feat(ale): a Breakout ladder that spans the range a player reaches

[drewstone]

The defect

The packaged ALE Breakout contract was derived from a reference that reached score 5 over 210 inputs, so its top achievement was score >= 4. Nothing in it could tell 7 points from 24. 0.8.0's achievement-only scoring stopped it ranking the worse player FIRST; it left every program that plays at all tied at the same three achievements, and the search with no gradient above them.

The measurement

Three deterministic controls, ale-py 0.12.1, seed 0, replay-attested. screen-blind repeats a fixed three-word cycle and is the strongest program a sweep of all 340 input patterns of period four or less found.

control	game score @300	@600	lives left	achievements before	after @300	after @600
screen-blind	7	7	0	3 of 5	3 of 7	3 of 7
steer-from-ascii	6	7	0 (dead at 375)	3 of 5	3 of 7	3 of 7
steer-from-ram	9	24	5	3 of 5	4 of 7	5 of 7

The RAM control now ranks strictly first at both budgets, and two rungs stay open above it.

The contract

A new reference, recorded once at seed 0 from a predictive paddle controller that reads the RAM ball and paddle channels, estimates the ball velocity from the previous decision, reflects the predicted path off the side walls, and steers to the crossing point. It reaches score 64 over 839 inputs.

The reference declares trigger points that double — 1, 2, 4, 8, 16, 32, 64 — and deriveContract samples the score that actually held at each, so the packaged checks are score >= 1, 2, 4, 8, 18, 32, 64. The fifth trigger reads 18 because the score steps 14 to 18 when the run clears a four-point row.

PackagedContract.calibrate passes with no declaration at all. All seven trivial baselines still score zero, at 210, 300, 450, 600, 900 and 1,200 inputs.

What the contract no longer states

• The two hash milestones. The substitution sweep measures 56 of 96 single-input substitutions still satisfying them, at 21 of 32 turns, for at least 5.22e11 distinct 32-input logs. No independent control has ever landed on that state, so they were two points of denominator only the reference could score against.
• life-lost. It is lives == 4, earned by dying. Removing it makes the whole-contract score equal the achievement score, so no consumer can pick the number that ranks a control that died above one that did not.
• The requires chain. engineState.score rises and never falls, measured over the reference and every baseline, so score >= 18 cannot pass before score >= 8. collapse.collapses is now false, and the seven rungs open at seven distinct inputs (32, 71, 137, 259, 404, 636, 839).

ale.test.mts keeps a demonstration contract derived from the same reference that carries all three, so the screen-frame and save-file tiers, the collision sweep, the attrition classifier and the collapse gate stay under test on the real emulator.

What breaks

The packaged reference file, the derived contract and its contract hash all change. A stored attestation against the old contract does not verify against the new one, and its milestone ids no longer exist. Pass the previous reference through makeAle({ game: 'breakout', reference }) to keep the old contract. No adapter behaviour changed: the worker, the evidence keys, the observation, the vocabulary and the seed handling are untouched.

Every pinned number that moved is updated in the same PR: ale.test.mts, README.md, docs/adapters.md, and the measured claims in calibration.ts, schema.ts and adapters/ale.ts. The version field is untouched.