test(cliffs): the cascade's whole gain is CROSS-CHUNK, and +8 was the wrong baseline (#84)

[wormeyman]

Priced the destroy cascade against the model that actually ships, rather than the post-filter baseline #143 used. No capture - scored on the error budget's own three regions.

model	port	surplus	missing	wrong orientation
shipped (rejectAtCrossingStage)	1547	22	6	21
destroy cascade	1545	20	6	18
cascade, forbidden to cross a chunk	1547	22	6	21

#143's net +8 does not survive, and the real figure is +2

The +8 scored the cascade against a post-filter model - kills applied, no cascade. But renderVulcanusCliffs ships rejectAtCrossingStage, which zeroes a rejected cell's four edge registers so its neighbours lose the shared edge, and that already reproduces most of the cascade. The honest gain is +2 positions and +3 orientations, recall untouched.

Corrected at both sites that quote it - cliffResidualCascadeAudit.spec.ts's header and the notes banner - rather than only in the new spec, since the banner is what a reader hits first.

The third row is the finding

Restricting the cascade to within a chunk reproduces the shipped model exactly, on all four counts. Everything the cascade buys is a cascade that crosses a chunk boundary - which rejectAtCrossingStage cannot do by construction, because each chunk owns a private copy of its shared edges. That privacy is precisely what keeps the pass chunk-local and worker tiling byte-identical.

All six disagreeing cells are onChunkBorder, and every one the game has an opinion about is a correction:

kind	cell	shipped	cascade	game
removed	(1606, 1598.5)	present	gone	absent (surplus)
removed	(-1050, 1022.5)	present	gone	absent (surplus)
reoriented	(1634, 1534.5)	west-to-east	none-to-east	none-to-east
reoriented	(1694, 1598.5)	north-to-east	north-to-none	north-to-none
reoriented	(1670, 1662.5)	south-to-west	none-to-west	none-to-west
reoriented	(1722, 1630.5)	east-to-south	east-to-none	not in game set

Why this is worth more than the two cells

That is the same place the residual's unexplained cells concentrate. The chunk-border enrichment has survived both of its plausible deflations - the orientation-reach rival (#134) and cascade double-counting (#143) - and stands at z = 2.67.

A chunk-local rejection model that cannot cascade across a chunk boundary is a mechanism of exactly that shape, and it is the first candidate that predicts border-only errors rather than merely being consistent with them.

It does not explain the enrichment yet - 6 cells here against 23 unexplained there, on different fixtures - and the spec says so. What is established is that the mechanism exists and is border-exclusive, which is what makes testing it against the residual directly the next move.

Adoption is deferred, and not for the reason expected

The reach is bounded at one hop (maxDepth 1, maxDist 4 tiles = one grid step), so a one-chunk halo is all it would ever need. The reach is not the obstacle.

The obstacle is a perf guard: test/cliffCellBounds.spec.ts pins the tiled-to-whole noise ratio below 1.1 precisely to stop that kind of inflation. A 128px worker tile is 4x4 chunks and would become 6x6 - ~2.25x the cliff-pass cell work by geometry (estimated, not benchmarked). Buying +2 positions for that is a trade to make with a benchmark in hand, not as a side effect.

Verification

pnpm run verify green - 1566 app (up 4) + 13 worker + 3 container tests, 0 warnings, exit 0. Measurement only; shipping accuracy unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HKjzpWJw1HbnigPP9udqN2