fix(cliffs): multisample offsets are in GRID UNITS - #18 root cause

[wormeyman]

Root cause of #18, found and fixed.

vulcanus-multisample-NOTES.md proved multisample(e,dx,dy) == e(x+dx, y+dy) at 150/150 comparisons, residual exactly zero. That is correct - for calculate_tile_properties, whose noise program has a 1-tile grid. It was never checked in any other channel, and the primitive's own docs say it evaluates "in a separate noise program with a larger grid" whose "sub-grids are copied to the main program". That phrase is load-bearing.

Asked through the cliff generator (grid = the 4-tile corner lattice) by routing a probe onto cliff_elevation with the rule collapsed, so cliffs mark exactly where the field crosses 71:

probe	column
x	70	baseline
multisample(x, 0, 0)	70	identity
multisample(x, 4, 0)	54	shifted 16 tiles, not 4
multisample(x, 0, 4)	70	null control

dx = 4 moves the field by 4 × the 4-tile grid step. The true rule is e(x + dx*G, y + dy*G) for the calling program's grid G.

So vulcanus_basalt_lakes_multisample's min over {0,1}² spans 4 tiles for cliffs and 1 tile for every per-tile consumer. min is erosion, so the cliff channel's elevation is far smoother. The port used the 1-tile field for both - too rough, over-placing ~40%.

Result

	before	after
wrong orientation	175 = 12.5%	37 = 2.4%
recall	0.806 / 0.938 / 0.853	1.000 / 0.973 / 0.965
[0,0] (worst region)	29.8% wrong	2.5%, recall 1.000
level sweep ratio	1.20-1.49 below elevation 120	1.00-1.09 at every level

VulcanusElevation now exposes cliffElevation beside elevation. Both hang off one stack and share every sub-expression below the multisample, so the cost is a second memo table - a private DAG for cliffs was tried first and is much dearer. Do not collapse them back together: they are different fields, not a cache miss.

Nothing here refutes the multisample port or the per-tile consumers - calculate_tile_properties and the tile renderer both live in the 1-tile channel where e(x+dx, y+dy) is exactly right.

Specs are inverted, not deleted

The inversions are the evidence: the s=0 smoothing arm is now exact, all four collapsed arms are exact, the level sweep matches at every level, and the corner-fields substitutions now move cells - those fixtures are the tile channel, so they are the right numbers for the wrong consumer, and their agreeing for months is how this stayed invisible. cliffResidual's Vulcanus block is retired: its mismatched population is now too small to compare.

The lesson

The third form of the same trap: not a fixture captured at the wrong site (#70's grid_offset), not a wrong value - a fixture captured through the wrong channel, agreeing with a port that made the same mistake. Ask which code path consumes a value, not only which coordinates it is sampled at. And no sweep inside the port could have found it: the smoothing, band and rule sweeps all searched a family that shared the defect.

Local test note

pnpm vp check: 310 files clean. The full suite currently shows 8 timeout-only failures on clean main as well - verified by stashing and re-running - after hours of Factorio captures loaded this machine. No assertion failures on either. CI is the arbiter.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GRx1CU29RKLsRuUAErEDRB