test(cliffs): the ore effect DECOMPOSES into rejections + cascade (#84)

[wormeyman]

With every route and every idea closed (#137, #138, #140), the premise itself became the suspect — the lesson of #136. It was re-audited and it survives, sharper than it went in. No capture: this is a fold of fixtures already on disk.

What the fold shows

• Pure SUPPRESSION. The resources-ON cliff set is a strict subset of the resources-OFF set in all three arms — zero cells are lost when the ore is removed. A perturbed field moves cells both ways; only a rejection moves them one way. That is the shape test, and it passes.
• Exactly ADDITIVE. 27 (calcite) + 4 (geyser) = 31 (all) — the same cells, disjoint sets, not merely equal counts.
• The orientation changes among SURVIVING cells are the destruction cascade. Take the ore-OFF world, destroy exactly the cells the ore rejects, run the port's destroyEnd cascade → you get the ore-ON world exactly, positions and orientations, in all three arms.

So there is no unexplained component of the ore effect at all. Every difference between the two worlds — the 5 / 4 / 1 orientation changes included — is rejection-plus-cascade.

Why this matters more than the numbers

It reframes what is open, and the old framing is what kept it stuck. "What mechanism lets the ore reach a cliff?" already has an answer: a rejection at the apply stage, exactly as #108 and #113 said, with the cascade as its only secondary effect. That is why five routes and three ideas all came back negative — they were hunting a pathway that is already known.

What remains open is narrower and far more tractable: which cells get rejected, and by what criterion. makeVulcanusOreRejection already reproduces that criterion at precision 1.000 without reaching recall 1, so the gap is a geometry question about specific cells, not a missing route through the generator.

A second, independent confirmation of the cascade

#139 confirmed the cascade against the game through a Lua runtime probe. This confirms the same model against ordinary map-generation output, via a completely different instrument. The two agree.

Coverage, measured rather than assumed

Planting one-sided destroyEnd no-ops:

planted no-op	this spec
every side	fails 3
south	fails 3
west	fails 2
north	passes
east	passes

These rejections only ever trim south and west ends, so this fold covers two of the four directions — it is not a whole-model guard alone. #139's probe covers north (a side === 0 plant fails both its ON arms). The spec says so explicitly, so a green run here is not misread as end-to-end cascade verification.

Verification

pnpm run verify green — 1543 app (up 9) + 13 worker + 3 container tests, 0 warnings, exit 0. Measurement only; shipping accuracy unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_012VRsPwmSwjnFKVSS6jx9Mo