test(cliffs): the residual is 31 DESTRUCTION decisions and nothing else (#84)

[wormeyman]

Answers the question #113 left open and named as unmeasured: are the 5 wrong orientations that survive the surplus cascade the fallout of the 5 false rejections, or a second defect?

They are the fallout. Measuring it the direct way settles considerably more than the question asked.

The result

Stop scoring predicates. Hand the port the game's own destruction set - destroy the raw cells the game does not have, keep everything else, let applyCliffs and the Cliff::onDestroy cascade run. Over all three oracle regions, with the real ore and the real lava in the world:

1531 of 1531. Zero wrong, zero surplus, zero missing.

So the exactness #113 found at [1500,1500] was not a property of that stripped-down levers world - it holds everywhere the port has ground truth. There is no orientation defect, no crossing defect, no field defect and no missing suppressor left anywhere in the Vulcanus cliff port.

Fitted vs predicted

• FITTED: the 225-cell destruction set - 225 booleans, the only thing taken from the answer.
• PREDICTED: all 1531 orientations, of which 14 are ones the cascade actively rewrites (19 ways each to be wrong); that the cascade destroys nothing beyond the 225, which would have shown as missing; and halo-independence, checked by re-running with our own lava+ore predicate outside the region instead of nothing (identical).
• Control: destroy the same 225 without telling their neighbours and the 14 come back wrong.

The 18 wrong orientations split cleanly into the two destruction errors

collides	matched	wrong	surplus	missing
lava + ore (what #113 scored)	1508	18	22	5
...plus the 22 the game destroys	1521	5	0	5
...instead sparing the cells the game keeps	1517	14	22	0
the game's set - both fixed	1531	0	0	0

Row 2 is #113's prediction confirmed. Row 3 is the half it could not measure.

The target is 31, not 27

Surplus and missing are outcomes and the cascade has already blurred them. One boolean per raw cell, before any cascade:

	count
raw cells inside the three regions	1756
the game destroys	225
we destroy	206
agree	200
we destroy, the game keeps	6
the game destroys, we keep	25

One of the 6 never shows as missing because a neighbour's cascade gets there first, and 3 of the 25 are cascade casualties anyway. Quote 31 for the rule and 27 for the rendered output - same defect, two stages.

Precision 200/206 = 0.971, recall 200/225 = 0.889, errors both ways - #111's finding restated on the stage that actually does the rejecting. #88 still stands: do not tune the box until it fits.

Two facts that narrow the next step

• All 6 false rejections are the LAVA box; the ore rule invents none. The whole two-sided error set belongs to one predicate, the tile half of Surface::wouldCollide.
• The raw set is a strict superset of the game's in all three regions (292/283, 1070/861, 394/387), not just at [1500,1500] where test(cliffs): two LEVERS - no entity suppresses a cliff, lava is 166/169 (#84) #111 measured it with the levers. Nothing left is a failure to generate a cliff.

Next

Surface::wouldCollide runs constCollideWithTile against the real surface; the port resolves tiles from our own Vulcanus tile model, and the two have never been compared inside a cliff's box - oracle-vulcanus-tile-names is a ~2600-tile spiral that does not cover these regions densely. A small capture over the 31 disputed boxes plus a control set would settle it either way.

Nothing shipping changes; this is measurement only. pnpm run verify green (1418 tests).

🤖 Generated with Claude Code

https://claude.ai/code/session_01KVXnqZPGsD7A2SaPVA9ArA