 194 → 234 → 158 → 38 → 24 over the iterations.)
Stands	yes, centre of mass 0.59 inside the footprint (38-sided hull)
Overhangs	19%
Pieces	1
Cavities	none

Re-import (dogfood/treehouse_reimport.aix: import("../out/treehouse/model.glb", resolution=160), 92 s to sample 217360 triangles at cell 0.064):

	Original	Re-import
Surface size	7.61 × 10.21 × 7.87	7.61 × 10.22 × 7.87
Volume	80.159	79.369 (−1%)
Centre of mass	(0.14, 7.268, 0)	(0.141, 7.292, −0.004)
Pieces	1	4: lantern loose (0.021 at (1.9, 3.88, 2.2)), two volume-0 specks at (2.382, 5.14/5.54, 0)
Watertight	24 edges	86 edges

What survived: the trunk, branches, canopy, deck, posts, hut with door and window, roof, ladder rails and the two ropes and seat. What went: the 0.08-thick rails (the +x pair became the two specks), the ladder rungs are dotted, and the 0.08 hook, which is why the lantern is a separate piece — all under 1.2 × the import cell (0.077). The GLB itself has all four nodes and 217360 triangles; the losses are the import's sampling, not the export.

3. Friction

1. check, first draft, line 71 (then mullion = ... | rotate(y=90)). Sizes showed mullion 0.5 × 0.5 × 0.07 (z -1.235..-1.165) — thin in z where I meant thin in x. My mistake, caught by the size line; not a tool fault, but it shows why the sizes matter.
2. check, first draft. warning: 5 steps are thinner than a grid cell (mullion, hook, ring_top, lan_base, house), the thinnest 'mullion' at 0.04 ... or set grid 512 covers them all. Naming house (a union of forty parts) as thin because a child is thin is noise, and "set grid 512" is a 17× sample-count recommendation offered as an equal to thickening four steps by 0.03.
3. Render 3 report, watertight row, lines 23–26 (the branch tubes). 194 edges ... mostly 17 at (0.476, 4.073, 0.217) in 'trunk', 'swing_branch'; 14 at (3.242, 5.777, 0.057) in 'swing_branch', 'rope_one'; 13 at (0.672, 4.406, -1.134) in 'plank_one', 'wood', then after fixing the joints render 5 gave 158 edges ... 18 at (0.914, 5.509, 0.191) in 'swing_branch', 'wood'; 13 at (-1.091, 4.409, -0.534) in 'branches', 'wood'. The named pair is "innermost step and its parent", not the two surfaces that clash, and every cluster centroid sat on the axis of a branch where no other part was. It took three full renders to decode that as rings of sub-cell slivers at the segment joins of a tube(..., smooth=4, taper=0.4); swapping the point list for curve([...]) (same points, same taper) dropped the count from 158 to 38. Neither docs/language.md nor the reference says a smoothed, tapered polyline tube leaves such rings; the docs pitch curve only as "no facets".
4. Renders 3 and 4, Cavities row. volume 0 at (1.039, 4.674, 0.008); volume 0 at ... (enclosed voids, not loose parts) — four sub-cell voids, described as enclosed cavities, at the same tangencies the watertight row was reporting. They vanished with the same fix. Two diagnoses for one cause, and neither said "these are grazing surfaces".
5. Render 5 report. Pieces 2 (the largest 79.355; then 0 at (1.836, 7.958, 0.003) in 'blob', 'canopy') — a volume-0 speck in the k-blend neck between two canopy spheres counted as a separate piece. The message for a real loose part and for a sub-cell speck is the same row; I fixed it by moving a blob 0.4, which was a guess that worked.
6. Report only lists three clusters. With 234 edges, I could not see whether the rest were on the same steps; a per-step tally would have pointed at "all four branch tubes" in one read instead of three renders.
7. ground() in a scene, line 40 of draft 2. tree = (wood + leaves) | ground() was going to lift the tree by the bark's displacement amplitude while house, ladder and swing stayed where their numbers put them; nothing warns that grounding one object of a scene moves it relative to the others. I caught it reading the check box (tree y -0.04..10.6) and replaced it with a cut at y=0. The docs describe ground() per shape only.
8. set zoom 1.3, line 10. No visible change between renders 7 and 9: the camera frames the box (tree box y −0.32..10.7, x −4.02..4.77) which the union(k=) and displace padded, not the surface extent (y −0.004..10.21, x −3.46..4.15) that the report knows. The doc's "stops where the box would touch the edge" is accurate, but about 8% of every side of the beauty frame is padding.
9. Re-import. warning: The model is 2 separate pieces: ... the loose piece is volume 0.021 at (1.9, 3.88, 2.2) in 'back' plus 2 tiny specks of mesh. The import's cell (0.064 at resolution 160, its longest side being 10.2) is coarser than the render grid (0.055), and nothing says which parts of the mesh fall under it: the thin-step check works on steps, and an import is one step. I had to work out from positions that the hook and the rails were the casualties.
10. Quick render, first draft. warning: 11 tiny specks of mesh ... in 'plank_one', 'joist' at the quick cell; gone at full grid. The doc says the pieces count is not judged on a quick sheet, but the specks warning is still printed there and read as a fault.
11. Timing. Each full render was 30–45 s at grid 200, of which exports were 9–17 s; the import sample was 92 s, run once for check and again for render (the doc does warn of this).

4. What worked well

check sizes found every misplacement before a picture was made; the sheet's four views were enough to judge the composition in one read; the slices settled the hut's hollowness and the plank gaps in one look; --focus hut at cell 0.0135 showed the door, mullion and eaves at a detail the main sheet could not; height() for the rope's top and offset() for the plank clearance did exactly what the docs say; union(k=) and curve() both cleaned the mesh as much as they cleaned the look; scene gave four GLB nodes with no extra work; the watertight row naming coordinates and steps, and the Pieces/Stands/Overhangs rows, made print-readiness a number I could drive down rather than a hope.

5. Three changes that would have helped most

1. Name both surfaces at a non-manifold cluster, and tally edges per step. "18 at (x, y, z) in 'swing_branch', 'wood'" should say what the branch met (the deck, the canopy, or itself), and a per-step count would have shown at once that one construct — the smooth=/taper= tube — carried them all. Add to the docs: a smoothed, tapered polyline tube leaves sub-cell rings at its joins; use curve() for a tapered branch.
2. Make imports report thin features. Print the import cell next to the model's thinnest parts (or default resolution from the render grid so the two cells match), and warn when the re-imported mesh loses pieces to sampling, so a loose lantern is not mistaken for a broken export.
3. Frame the beauty camera on the surface extent, not the loose box (or let zoom reach the surface extent), since a blended, displaced tree pads its box by up to 10% a side and the report already measures the true extent.

Files: /home/user/aixle/.claude/worktrees/agent-a11fd7f11ce175f69/dogfood/treehouse.aix, /home/user/aixle/.claude/worktrees/agent-a11fd7f11ce175f69/dogfood/treehouse_reimport.aix; outputs under /home/user/aixle/.claude/worktrees/agent-a11fd7f11ce175f69/out/treehouse/ (beauty.png, model.glb, model.stl, report.md), out/treehouse_hut/, out/treehouse_reimport/. Nothing committed; nothing outside dogfood/ and out/ touched. 11 renders in total.