the ED station is 12,406 triangles over its 180,000 budget after the fitted-eyelash trade

[gidich]

The ED four-actor station is over its authored Quest triangle budget after #683 traded helper
eyelashes for fitted ones.

Measured on main, 2026-08-27

ED four-actor station: 192,406 triangles of ACTORS ALONE
authored budget:       180,000  (asset-generation-pipeline.md:87)
over by:                12,406  (6.9%)

the-station-fits-its-budget-and-still-looks-like-people.test.ts is RED. Two further contracts are
red from the same change:

• mpfb2-actor-is-stripped-of-helpers clauses (1) and (5): aisha BODY triangles now measure 36,972
  where the documented stripped body is 26,756, and the destructive probe reads 79,362 against the
  same 26,756 expectation. The lash geometry is being counted as body.
• every-actor-has-brows-lashes-and-teeth clause (3), which the worker records as pre-existing.

This was predicted and declared, not discovered

#683's worker named all of it in its OUT-OF-SCOPE section before I accepted the land: "The #695
budget contract is now RED (predicted by this card's own cost analysis)." The trade was 368-tri hm08
helper lashes for 16,632-tri fitted mindfront_eyelashes_01 across 11 actors.

Accepting that land was my call and the regression is a consequence of it, not a surprise. The
fitted lash is the correct D1 direction — wire the proven CC0 pack rather than keep hand-rolled
helper stubs — and the budget is a real Quest constraint. Both are true, which is why this is a card
and not a revert.

What must NOT be the fix

Raising the 180,000 figure. It is authored in asset-generation-pipeline.md:87 as a hardware
constraint, and moving a threshold to clear an observation buys the observation instead of the
property.

Open questions this card does not answer

Whether the lash count should be decimated per-actor for the ED station, whether lash geometry should
be excluded from the BODY measurement the way garment triangles already are per #321, or whether
the fitted pack has a smaller variant than mindfront_eyelashes_01 — the worker records that as
already the smallest the pack offers at 8,316 quads.

Also unanswered: whether the fitted lashes are visibly better at conversational distance. #683's
own prompt requires a reviewer-graded pixel verdict and it has not been given. If they are not
visibly better, the trade is bad on both axes and reverting becomes the honest answer.

No ## done_when yet — the remedy is a design choice among the three above, and inventing a contract
before that choice is made would encode a guess.

OPERATIONALIZED 2026-08-28

The revert-or-keep fork is closed. At a 1024 head framing with the head at 0.80 of frame, the fitted
lashes read as lashes — strand-textured fringe, inner to outer corner, tapering, bilateral, no iris
occlusion. Graded on #683. So the pack stays and the overage is recovered by decimation.

Measured on HEAD, verified by direct survey of the four ED GLBs

mpfb-gown-adult-patient        total  69,547   lash  4,603
mpfb-clinical-nurse-adult      total  41,866   lash  3,237
mpfb-family-partner-adult      total  36,582   lash  3,239
mpfb-clinical-physician-adult  total  44,620   lash  3,297

ED TOTAL      192,615
ED LASHES      14,376
NON-LASH      178,239
BUDGET        180,000  ->  lash allowance 1,761 across all four

That is roughly 440 triangles per actor against ~3,600 today: an aggressive rung, not a comfortable
one. It must be measured and graded, not assumed to work.

THIS CARD ABSORBS #748, because the contract cannot pass otherwise

the-station-fits-its-budget-and-still-looks-like-people has three standing clauses and no planted
RED. Two are red: clause (1) on the budget, and clause (3) because
mpfb-gown-adult-patient ships 4 UV-bearing primitives where 16 are expected (#748, bisected to
c66c012c inside #714).

A run: on that file cannot pass while either stands. Both need a re-bake of the same four assets,
so splitting them would leave this card with a contract nobody could satisfy. Fix both.

The levers, and what is refused

Selective lash decimation as an ED-specific LOD. Do NOT replace the source pack —
mindfront_eyelashes_01 is already the smallest variant it offers. Do NOT spend already-graded
hands, face or garment detail to subsidise eyelashes unless a lash-only sweep fails. Do NOT raise
the 180,000 figure; it is authored at asset-generation-pipeline.md:87 as a hardware constraint.

If no lash-only result at or under 1,761 remains legible, say so and stop — that is a successful
finding. The named fallback is a hybrid supporting-actor LOD, which requires REPLANTING the
body-count counterweight in an-actor-wears-fitted-eyelashes-not-helper-stubs.test.ts, never
weakening it during execution.

