The teeth retreat spent the tongue clearance — two assets now have 1.0 mm where they had 7.3 mm

[gidich]

#739 cleared every humanoid's teeth from its face by spending the tongue clearance. Two assets now
have 1.0 mm between the teeth and the tongue, where they had 7.3 mm, and nothing has checked whether
that survives the jaw opening.

Measured 2026-08-28 on main at 0a982e8e

Teeth-to-tongue gap, before #739's retreat and after, in millimetres:

asset	before	after
mpfb-clinical-physician-adult	7.3	1.0
mpfb-clinical-nurse-adult	8.0	1.1
mpfb-gown-inspect	11.7	1.4
mpfb-ob-patient-aisha	11.7	1.4
mpfb-viseme-inspect	11.7	1.4
mpfb-family-partner-adult	7.5	2.0
mpfb-gown-adult-patient	7.4	3.7
mpfb-peds-patient-child	5.6	4.0
mpfb-peds-nurse-kevin	7.5	5.5
mpfb-street-adult-male	7.5	5.5
mpfb-peds-parent-aisha	4.5	4.5 (untouched)

#739's counterweight requires the gap to stay positive and it does, on every asset. This card is not
about that clause failing. It is about the headroom behind it.

Why the margin is this thin, from the fix's own arithmetic

The retreat was |marginAtCap| + 0.0005. The physician's pre-retreat teeth-to-tongue gap was 1.5 mm
and its cap margin was -5.8 mm, so clearing the face at all consumed almost the whole gap. A safety
of 1.5 mm or more would have zeroed or inverted it. 0.5 mm was the largest value that fits the
population, and it leaves the two tightest assets at about a millimetre.

What is not known, and it is the reason to file

The tongue bound was measured at rest. Nothing has checked it under the jaw rotation
(JAW_OPEN_TEETH_CLEAR_RADIANS, up to 8.64°) or under any of the 47 body morphs. The teeth and the
tongue are both skinned to the same rig, so a jaw rotation moves both — but they are separate meshes
with separate weights, and 1.0 mm is inside the range where a weighting difference matters.

The measurement is cheap and the instrument exists: every-humanoids-teeth-stay-behind-its-own-face
already applies morphs statically and derives per-asset bands.

The alternative this makes worth considering

Five assets resolve to two head geometries — gown-inspect, ob-patient-aisha and viseme-inspect
share one; peds-nurse-kevin and street-adult-male share another. Deduplicating them would turn
eleven per-asset retreats into a smaller set of geometry fixes, and a geometry fix has room to move
the tongue as well as the teeth rather than only sliding one past the other.

Not measured

Whether the tongue is even visible in any shipped frame. If it never renders, a 1.0 mm gap is a
bookkeeping concern rather than an appearance one, and that changes the priority entirely. I have not
graded a frame with the mouth open wide enough to show it.

factory_step: body_param

[gidich]

Graded one of the two tightest assets. At rest the retreat is invisible, which lowers this card's
urgency without settling it.

MY GRADE — mpfb-clinical-physician-adult, front lit, native resolution

Captured on merged main at 0a982e8e with pnpm asset:model-vetting:glb-grade --glb apps/ui-xr/public/generated-humanoids/mpfb-clinical-physician-adult.glb. This is the asset whose
teeth-to-tongue gap went 7.3 mm to 1.0 mm.

The figure reads as an upright clothed adult: white lab coat over teal scrubs, dark shoes, arms at
the sides. The mouth is closed and reads as lips. No teeth are visible and no tongue is visible.
So the 6.3 mm retreat produced no visible change at rest, and the 1.0 mm gap it left is not something
a learner sees in this pose.

That answers the priority question this card raised: at rest, a 1.0 mm teeth-to-tongue gap is
bookkeeping. It does not answer whether the tongue appears with the mouth open — this capture is
a static rest pose and drives no viseme.

Out of scope here, and I am not fixing it

The same frame shows torn geometry where the lab coat collar meets the neck: dark speckled fragments
around the throat and a ragged edge where the teal scrub shows through. The scrub V-neck has a torn
boundary at the shoulders, the coat's front edge and hem carry thin jagged slivers, the eyebrows are
dark speckle (#597), and the shoes read as dark blobs.

The collar tearing is the class #73 recorded for garments meeting the neck. It is not filed from
here — a single asset's grade is not a population claim, and I have not measured it.

What this changes for this card

Lower priority than filed, and the remaining question is narrower: does the tongue show at an open
mouth on a 1.0 mm gap, and does the gap survive jaw rotation. Both need a driven capture of this
asset, not a rest pose.

Method note

--output-root is resolved relative to the repo root, so an absolute path lands inside the tree. I
wrote a stray directory there and removed it. Use a repo-relative output root or the default.