fix(moonbreak): estimate losses from the real waves, not a pooled average

[rolljee]

The inconsistency flagged at the end of #56.

The bug

computeMoonbreak builds its probability from each attacker's own six-wave split — remainder waves of base + 1, the rest of base. Its loss estimate did something else: it pooled the whole attack into one average wave size, totalRip / min(totalRip, attackers × 6).

The two halves of the same calculation therefore modelled different attacks, and the losses came out blind to how the fleet was divided:

fleets	total	probability	mean losses (before)	mean losses (after)
[1, 100]	101	78.96 %	22.04	26.74
[50, 50]	100	87.05 %	21.90	22.06

Same answer for two attacks that are nothing alike. Spread evenly, the waves are smaller, the moon falls sooner and fewer waves ever fire — so it should cost less, not the same.

The fix

Walk the same wave sequence the probability is built from: each attacker's six waves, at their own size, in firing order. An attacker with fewer than six Deathstars pads with empty waves, which neither threaten the moon nor lose a ship, so they cost nothing to carry.

Impact on the numbers people already trust

Even fleets — what the module was checked against when it was ported from og-bot-discord — barely move:

case	before	after
[6] on 8944 km	2.48	2.48
[25, 25]	13.44	13.53
[25, 25, 25, 25]	16.72	16.85
[100]	27.66	27.78

Lopsided ones move a lot, which was the point.

One consequence worth stating

Losses now depend on the order the attackers are listed, because the fleet that opens faces a moon still standing: [100, 1] loses 27.88 where [1, 100] loses 26.74. The probability is unchanged either way. That is real — attackers fly one after the other — and the help text now says so in both languages, rather than leaving it to be discovered.

Not changed

The variance still sums the per-wave variances, treating waves as independent although they share the survival chain, and the spread is still summarised as a gaussian. Both are approximations kept from the model this was ported from; they are now documented as such in the code. Changing them is a separate decision about the statistics, not about which attack is being modelled.

Also here

ogamejs bumped to 4.0.2, which refuses rate terms that cannot describe a trade (rolljee/ogamejs#40) — the follow-up noted in #56.

Tests

New coverage for the wave decomposition (attackerWaves, attackWaves: extras first, empty padding, firing order, every Deathstar accounted for) and for the two behaviours that motivated the change — a lopsided attack costing more than an even one, and the leading fleet costing more than the trailing one. Existing expectations updated where the numbers moved. 415 pass, build OK.

🤖 Generated with Claude Code