test(cliffs): the field is RIGHT - the fine sweep refutes #106's conclusion (#84)

[wormeyman]

#106 was wrong about the attribution, and this corrects it. That PR concluded the port's grid-4 cliff-elevation field is wrong at [1500,1500]'s high bands and handed over "it lives in the mountains branch of vulcanus_elev". The fine sweep was run to quantify that. It refutes it.

The measurement

Sweeping cliff_elevation_0 across [700,900] step 5 under the same collapsed rule turns each placed cell's orientation into one-sided constraints on its corners - a crossing at L says "this corner > L, that one < L", and the sign says which. 41 levels bracket a corner to the step.

Only positive observations are used, which is what makes it sound: an absent cliff is ambiguous (the lava/ore rejections drop whole cells) but a present crossing is not, because fixImpossibleCellsSweep only ever writes 0 - it can delete a crossing, never invent one - and the rejections never touch the edge registers.

corners with a two-sided bracket	998
mean bracket width	5.72
port's value INSIDE the game's bracket	996
worst miss of the other two	6.7e-4
disputed-edge corner slots with a bracket	26 of 72
of those, containing the port's value	26 of 26

The two misses are the port sitting on a swept level, where crossesCliff's strict dA < 0 && dB > 0 yields no crossing and so no observation - the bracket's open endpoint, not an error. At 4e-4 of a 5-unit bracket there is no room for it to be anything else.

The field is exonerated by direct measurement, and the (i*4, j*4) lattice is re-confirmed on the way: a wrong sampling site could not put 996 of 998 values inside 5-unit brackets.

Where #106 went wrong

It established what the residual is NOT - not smoothing, not the gate, not the repair, not the rejections, not a boundary tie - and then treated the field as the last man standing. But it never measured the game's field; it measured that OUR value sits a median 18.8 from the level at the disputed edges, which is a statement about our field. "Everything else is excluded, so it must be X" is only as good as the list, and the list was not closed. The lower bound was real; the attribution was not. Same shape as #88, where the best-scoring model was the wrong one.

What is actually left

Across all 41 levels the game's cell code is the port's with edges removed in 1231 of 1235 disputed cells - the port finds crossings the game does not, essentially never the reverse.

The lead is the coverage number: only 26 of 72 disputed corner slots get a two-sided bracket, and 661 of the 1,659 corners whose value lies in [700,900] get none - a corner goes unbracketed when the game emits no entity beside it at any level, which is what the rejections do to whole neighbourhoods. One chunk at L=790 shows it directly: cell 1634,1706.5 keeps its tight edge (714.4 → 795.9, margin 5.9) and loses its wide one (721.0 → 875.4, margin 69), and the cell sharing that wide edge is absent from the game entirely.

That points at the emission/rejection path rather than the field. Stated as a hypothesis from one chunk plus coverage statistics, with the control named: whether disputed edges sit adjacent to rejected cells above the base rate of all crossing edges.

Determinism: the two fixtures overlap at L=790 and agree 494/494.
Vacuity: shifting the field by 10 fails 3 of the spec's 5 tests.

Item of #84; the issue stays open.

🤖 Generated with Claude Code

https://claude.ai/code/session_011DckiyMmb4zDPBV8NSyxTg