fix(cliffs): sample the fields at the game's lattice - Nauvis is now EXACT (#18)

[wormeyman]

The port sampled cliff_elevation and cliffiness at (i*4, j*4 + 0.5), adding the cliff prototype's grid_offset {0, 0.5} to the sample position. That offset is a centre offset. The game samples the bare (i*4, j*4).

Result

	recall	precision	ratio
Nauvis seed 123456	0.943 -> 1.0000	0.943 -> 1.0000	1.000
Nauvis seed 777771	0.942 -> 1.0000	0.942 -> 1.0000	1.000
Vulcanus [0,0]	0.792 -> 0.806	0.685 -> 0.681	1.184
Vulcanus [1500,1500]	0.870 -> 0.938	0.719 -> 0.779	1.203
Vulcanus [-1200,800]	0.803 -> 0.853	0.866 -> 0.912	0.935

Nauvis is exact - 282/282 and 52/52 against find_entities_filtered, nothing missed and nothing invented. That is the ~6% residual which has been open since 2026-07-20 and had five separate causes proposed and falsified for it (fixImpossibleCells, wouldCollide, field precision, boundary proximity, the smoothing knot model).

Vulcanus improves on recall in all three regions but its over-placement is unchanged - 1.18x/1.20x/0.94x - so #18 stays open, now scoped to that rather than to recall.

Evidence

Two independent sources, neither of them our own tests:

• The binary. CliffGenerator::crossingsForChunk (arm64 0x10160c9cc) reads grid_size at [proto+0xb60]/[0xb68] and never grid_offset at [0xb70]/[0xb78]. Sample origin is chunkPos << 5 converted to float; step is grid_size. It computes no other coordinate - it consumes NoiseCache registers.
• The game data. base/prototypes/entity/entity-util.lua:305: "cliffs are auto-placed with centers at (0, 0.5) offset from the grid".

Then confirmed by capture: 12,675 corners re-sampled at the game's lattice, seed 123456, Factorio 2.1.12 with refs:sync --check in sync, scored against the game's 885 real cliff-vulcanus entities. A control arm using the old fixture reproduced the shipped numbers exactly, which is what makes the comparison readable rather than a harness artefact.

Why it hid for two months

The error moves no placed cliff. Cell centres are derived from their own constants, so x mod 4 == 2 and y mod 4 == 2.5 held, the preview-agreement check held, and PR #57's substitution of the game's own field values moved zero cells - because that fixture had itself been captured at the port's assumed site (capture.ts hardcoded the same 0.5).

A substitution test can falsify a value. It cannot falsify the site it was sampled at. Both "our values are right" and "our sample site is wrong" produce zero flips.

Fixtures

The Vulcanus corner-fields oracle is re-captured at the game's lattice. The previous capture is kept as -legacy-y0.5: it is valid ground truth for the site it names, and keeping it is what lets a future spec demonstrate the blind spot rather than just describe it. Both carry provenance entries.

Test changes are re-measurements, not accommodations

• cliffPlacement.spec.ts now pins Nauvis at equality rather than >= 0.85. Placement is deterministic given the seed, and slack is how a 0.943 sat unexplained for two months.
• cliffResidual.spec.ts's "the wrong cells sit on band boundaries" block asserted the residual exists. There are no wrong cells left, so it now asserts the set is empty (and still partitions on the same distance function, so an empty set cannot mean a loop that never ran).
• cliffSmoothing.spec.ts's plane gains a .37 offset. At j*4 the old constant made corner elevations land exactly on band edges - measured: all 7 disagreeing cells at distance 0.000e+0, with the bilerp computing 7.105e-15 - so the test was measuring float rounding, not smoothing. It passed before only because the old +0.5 knocked the plane off the boundaries by accident.
• vulcanusCliffs.spec.ts's independent blend now takes corner indices instead of deriving positions from centres - the exact mistake being fixed.
• vulcanusOreCliffSeparation.spec.ts: full-ore false positives 47 -> 46, per-region split shifted. The game still places zero there, so that finding is untouched.

pnpm run verify green: 1255 passed / 3 skipped, worker 12, container 3.