test(cliffs): the two WEST observations are independent (#84)

[wormeyman]

#150 localised the border residual to the west chunk edge; #142 separately found all six ore-recall cells with their nearest resource to the west. If those were one phenomenon, two open threads would collapse into one.

They are not. No capture - scored against the game's own dumped resource entities, not the port's model.

The west edge is not about ore

	within 32 tiles of a resource	nearest resource lies WEST
base (8,393 raw cells)	21.3%	47.1%
west-edge survivors (9)	2 of 9	44.4%
other survivors (15)	3 of 15	60.0%

West-edge survivors are no closer to resources than the other survivors, and their nearest resource points west slightly less often than the base rate. Whatever produces the west-edge concentration, it is not the ore.

A signal that looks real at 3 sigma, and is not

At a 64-tile radius the survivors do look resource-proximate: 70.8% against a 40.2% base, z = 3.06. Two things kill it, and both are asserted in the spec so it cannot be rediscovered and believed:

• The scale is wrong for a cell-level cause. Nothing appears at 16 tiles (z = 0.81) or 32 (z = −0.05). An effect absent at short range and present only at 64 tiles is describing a region, not a cell.
• n_eff is ~9, not 24. The 24 survivors sit in 9 regions, and the within-64 outcome is nearly all-or-nothing per region - four regions have ALL their survivors inside 64 tiles, two have NONE. The cells are not independent draws, so a binomial z over 24 of them is inflated.

That is the same failure as the old ore/cliff "100× below chance" claim, which assumed tile independence. Worth catching twice.

Where #84 stands

Five mechanisms are ruled out for the chunk-border residual - orientation-reach (#134), cascade double-counting (#143), the cross-chunk cascade (#149), updateConnections (#150), the sweep's edge order (#151) - and the ore is now ruled out as the cause of its west signature.

The west concentration itself stands at z = 2.6-2.9 across every sweep-order arm, on 17-24 border cells over 14 regions. That n has never been raised. Raising it out of sample is the honest next step before another mechanism hunt - this repo has been burned before by a signal that looked solid at n=14 and needed raising rather than slicing.

Verification

pnpm run verify green - 1584 app (up 6) + 13 worker + 3 container tests, 0 warnings, exit 0. Measurement only; shipping accuracy unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HKjzpWJw1HbnigPP9udqN2