test(cliffs): the orientation residual IS part of the over-placement (#84)

[wormeyman]

The wrong orientations and the surplus cells have been tracked as two separate defects since #83. They are one defect.

The structure

placedCells builds one edge register per chunk, so v[cy][cx] is cell cx's left edge and cell cx-1's right edge - the same array slot, not two equal values. A spurious crossing therefore cannot be confined to one cell. It corrupts the orientation of the real cell on one side, and on the other it manufactures a whole cliff the game never placed.

Measured over all three oracle regions, rejections off so the geometry is not masked:

matched cells	1531
wrong orientations	37
of those whose disputed-edge neighbour the GAME places	0
distinct phantom neighbours	34
of those the PORT places, i.e. that are surplus cells	34 of 34

On the shipping path:

region	matched	wrong	surplus	surplus that ARE phantoms
[0,0]	281	5	2	2 of 2
[1500,1500]	858	25	22	10 of 22
[-1200,800]	386	3	1	0 of 1

At [0,0] the spurious crossings are the whole of the over-placement. And the reason 33 wrong cells do not imply 33 surplus is that the lava and ore rejections already remove 19 of the phantoms - the rejection hides the phantom while leaving the neighbour's orientation wrong, which is exactly how the two counts drifted apart and came to be read as unrelated.

What it changes

Not the design of the blocked grid-4 cliff-elevation capture - its value. That capture was being weighed against "33 wrong orientations, ~1.6% of cells", which reads like a rounding-error chase. It is also worth 12 of the 25 surplus cells, and all of [0,0]'s.

Ruled out on the way, each with a discriminating control

• No gate in crossesCliff is marginal on the disputed edges. Scored against every crossing edge of every matched cell as control: cliffiness gap above > 0.5 (disputed min 6.8e-3 vs control 1.0e-3), the a < 0 || b < 0 early-out (disputed min elevation 15.6), boundary < e0 (disputed min 2.73 above e0). Corroborates test(cliffs): the orientation residual is not a boundary tie (#84) #102's band-margin result by a different route.
• The asymmetric smoothing knot lattice is not a misreading. In-chunk knots at 0, 4 and 7 (hi clamped to CHUNK_CORNERS - 1) is odd enough to look like an off-by-one. The natural alternative - knots every 4 corners globally - is 8x worse: 312 wrong against 37, matched 1531 -> 1173, missing 38 -> 396. Raw (s = 0) is worse still at 677. The disassembly reading now has a measurement behind it.
• cliffiness_basic stays exonerated, and soundly: it contains no multisample, so unlike elevation it has no grid-4 variant to have been captured in the wrong channel.

Verification

pnpm run verify green - 160 files, 1369 tests. The new spec was checked for vacuity by pointing the neighbour lookup at the wrong axis: 4 of its 5 tests fail.

Item of #84; the issue stays open.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FpXiK2MPqPFVtm5CMPNXR2