test(cliffs): falsify the smoothing, and close fixImpossibleCells (#18)

[wormeyman]

Chases the Vulcanus cliff_orientation residual (#18). Does not fix it - it falsifies nine more candidates, closes the one the notes left open, and localises what remains precisely enough to say what the next step is.

The harness

cliff_smoothing is a transform on the elevation field, so a candidate smoothing can be pre-applied to the field and the placement then run with smoothing: 0. That exercises the shipping rule over a candidate transform with no edit to cliffPlacement.ts, and one shared field cache makes each arm ~1.5s for all three regions. The control arm reproduces 1400 matched / 175 wrong to the digit, which is what licenses everything below.

What was falsified

candidate	result
smoothing knot geometry	sharp optimum in 4 dimensions at span 4 / clamp 7 / offset 0 / s=1
clamp 8 ("clean" bilerp) / clamp 6	33.6% / 40.8% vs 12.5%
any anchor shift	51-72%
interpolation curve	linear beats smoothstep (29.3%), nearest (61.8%), sqrt (50.5%)
smoothed cliffiness	20.0% - it stays raw
cliff_elevation_0, cliff_elevation_interval	bottom out exactly at 70 and 120
crossesCliff structural variants	none help; "different band" scores exactly 175 because it is provably identical
fixImpossibleCells clear order	ported L,T,R,B is the best of all 24

Two of these are worth more than a row. The clamp-to-7 asymmetry that reads like a misread is 2.7x better than the clean reading, so the disassembly is corroborated by something other than itself for the first time. And fixImpossibleCells is closed by size: across all three regions it finds only 35 illegal cells and clears 59 edges, so it is too small by 5x to explain 175.

The fields were compared as values, not outcomes

Every previous exoneration substituted the game's fields and asserted the placement did not move - weaker than it sounds. Compared directly over all 12,675 corners:

	median	max
elevation |ours - game|	2.7e-4	4.8e-2
cliffiness |ours - game|	0	6.4e-6

The corrections the failures need are ~3.6 elevation units - four orders of magnitude larger.

The new spec

test/cliffEdgeConsistency.spec.ts uses no field, no rule and no port output: the game's own adjacent cliffs must agree on the edge they share, because the engine stores one value per edge. They do - Nauvis 147h+166v, Vulcanus 805h+834v, zero mismatches - which clears the code packing, the position→cell lattice and the orientation table from the fixture alone. It carries a planted table corruption to prove it discriminates, and asserts the 8 off-lattice crater-cliffs so the lattice check cannot pass vacuously.

What remains

Over the 5891 edges where the cliffiness gate is open, our smoothed elevation predicts the game's crossing 88.7% of the time and the raw one 60.3%. So smoothing is doing real and largely correct work, and the last ~11% is a difference between the game's smoothed cliff elevation and a linear bilerp of the raw field - at corners where the raw field agrees to 0.05.

The next step is the binary, not another behavioural sweep. That space has been searched and it bottoms out here; crossingsForChunk @ 0x10160c9cc wants re-reading for its smoothing loop specifically.

pnpm run verify: 306 files clean, 1288 passed / 3 skipped, preview 12 passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GRx1CU29RKLsRuUAErEDRB