Props and rigs

The route’s first five subjects ended at a textured mesh. A game needs more than that: a figure has to be posed, and it has to hold things. This page covers the stage after paint — and it opens with a negative result, because that is what the stage has produced so far.

The subject that will not rig

A harbour guard was reconstructed to test posing and props. It reconstructs cleanly and then fails auto-rigging. Four candidate causes were eliminated by measurement:

candidate	how it was tested	outcome
untextured input	submitted a textured plate and an untextured one	eliminated — both fail
component count	compared against a subject that rigs, by shell count	eliminated
mesh topology	manifold edges, shells, watertightness	eliminated
the pose	a narrow A-pose and a wide one, measured arm-to-torso separation	eliminated

The discriminator is unidentified. canon/BASE-FIGURE.md carries this table and declines to name a cause.

That refusal is the interesting part. This repo did name a cause — arm-to-torso separation — and wrote it into canon. It was falsified inside the same hour: a true A-pose failed at separation 0.5412 while a wider pose passed at 0.5483, so the variable moved the wrong way. The canon document now carries the table and no cause, which is the honest state rather than the most plausible survivor.

What a rig that exists can take

Where a rig does exist, posing has a measured ceiling rather than a fix:

roughly 20 degrees of arm rotation stays clean
roughly 110 degrees shreds the shoulders

Rigid plate binding was built to repair that and does not. All three repairs were measured and all three failed — rigid under 5% is indistinguishable from the original, rigid under 20% is worse, and eight smoothing passes still tear. The tool lives in tools/superseded/ where anyone can run it and watch it fail the same way.

tools/pose_rig.py aims named bones at directions stated in the call rather than at guessed Euler axes, and bakes the deformed mesh through a duplicate-and-apply rather than leaving it in a modifier stack. It filters to armature-skinned meshes, because an auto-rig return can contain geometry that is not part of the figure at all.

Reading what an auto-rigger gave back

tools/verify/rig_report.py opens the GLB container directly — JSON chunk, accessors, JOINTS_0 / WEIGHTS_0, skins — and reports three things:

mesh identity against the file that was submitted (--against)
skeleton joints by name, so a 23-joint biped is recognisable as one
whether skinning is present at all

--require-same-shape compares vertex and face sets, not counts. An earlier --require-same-mesh compared counts and was brittle for the reason any count is: two different meshes can agree on a number. The question “did the rigger return my mesh” is order-invariant, so the check is too.

A prop has a real size, and it is declared

Every reconstruction this route makes comes back normalised to 1.002 on its longest axis. Measured on four unrelated assets — a body, an A-pose body and two shields — to three decimals. So a prop mesh carries no true scale at all, and a heater shield loaded as it arrives is 1.8 m tall: a door, not a shield.

The figure is correctly sized only by coincidence, because a human is what the unit box happened to be fitted to.

The scene unit system, stated once rather than assumed: 1.0 unit = 1.8 m.

# build a library prop at true size
python tools/prop_scale.py --glb raw.glb --out props/shield.glb \
    --name heater_shield --real-mm 758 --axis height --attach forearm_l

# check one
python tools/prop_scale.py --verify props/shield.glb

--real-mm is authored, not read off a normalised mesh. The size lands in a .prop.json sidecar beside the GLB, and --verify checks the file against its own declaration rather than trusting it.

The refusal that matters: a prop whose longest axis measures ~1.002 is an unscaled reconstruction wearing a prop’s sidecar, and --verify exits 2 while reporting the real size in millimetres. That is the exact mistake the contract exists to prevent — a file that looks like a prop, declares a sensible size, and is a door.

The predecessor to this contract was a HEIGHT_FRAC = 0.42 in a compose script: a size chosen by eye, per placement, with nothing recording what the object was supposed to be.

A prop that touches is not a prop that is held

tools/verify/prop_contact.py measures the gap between a prop and the figure. It reports two numbers, and the second one is the point:

minimum point-to-surface distance — do these two surfaces meet anywhere
contact AREA — how much of the prop is actually against the figure

A minimum of zero says two surfaces meet somewhere. One vertex meets somewhere. The asset that prompted this tool had a minimum of 0.00072 and a contact patch of 97 points out of 624,510 — it passed “touching” and was plainly not being gripped.

The floor is an area, not a count, because a count is a fraction of however many samples the prop happened to get, and is therefore a property of tessellation rather than of contact. The tool samples vertices and area-weighted surface points for that reason.

python tools/verify/prop_contact.py --glb posed.glb --prop shield \
    --tol 0.002 --min-area 0.0005 --json-out contact.json

Both thresholds are flags rather than constants, and both are set independently of the result they judge. A gate whose threshold is buried in its source cannot be stated in a report, and a threshold chosen after seeing the number it judges is not a threshold.

Defaults: --tol 0.002 units, --min-area 0.0005 units squared — about 16 cm squared on a 1.8 m figure.

Base figures carry no props

Standing rule: a base figure is generated arms to the side, holding nothing. Props are separate GLBs, stored at true size, attached afterwards.

The first reason written for this rule was wrong and the correction is kept in place beside it in canon/BASE-FIGURE.md. The rule survived the correction because it is right for a different reason than the one first given: a prop modelled into a figure cannot be swapped, cannot be sized, and cannot be checked — all three of which the contracts above exist to do.

Clause 7 — five digits per hand

A six-fingered hand reached an accepted sheet past four gates that could not see hands. The Director found it by looking.

Canon clause 7 exists because clause 2 caused it, and it is worth stating plainly: no metric in this repo could separate a five-fingered hand from a six-fingered one, because every one of them was measuring colour statistics, silhouette agreement or surface coverage. Hands are now an acceptance item at the Director’s zoom.

This is the same lesson the repo keeps re-learning in new clothes: a measurable proxy is not a conservative substitute for asking.

What this stage does NOT claim

No prop has been mounted and accepted. Both contracts are tested against synthetic meshes at a known separation. Nothing has been placed in a hand and ruled on at the Director’s zoom, so the contact gate has no scalp on a real asset.
The rig question is open. Four causes are eliminated; the fifth is not known.
The second subject is not textured. Geometry, rig and props are measured; the texture route has not been run on it.