The verdict this card cannot close by itself

My grade was on mpfb-ob-patient-aisha at the full 16,632-triangle source, which is not an ED
actor. After optimisation the orchestrator must grade an actual ED actor at the same 1024/0.80
framing via isolated-subject-harness.ts --head-crops. A machine clause cannot assert legibility.

Report these

If any proof cannot pass as written, or passes trivially, or asserts the opposite direction from the
defect, say so in your first report at the moment you find it. If satisfying a clause would make the
actors visibly worse, say so and satisfy it anyway. Name out-of-scope wrongness with the object and
what it looks like.

done_when

• run:pnpm exec vitest run tools/openclinxr/evidence/the-station-fits-its-budget-and-still-looks-like-people.test.ts
• run:pnpm exec vitest run tools/openclinxr/evidence/an-actor-wears-fitted-eyelashes-not-helper-stubs.test.ts
• changed:apps/ui-xr/public/generated-humanoids/mpfb-gown-adult-patient.glb

factory_step: body_param

[gidich]

The pixel grade this card needed now exists, recorded in full on #683. Summary: at a 1:1 native crop the fitted lashes are present, symmetric and distinct from the brow, but they render as a speckled stipple rather than legible lashes. So the 16,264 additional triangles per actor currently buy distinct geometry that does not yet read as eyelashes. That does not settle the trade on its own — the stipple may be material, alpha or normals rather than geometry — but the budget decision now has its missing input instead of resting on a triangle count alone.

[gidich]

CORRECTION — I listed three regressions and only ONE is new

This card said #683 caused three reds. Verified against the tree, two of them predate it.

Measured on the two GLBs directly, git show f82e1cc2^ versus main:

BEFORE #683:  74,642 triangles, 10 meshes
AFTER  main:  90,906 triangles, 10 meshes
delta:       +16,264

That delta is exactly the lash swap the worker predicted, 16,632 fitted minus 368 helper, and the
mesh count is unchanged. So the lash trade is the whole triangle change for this actor.

mpfb2-actor-is-stripped-of-helpers was ALREADY red. Its classifier reads 79,362 now against a
documented 26,756. Subtracting the same 16,264 gives 63,098 before #683 — still far over its 28,000
threshold. It counts fitted brows and retained facial features as body, so it was red before this
card existed. #683 worsened the number and did not cause the failure. That is a stale instrument
exposed, not a new regression.

every-actor-has-brows-lashes-and-teeth clause (3) was already red from #695's intentional ED
body decimation, which the #683 worker recorded as pre-existing and I repeated as new.

Only the station-budget clause is a verified new product regression: 178,969 to 192,406 against
180,000 authored, over by 12,406.

And the comparative pixel grade now exists

Recorded in full on #683. Control and treatment rendered through the same tool at the same
framing, cropped 1:1 at an identical box: the 368-triangle helper lashes read as a cleaner, more
defined upper-lid line than the fitted pack, which reads as speckle.

So the 16,264 triangles per actor are currently unpurchased. The confound is that the two bakes
differ by more than lashes — the control is 1.4 MB larger, and the worker's report names a
height-solve re-bake and four decimations — so this does not by itself justify a revert. What it
establishes is that no fidelity gain is visible at this framing.

What this card still needs

A decision among decimating the lash per actor, excluding lash geometry from the BODY measure the
way #321 already excludes garments, or replacing the pack. Not raising the 180,000 figure, which is
an authored hardware constraint.

[gidich]

Correction: the pixel verdict I gave this card is WITHDRAWN. I reported that the fitted lashes read as speckle and that the 368-tri control looked more defined. Measured on the capture itself, the iris spans 15 pixels in the 4096 frame and the lash band occupies three to five — individual strands are sub-pixel, so speckle is what any lash geometry looks like at that framing, including a correct one. I described texture from an image that cannot carry it, and that verdict was about to justify a revert. What survives: the structure pass shows dense lash geometry in the treatment that the control lacks, and the triangle delta is exactly 16,264 with mesh count unchanged. This card's budget overage of 12,406 triangles is independently verified and does not depend on the grade. A head-framed capture is the instrument that would settle appearance; the full-body 4096 is not.

[gidich]

The pixel verdict this card needs cannot currently be produced. Filed as #745: the isolated AO harness returns an error surface, head-focus-inspection.ts times out waiting for window.__openClinXrIsolatedSubjectEvidence on this exact actor, and the one harness that runs frames the whole body — where the iris measures 15 pixels and no lash verdict is available. head-focus-inspection.ts's own header records that this framing was already proven insufficient by #350. The capability was built and has decayed.

