test(cliffs): the border enrichment is NOT a cascade artifact (#84)

[wormeyman]

The ore thread's two wins (#141, #142) came from auditing populations rather than hunting mechanisms, so the border residual gets the same treatment.

The obvious deflation — after #142 showed four of ten ore "misses" were cascade casualties — is that the unexplained cells are largely casualties too. The residual is counted at the output (cells the game killed that our predicates don't), so a casualty of a correct kill scores as an independent defect.

It is not that

Applying the destruction cascade — confirmed against the game in #139 and again in #141 — to the port's own kill set:

	unknown	on border	z	false rejections
before cascade	33	22	2.36	12
after cascade	23	17	2.67	14

The cascade explains 10 of the 33, and the enrichment concentrates rather than dissolving — the share rises too, so it is not the z moving on sample size alone.

The border signal has now survived both of its plausible deflations: the orientation-reach rival (#134) and cascade double-counting here.

A shipping gain, measured and deliberately not taken

The same run says the port is leaving cells on the table: 10 fewer missed for 2 more false rejections, net 8. renderVulcanusCliffs does not cascade at all today — applyCliffConnections exists but only specs use it.

It is recorded rather than adopted on purpose, because adopting it:

• changes rendered orientations as well as positions,
• needs cliffErrorBudget.spec.ts moved in lockstep (that file's header records the day it drifted), and
• applyCliffConnections additionally bundles the updateConnections model, which is an explicit upper bound and is not what this audit applied.

Look at the 2 new false rejections before the 10 wins. They are cells the game KEPT that our cascade removes, and #134 recorded a gate the port does not model: Cliff::destroyEnd refuses to forceDestroy when entity flag bit 4 of +0x6e is set, returning with the orientation UNCHANGED rather than merely undestroyed.

Scope, stated because the headline differs from the published 44

This pairs 14 regions across three fixtures, not the 15 the 44-cell figure covers — [1500,1500]'s ON/OFF pair lives in a fixture with a different case shape.

The overlap is the correctness check, and it passes: on the eight-region border batch this harness reproduces 17 unknown, 12 on a border, exactly what cliffResidualBorderEnrichment publishes for that batch. Without that, every other number here could be measuring something else.

Verification

pnpm run verify green — 1554 app (up 5) + 13 worker + 3 container tests, 0 warnings, exit 0. Measurement only; shipping accuracy unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_012VRsPwmSwjnFKVSS6jx9Mo