every appearance-grading harness is broken, and four cards are blocked on a pixel verdict

[gidich]

Three capture harnesses that produce appearance evidence all fail, and every card waiting on a pixel
grade is blocked behind them.

Measured

harness	subject	failure
capture-room-ao-isolated.ts (#716)	rooms	four byte-identical 8,904 B frames reading WebXR unavailable; provisioning the gitignored cagematch tree did not change it
head-focus-inspection.ts	mpfb-ob-patient-aisha.glb	page.waitForFunction: Timeout 120000ms exceeded — isolated-subject.html loads (domcontentloaded succeeds) but window.__openClinXrIsolatedSubjectEvidence never appears
model-vetting-glb-grade-capture.ts	any GLB	works, but frames the whole body and takes only --glb and --output-root

The third is the one that runs, and it is the wrong instrument for a face: on aisha the iris measures
15 pixels in its 4096 frame, so a lash band is three to five and no quality verdict is available
from it.

What is blocked behind this

• #683 — its own brief says "Reviewer-graded pixels decide whether this landed", and the verdict
  I gave from the whole-body capture had to be withdrawn as below the resolvable scale.
• #737 — the revert-or-keep fork for the lash trade needs that same verdict.
• #526 — landed and ungraded since the room bake.
• #643 — the showcase, which sequences behind #526.

The repo already knew this framing is insufficient

head-focus-inspection.ts's own header states it: "the Anny head lands at ~100 px in a 1024 px
frame under whole-subject framing — a framing this loop has repeatedly proved cannot support a face
verdict (#350: 10–20 px slivers hid in exactly that frame)."

So #350 established the principle, #354 built an eye-focus station, #358 measured it — and the
station now times out. The capability was built and has decayed.

claimScope / notEvidenceFor

• claimScope: three harness invocations on this machine today, and the pixel measurement of the
  one that runs.
• notEvidenceFor: a shared root cause. Two of the three fail differently, and I have not traced
  either. #715 fixed a ui-xr boot regression earlier; whether these are downstream of that, of a
  stale dist, or of something else is NOT DETERMINED.

No ## done_when yet: the remedy depends on which harness is worth repairing, and naming a contract
before that is chosen would encode a guess.

[gidich]

CORRECTION — the premise of this card is too broad and is withdrawn

I titled this "every appearance-grading harness is broken" after testing ONE head-framed path
(head-focus-inspection.ts) and seeing it time out. That is a generalisation from a single failure,
and it is wrong.

A working head-framed path exists and produced a verdict in 3.9 seconds:

pnpm -s exec tsx tools/openclinxr/evidence/isolated-subject-harness.ts --head-crops

.openclinxr/evidence/head-focus/mpfb-head-front.png
1024x1024, frameSpanFraction 0.80, 1 dev-server boot, 1 browser launch

The head box is derived from the silhouette width profile cut at the neck — 40,855 matched vertices,
neckPositionMeters 1.4631 — not from literal coordinates. At that framing the lashes resolve as
strands and #683's appearance question is answered, recorded there.

There is also a second close path I had not tried: _isolated-humanoid-lab supports
framing=face|head|closeup via __isoSetFraming, and humanoid-vision-score.ts already drives it
at two framings.

What survives, narrowed to what was actually measured

Two harnesses fail, and neither failure was traced:

• capture-room-ao-isolated.ts returns four byte-identical 8,904 B frames reading WebXR unavailable. Provisioning the gitignored cagematch tree did not change it.
• head-focus-inspection.ts times out after 120 s waiting for
  window.__openClinXrIsolatedSubjectEvidence; isolated-subject.html reaches
  domcontentloaded and the global never appears.

Both are real. Neither supports the claim that grading is impossible, because a third path works.

The cards I said were blocked

#683 is graded. #737 has its verdict. #526 and #643 are room captures, and the AO harness
failure above is specific to them — but pnpm asset:ui-xr:environment-room-capture produced 15
usable room images earlier in this run, so those are also less blocked than I claimed.

The lesson, which is mine

Two failures and a title. The rule this loop keeps relearning is that a single failing instrument is
evidence about that instrument, and I wrote it as evidence about a capability. Retitling would hide
that; the correction stays here.