[gidich]

The appearance verdict this card's fork needed now exists, recorded in full on #683. At a head-framed 1024 capture the fitted lashes read as lashes: a strand-textured fringe on both upper lids, inner to outer corner, tapering at the outer corners, bilateral, no iris occlusion, clearly not an eyeliner band. Every criterion set for the verdict passes. What is still unmeasured is the SIZE of the improvement — I captured no control at this framing, so "legible" supports keeping the pack while the question of what 16,264 triangles per actor buys over the 368-triangle helper remains open. On the revert-or-keep fork this points to keep, and to resolving the 12,406-triangle overage by decimation or LOD rather than by reverting the trade.

[gidich]

Factory: Dispatched

IN-SCOPE: ED station back under its authored budget by lash-only LOD; gown UV regression (#748) traced and the dropped data restored.

• New tools/openclinxr/asset-pipeline/anny/optimize_glb_lash_lod.mjs: meshopt simplifyPrimitive scoped to /lash/i primitives only. The v4 simplify transform has NO primitives filter — measured: passing it the whole document decimated every mesh (gown 69,547 → 11,121). Scoping is the whole point.
• Rung (ratio 0.12, error 0.005) applied to the four ED actors, lash meshes only: gown 4,603 → 386, nurse 3,237 → 388, family 3,239 → 387, physician 3,297 → 395. Lash total 1,556 ≤ 1,761 allowance; ED total 179,795 ≤ 180,000 (margin 205). Bodies at recorded counts (14,928/12,892/12,884/12,890), brows/teeth/tongue, mesh counts and UV prims byte-unchanged. Per-actor rung reports: tools/openclinxr/evidence/issue-737/lash-lod/*.json.
• Clause (3) / the gown patient lost UVs on 12 of 18 primitives and the station clause that refuses that is red #748 traced: no committed bake path drops UVs. Measured: stage-2 import→export preserves all 16 on the decimated input, on the full-res input, and through the Wire meshopt decimation into the humanoid pipeline — the silhouette holds at 2.81x and the station comes under budget #695 meshopt rung (all 16 kept). The 4-UV shipped state is a one-off artifact of the c66c012 session (its on-disk input is unrecoverable; export options are identical across 57af536e/c66c012c/current). restore_gown_uvs.mjs restores the exact dropped TEXCOORD_0 attributes from the every shipped garment lost its base-colour texture — 22 of 22 materials render from a flat factor #740-era blob (a55067d) — hash-verified byte-identical POSITION on all 17 non-garment primitives, so the UVs map 1:1. The graded The bodice fold amplitude is 34 mm on a 10-22 mm standoff, so the wave valleys land inside the body #714 garment (29,185 tris) is untouched (it differs between the two states and carries no UVs in either).
• I did NOT re-bake the gown through stage-2: a re-bake from the decimated base produces a different garment (47,845 tris derived from the 14,928-tri body) that would need decimation to ~21,800 — spending garment detail the card forbids unless the lash-only sweep fails. The sweep did not fail (1,556 ≤ 1,761), so the restore path preserves the graded asset.
• Both contract suites green: the-station-fits-its-budget-and-still-looks-like-people 3/3; an-actor-wears-fitted-eyelashes-not-helper-stubs 5/5.
• Head crops for the orchestrator's lash grade (1024×1024, focus=head, same The eye-focus station frames the head only on MPFB and silently falls back everywhere else — blocking the hair direction comparison #358 product path): tools/openclinxr/evidence/issue-737/head-crops/ed-gown-patient-head-front.png, ed-clinical-nurse-head-front.png, ed-family-partner-head-front.png.

OUT-OF-SCOPE:

• Physician head crop REFUSED by the lab: focus=head unresolvable on mpfb-clinical-physician-adult — measured neck-constriction ratio 0.871 vs the 0.85 threshold (hair-excluded silhouette; the hair top extends the band). Same result in Node and browser; the lab's hair-exclusion (Fitted hair defeats head-box derivation: aisha returns null, and removing the hair mesh alone fixes it #394) is deliberate product code, not mine to change. Recorded in head-crops/refusals.json. Its 395-tri lashes are ungraded.
• Pre-existing reds, verified identical by stash (not caused by this slice): the-gown-shards (3) and the-gown-shell-clearance (4) pin the gown at the Wire meshopt decimation into the humanoid pipeline — the silhouette holds at 2.81x and the station comes under budget #695-era sha 7bd12d06… (7,116,988 B; ships as dd2340a9… 9,280,276 B since Eyelashes ship at 368 tris from hm08 helpers while brows ship at 35,334 fitted — the eyelashes01 pack is staged and unused #683); the-gown-mhclo-fits (1)(2)(3) reads .openclinxr-local/provider-cache/…/crudegown.mhclo which is gitignored and absent in this worktree; every-actor-has-brows-lashes-and-teeth (3) and the-eyebrows-cost… (2)(4) pin the pre-Wire meshopt decimation into the humanoid pipeline — the silhouette holds at 2.81x and the station comes under budget #695 body (26,756) and pre-Eyelashes ship at 368 tris from hm08 helpers while brows ship at 35,334 fitted — the eyelashes01 pack is staged and unused #683 lash cost (368); pnpm architecture markdown-references fails on unresolved refs in PROJECT_STATUS.md / doc-authority-registry pointing at gitignored/absent paths (no Markdown touched here).

