Labels measured — recoverable, and the fixture refused the real question

Fourth result under renderer-as-laboratory. Labels are the first genuinely semantic UI layer: the text carries identity, and WorldScene.tsx:352 puts failing on the label colour. Unlike the camera, a label can make a claim.

tests/label-attachment.test.ts.

---

The task

Given a label on screen, can a viewer attach it to the right building?

Two failure modes, both computed in closed form and neither needing a label width — which matters, because distanceFactor scaling is a drei internal and approximating it would put the result at the mercy of a guess:

• Misattachment — the label's anchor falls inside a nearer, different building's silhouette, so a viewer reads the name as belonging to that one.
• Orphaning — occlude={false} (:348) draws every label through geometry, so a label for a hidden building floats with no referent.

Result

18 buildings · 2 occluded · 18 labels projected

misattached at the default vantage:  0
orphaned (referent hidden, drawn):   2/18  (11.1%)
108 poses · worst single pose:       2 misattached
labels misattached at EVERY pose:    0

Misattachment is real and recoverable. Up to 2 labels sit on the wrong building at some poses; none does so at every pose. Same verdict shape as the camera — a navigation cost, not a claim failure.

Orphaning is not a defect on its own. A floating name may be the honest way to say "something is here you cannot see". It becomes a false claim only if it lands on another building, which the misattachment test covers separately.

Element	Semantic ownership	Perceptual effect
label text	derived — identity	preserves — 0 misattached at every pose
label colour	derived — failing	unmeasured; shares ALERT with the beacon
label anchor height	arbitrary (h + 4.5)	preserves
occlude={false}	arbitrary	preserves — orphans 11.1%, none misread

---

The finding the fixture produced instead

The density question — do labels collide when the estate is crowded? — could not be asked. Scaling the input from 9 to 80 repositories:

repos=9   buildings=18   misattached=0
repos=20  buildings=18   misattached=0
repos=40  buildings=18   misattached=0
repos=80  buildings=18   misattached=0

The building count does not respond to the input at all. Tracing it: deriveStructure iterates PRODUCTS (src/lib/hq/estate.ts:82) and AGENTS (src/lib/hq/roster.ts:67) — fixed constant tables. The repos argument only supplies stats for products that already exist, matched by repos.find((r) => r.slug === product.repo).

The world does not construct itself from the repository. It is a fixed estate of predefined products and agents, onto which repository statistics are attached.

That is a much larger finding than the label result, and it bears directly on the stated milestone — "point Cognithamus at a real repository, watch the world construct itself". Today, pointing it at a different repository changes the heights and colours of a predetermined city. It does not change what buildings exist.

The repository adapter built in repository-reconstruction.md does produce subjects from arbitrary input — 117 of them from an unprepared repo — but nothing connects those subjects to deriveStructure. The two halves exist and do not meet.

This was not derived. The fixture rule forced it: can the fixture express the phenomenon? It could not, and the reason it could not was the finding.

---

Standing

Measurable defect exposed?	yes — the estate is constant, not derived from input
Measurable improvement?	none made; measurement first
Freedom removed without evidence?	no
Fixture rule applied?	yes, and it produced the result

What is still not established

• Label-to-label overlap is unmeasured. Two names on top of each other is a distinct failure from misattachment, and it needs a label width — i.e. drei's distanceFactor semantics, or a rendered surface.
• Label colour is unmeasured. It carries failing and shares ALERT with the beacon; whether two carriers of one fact is sufficiency or competition is the beacon investigation's question.
• Density remains untested, and it is the condition under which labels are most likely to fail. It cannot be tested until the estate responds to input.
• 108 poses at 20° steps; a misattachment could hide between samples.

219 → 223 tests, tsc clean.