Let the AI pay for converge and sunburst when X is the colour knob

[liamiak]

On a converge or sunburst card whose colour count lives in a second SVar, X has one job: buy colours. The AI was not announcing it, so those cards were cast for the minimum.

ComputerUtilCost.setMaxXValue ends with sa.setXManaCostPaid(x), leaving X at the maximum. PermanentAi then measured its converge baseline there:

final int xPay = ComputerUtilCost.setMaxXValue(sa, ai, false);   // X is now the max
if (source.hasConverge()) {
    int nColors = ComputerUtilMana.getConvergeCount(sa, ai);     // measured at max X
    for (int i = 1; i <= xPay; i++) {
        sa.setXManaCostPaid(i);
        int newColors = ComputerUtilMana.getConvergeCount(sa, ai);
        if (newColors > nColors) { nColors = newColors; }
        else { sa.setXManaCostPaid(i - 1); break; }
    }
}

The baseline is already the best colour count available, so the first step of the walk always compared worse and X collapsed to 0. TokenAi had a different miss: its X block is guarded by "X".equals(tokenAmount), so a card counting colours in Y never reached any X handling at all.

Both want the same thing, so it is now named once - ComputerUtilMana.setXForBestConverge - and PermanentAi shrinks by 10 lines calling it.

Effect

All five cards in the pool whose cost has X and for which Card.hasConverge() is true, with five differently coloured lands untapped:

	X before	X after
Engineered Explosives	0	5	0 charge counters, so it could only ever destroy 0-drops
Chamber Sentry	0	5	entered as a 0/0 and died
Skyrider Elf	0	3	entered as a 0/0 and died
Sweep the Skies	unannounced	4	converge 1 -> 5, so five thopters rather than one
Prismatic Ending	unannounced	unannounced	not fixed, see below

Only Engineered Explosives is flagged. The other three that this fixes are in AI decks today - two of them entering as 0/0s and dying.

Edge cases behave: five Wastes gives X=0, and Plains/Plains/Island gives X=2 rather than paying for a colour already covered.

Engineered Explosives unflagged

Casting it for X=0 was the only thing wrong with it. The activated half needs nothing: DestroyAll routes it through the same DestroyAllAi.doMassRemovalLogic as Wrath of God, and it matched Wrath's decision on every board I tried. Two charge counters, everything at cmc 2 so both cards hit the same permanents:

AI creatures	opponent creatures	Engineered Explosives	Wrath of God
161	307	declines	declines
161	468	plays	plays
322	629	plays	plays

It does wipe at a cost to itself when the swing justifies it, on the shared CREATURE_EVAL_THRESHOLD - not only when it risks nothing.

The flag was also self-fulfilling: AiController.getSpellAbilityToPlayImpl strips every ability of a flagged card before evaluating anything, so "the AI never plays it" held regardless of whether its AI worked.

Prismatic Ending, not fixed

Announcing X takes its converge from 1 to 5 and it still declines, so X is necessary but not sufficient there. Its condition is ConditionPresent$ Permanent.nonLand+cmcLEY, and Count$Converge reads the colours actually paid:

if (sq[0].contains("Converge")) {
    SpellAbility castSA = c.getCastSA();
    return doXMath(castSA == null ? 0 : castSA.getPayingColors().countColors(), expr, c, ctb);
}

Before the spell is cast getCastSA() is null, so Y is 0 and cmcLEY matches nothing above a 0-drop. The AI's getConvergeCount predicts converge by simulating payment, but the condition path uses the rules-side count; bridging the two is a separate piece of work and I have left the card flagged.

DamageAllAi:30 and DestroyAllAi:64 have the same X-only assumption, but no card in the pool currently reaches them through it.

Testing

Two tests. Engineered Explosives plays a real turn and lands with five charge counters - that one has to go through the full AiController path, since it is the unflagging that needs proving. Sweep the Skies asserts the announced X instead, because TokenAi gates token spells behind a random roll and a played-turn assertion on it is flaky; both assertions were run repeatedly each way.

Each test fails without its own half of the change. 337 tests, 0 failures, checkstyle clean.

🤖 Implemented with the assistance of Claude Code (Opus 5).

[tool4ever]

I don't think the reasoning for this was zero X since that's another regression and more like AI unable to coordinate which one is actually a smart choice

[liamiak]

You're right — closing into #11486.

Engineered Explosives needs board-aware counter selection rather than most-colours: its counters
pick which mana value the wipe hits, so five is usually the worst live choice. It stays flagged.
#11486 carries the other three cards and unifies the duplicated simulation you flagged there.