test(cliffs): the cross-chunk cascade does NOT explain the border enrichment (#84)

[wormeyman]

#148 raised the missing cross-chunk cascade as the first candidate that predicts border-only errors rather than merely being consistent with them. This tests it directly against the residual, on the 14 regions the enrichment is measured over. No capture.

It loses.

model	unexplained	on border	z
shipped (rejectAtCrossingStage, chunk-local)	25	19	2.99
cross-chunk destroy cascade	23	17	2.67

It explains 2 of the 25 - both border cells, so it does bite exactly where the signal lives - and leaves a residual still enriched at z = 2.67. A mechanism that explained the enrichment would drive that toward the 46.2% base rate. It does not come close. The border enrichment remains open.

At the plain surplus level the same run gives 90 → 84, and all 6 cells the cascade removes are on chunk borders - #148's border-exclusivity reproduced on 14 regions rather than 3.

The answer was already in committed data

cliffResidualCascadeAudit's published "after cascade" row - 23 unexplained, 17 on border, z 2.67 - already had a cross-chunk cascade in it. That harness works on a flat cell map with a 64-tile halo and never restricts propagation to a chunk, so its post-cascade number was, unrecognised at the time, exactly the number that refutes the candidate. It was sitting in the repo a day before the candidate was proposed.

The refutation was available by asking what model the published number was already computed under - a cheaper question than the experiment, and worth asking first next time.

What was genuinely missing was the other row: nobody had ever scored the residual under the model that actually ships. That is the 25 / 19 / 2.99, and the 25 → 23 delta is the precise size of the mechanism's claim on the residual. So the test still earned its keep.

The control is load-bearing

The cascade row reproduces the published 23 unexplained, 17 on a border on the same 14 regions. Without it, the shipped row could be measuring a different quantity and the 2-cell delta would mean nothing - both models have to be scored by one definition of "unexplained", which here is #143's: a cell the game killed that neither our own kill set nor the game's own ore lever accounts for.

Bookkeeping

#148's spec header updated to carry this outcome, so its candidate claim does not stand unqualified for the next reader.

Where #84 stands

The border enrichment has now survived three attempts: the orientation-reach rival (#134), cascade double-counting (#143), and the cross-chunk cascade here. It is 23 cells at z = 2.67, and there is no live mechanism for it.

Verification

pnpm run verify green - 1570 app (up 4) + 13 worker + 3 container tests, 0 warnings, exit 0. Measurement only; shipping accuracy unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HKjzpWJw1HbnigPP9udqN2