CLAIM: ED four-actor station is 179,795 triangles (≤ 180,000 authored budget) with 1,556 total lash triangles (≤ 1,761 allowance) via lash-only meshopt LOD at (0.12, 0.005); the gown's UV primitives restored 4 → 16 from byte-identical #740 data; both named contract suites green; bodies, brows, teeth, tongue and mesh counts untouched.

NOT TESTED: whether the decimated lashes still read as lashes at conversational distance — text-only worker; the orchestrator must grade the head crops at the 1024/0.80 framing (the #683 verdict was on full-res aisha, not an ED actor). The physician's lashes have no head crop (lab refusal above). The rung's behavior under runtime skinning (weights were re-mapped by simplify) is unverified. The eyebrow re-export shrink (13,187→12,122) was observed in stage-2 round-trips but NOT shipped — only relevant if the gown is re-baked again.

[gidich]

ACCEPTANCE GRADE — the LOD is accepted, and the benefit case does not rest on this actor

Landed and verified by my own survey of the four ED GLBs, not taken from the report:

mpfb-gown-adult-patient        total 65,330   lash  386   uv 16/18
mpfb-clinical-nurse-adult      total 39,017   lash  388   uv 17/17
mpfb-family-partner-adult      total 33,730   lash  387   uv 16/17
mpfb-clinical-physician-adult  total 41,718   lash  395   uv 18/18

ED TOTAL 179,795 against the 180,000 budget  — under by 205
ED LASHES 1,556 against the 1,761 allowance  — under by 205

Both red clauses of the-station-fits-its-budget-and-still-looks-like-people are green.

The grade, on the WORST rung rather than a convenient one

The gown patient took the largest reduction — 4,603 to 386 triangles, 11.9x. Grading "an ED
actor" would have let the mildest be chosen, so this is the one captured, control and treatment, at
the same 1024/0.80 framing through renderHeadFocusCrops with a per-subject spec.

before (4,603 lash tris):  thin dark upper-lid line, no strand fringe
after  (386 lash tris):    thin dark upper-lid line, no strand fringe
difference:                NOT PERCEPTIBLE at this framing

The LOD cost nothing visible. It is accepted.

The part that matters more, and it qualifies my earlier grade

This actor's lashes were not legible before the decimation either. My earlier verdict — that the
fitted pack "reads as lashes" with a strand-textured fringe — was on mpfb-ob-patient-aisha at the
full 16,632-triangle source, and it does not transfer here.

So the fitted pack's demonstrated benefit rests on one non-ED actor at full resolution. On the ED
cast, at the framing a learner would see, the lash is a thin dark lid line at both 4,603 and 386
triangles. That does not undo the acceptance — 386 costs almost nothing — but it does mean the
station now carries a feature whose visible value on these four actors is unestablished.

Out-of-scope, seen and not fixed

• The rectangular face-texture seam is present on this actor too: lighter geometry outside a
  rectangular texture patch, with a hard vertical edge beside each cheek.
• The hair is a hard-edged shell with a straight cut across the forehead and a visible seam.

NOT TESTED

The other three ED actors at this framing. Whether the difference is perceptible in the runtime at
conversational distance rather than in an isolated 1024 head capture. Why aisha shows a strand
fringe and this actor does not at a comparable lash count.

[gidich]

The face-seam observation in my acceptance grade is WITHDRAWN. I recorded a rectangular texture patch with hard vertical edges beside each cheek. Measured: the interior luminance steps sit at x≈354 and x≈655 on every row of both actors, dark-to-light — that is hair meeting skin, and the vertical step at y=549 is the fringe. There is no interior seam; the rectangle is the gap between the bob haircut, not a boundary in the skin texture. Every triangle-count, UV and budget finding in that grade stands — those were measured. Only the seam observation is withdrawn, and it did not reach a card.