The audits re-run at real density

Every previous verdict was measured on 18 buildings from the fixed estate. Density was the variable that could not be tested, because the estate did not respond to input. It responds now.

tests/density-audit.ts, run against a generated world (113 buildings, 43 districts, 1,327 links).

---

Result

	18 buildings	113 buildings
default vantage occludes	11.1%	50.4%
unreachable from any pose	0	0
labels misattached at default	0	16.8%
worst single pose	2	100
labels misattached at every pose	0	0
position: proximity reversals	—	54.5%
representation frontier	matrix	matrix, still derived

The verdicts survive in kind and degrade severely in degree.

Nothing became unrecoverable. Everything became much harder to recover. Half the estate is hidden at first sight, and at the worst pose every label on screen is attached to the wrong building.

The correction I nearly published

The first run reported 12 buildings unreachable from every pose — which would have been a renderer claim failure, the most serious category available.

It was my sampling. 108 poses at 20° azimuth, 3 elevations, 3 distances:

432 poses (20°, 6 elevations, 4 distances) → 1 unreachable
1728 poses (5°)                            → 0
4320 poses (2°)                            → 0

A sampling artifact, refuted by refining the instrument rather than by argument. The camera verdict stands: exploration recovers everything.

This is the second time in three rounds that the first number was an artifact — after the collinear fixture. Both were caught by testing the instrument rather than the subject.

What actually degraded

Occlusion, 11.1% → 50.4%. Recoverable and expensive. A viewer must now move to see half the world, where before they had to move to see a tenth of it.

Labels, 0% → 16.8% at the default vantage. This is the sharper result, because a label is a semantic element: at density it names the wrong building one time in six on first sight. Still recoverable — zero labels are misattached at every pose — but "recoverable" now means "the viewer must orbit to disambiguate every sixth name".

Position, unchanged at 54.5%. The proximity claim is as bad at 113 buildings as it was measured before: worse than a coin flip, so the arrangement carries no information. Density did not make this worse because it was already at floor.

Representation, unchanged. The matrix remains uniquely non-dominated for a symmetric binary relation.

---

Standing

Measurable defect exposed?	no — the candidate was an instrument artifact
Measurable degradation quantified?	yes — occlusion 4.5×, label misattachment 0 → 16.8%
Prior verdicts overturned?	no, and they now carry numbers at real density

What is still not established

• Label-to-label overlap is still unmeasured, and it is now the most likely real failure: at 113 buildings the labels almost certainly collide with each other, which is a different defect from misattachment and needs a label width.
• The "recoverable" verdict is doing more work than it used to. Recoverable at 11% occlusion and recoverable at 50% are the same word for very different experiences, and this framework has no measure for effort-to-recover.
• Occlusion remains centre-point only; a building 90% covered counts visible.
• One repository. xi-hq was not audited at density, only generated.

231 tests, 26 DOM, tsc clean.