SPRINT14 gate 3.1. The editor is about to offer these GLBs to an author by
name, which turns every silent field in them into a trap for whoever places
one. Audited the placeable set, fixed the gaps in the factory (nothing
hand-edited), and added the two props D needs to author a yard worth playing.
THE AUDIT, and what it found:
- Four `*_anchor` nodes carried no rating_hint at all — door_anchor,
pickup_anchor, grip_anchor, window_light_anchor. adoptAnchor does
`rating_hint ?? 1`, so silence there does not mean "unrated", it means
RATED PERFECT: a shed door skin outranking a concreted post, conjured out
of a missing field. They now deny it explicitly (`tie_off: false`).
- The carport's price resolved only by coincidence: collateralFor() matches a
STRUCTURE whose site id equals the anchors' collateral string, and site_02
happens to name its structure "carport". The editor will generate
"carport_2" for the second one, and the trap would have gone free — the
gutter bug reborn in the sprint meant to bury it. The GLB now carries
`collateral_key`, like the house does for "gutter". Lane A: the runtime
half is yours (collateralFor could fall back to it).
- The trampoline is unpriced, and now says so rather than omitting it.
Three new rules make the audit permanent, and each is proven able to fail:
no silent anchors, no collateral string nobody prices, no baked anchor_type
outside the checked ANCHOR_TYPE enum. Rule 2 found the carport gap on its
first run — an unfaked negative control.
THE PROPS:
- swing_set_01 (+ wrecked) — the palette's honest middle. It had a ceiling
(gum fork 1.0) and a trap (carport beam 0.22) and nothing in between. Two
apex anchors at 0.45, typed `swing_frame` (NOT `post`: the type string is
what the player reads before committing, and "post" promises concrete),
and a crossbar that is the most anchor-looking object in the game and is
not one. Priced at $140 — proposal, reasoning baked beside it, A rules the
number: gnome 25 < gutter 90 < swing 140 < carport 180.
- tree_jacaranda_01 — a second species whose LADDER is the feature.
Gum: 1.00/0.88/0.76, climb freely, pay 24%.
Jacaranda: 0.95/0.52/0.40, climb at all, pay 58%.
That makes tree choice a real decision instead of "is there a tree".
Unpriced by ruling, same as the bike: no limb-failure event exists to see.
Both wreck/orientation claims are measured in browser coords, not reasoned —
`wreck_falls_toward: "+Z"` is baked AND asserted against the mesh.
Selftest 376/0/0. Two consecutive full factory runs: 44 files byte-identical.
Textures (SPRINT2 §Lane E-2). sail_weave.png is a seamless 512² knitted
shade-cloth weave with the stripe banding real cloth has; the script proves the
wrap by evaluating a second tile and requiring an exact match, because Lane B is
being told to set RepeatWrapping and a bad wrap would seam across the whole
sail. Luminance rides in a narrow band so it multiplies the base colour instead
of replacing it — high contrast reads as burlap, not HDPE. Dropped the per-pixel
noise: invisible at ±0.012 and incompressible, it cost 305 KB of the 323.
sail_tears.png is a strip of 4 escalating rips. Each is a lens, not a slit —
fabric under tension parts widest in the middle and tapers to a point, and a
constant-width gap reads as a drawn line. Threads bridge the gap, scaled to the
local width; without them a dark lens is a hole rather than a tear.
Dressing (§Lane E-3): wheelie_bin_01 (240 L, 12 kg, `lid` on its own node so it
flaps before the bin goes over), washing_line_01 (a Hills Hoist — the `head`
freewheels, giving a second wind tell at head height), garden_gnome_01
(collateral bait: a smashed gnome reads where a damage number doesn't).
Bin and tramp land in models/debris/ for Lane C to glob. All 19 assets pass, 28
output files byte-identical across two runs, selftest 129/0/0.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Verifying the GLBs in three.js (not just Blender) showed four assets reporting
inflated bounds: tramp_01 came back 3.29 x 1.27 m against a true 2.96 x 0.78.
Box3.setFromObject expands each mesh's LOCAL box by the world matrix, so a node
carrying a rotation over-reports — the same trap Blender's obj.bound_box sets,
and what three uses for frustum culling. Joined nodes inherited parts[0]'s
rotation, which for an arc is half a segment step off-axis.
Applying rotation at join time makes every local box axis-aligned, so the
default Box3 path is now correct for consumers and culling is tight. World
geometry is unchanged; the exported dims are identical.
Adds tools/assetcheck/, the three.js harness that caught this. It's the only
check that can: a Blender round-trip exports Z-up->Y-up and imports Y-up->Z-up,
so a broken export_yup flips back and passes green.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
16 generated assets plus the grass billboard atlas and the four debris models
copied verbatim from the 3D-STORE library (copies rule, §0). All meter-scale,
Y-up, and far under the 15k tri budget — garden_bed is heaviest at 2,580.
Regeneration is byte-deterministic: two consecutive runs produce identical
files, so re-running the factory causes no churn.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
