Let the AI use Goblin Welder

[liamiak]

Goblin Welder was AI:RemoveDeck:All, which is not only a deckbuilding flag:
AiController.chooseSpellAbilityToPlay strips every ability of a flagged card from the candidate
list, so on master the AI never activates it and no PumpAi code runs.

Lifting the flag alone is worse than leaving it. It targets both the artifact sacrificed and the
graveyard card replacing it; with no AILogic the generic pump targeting takes the most expensive
artifact on the board regardless of owner. With only the flag removed:

Board (Welder plus)	Flag lifted, no AILogic
ours: Chromatic Star, Sol Ring; Wurmcoil in the yard	sacrificed the Sol Ring
ours: Sol Ring; only an Ornithopter in the yard	still welded — Sol Ring → Ornithopter
theirs: Wurmcoil, Blightsteel in their yard	sacrificed their Wurmcoil, handing them the Colossus and two tokens

AILogic$ Weld picks the pair as one trade, delegating each choice to what already exists:

• ours, going away — getCardPreference(… "SacCost" …), else getWorstAI. That keeps Sol Ring
  over Chromatic Star, which carries SVar:SacMe:1.
• theirs, going away — getBestRemovalTargetAI; welding against an opponent is removal.
• coming back — pumpMandatoryTarget: best from our graveyard, cheapest from theirs. The
  sub-ability needs the tag too — pumpTgtAI resets targets before dispatching, and the generic path
  would take their best.
• which trade — the largest mana-value delta, positive being the whole go/no-go test.
  evaluatePermanentList values a non-creature permanent at CMC + 1; ZoneExchangeAi gates its
  swap the same way. A point off an opponent counts as a point to us, as GameStateEvaluator does
  for board presence.
• timing — the shared AtOppEOT rule.

Only this card carries the AILogic, so nothing else through PumpAi changes.
AI:RemoveDeck:Random goes too, since AI_CAN_PLAY needs both flags clear;
DeckNeeds:Type$Artifact replaces it.

Reasoned, not measured: the opponent's end step over acting earlier. Known limit: one candidate pair
per player, not every combination.

Six tests, second commit, droppable alone: two fail without the fix, two caught earlier versions of
this logic, two guard the "no trade" branch. Full desktop suite: 363 tests, 0 failures, 6 skipped.

Written with Claude Opus 5 (also recorded in the commit co-authors).

[liamiak]

Rebased onto master and reworked. Two things changed the shape of this:

• AI:RemoveDeck:All turns out to gate the ability at runtime, not only at deckbuilding — AiController.chooseSpellAbilityToPlay strips a flagged card's abilities from the candidate list, so no PumpAi code was ever reached. That makes removing the flag on its own a regression rather than a no-op, and the description now shows what the generic targeting does with the flag lifted and no AILogic.
• The pair taken is now the one with the largest mana value delta rather than the first profitable one found, so a bigger swing against an opponent is no longer passed over for a smaller upgrade of our own.

AI:RemoveDeck:Random goes as well, since AI_CAN_PLAY needs both flags clear; DeckNeeds:Type$Artifact replaces it. Six tests, two of which fail without the fix.

[liamiak]

Rebased onto master — PumpAi.checkPhaseRestrictions was reworked since (#11739, #11742) and the weld path sits above that change. Full desktop suite: 371 tests, 0 failures, 6 skipped.

I also ran this in real games rather than only unit tests: 20 headless mirror matches with a mono-red artifact deck. The AI chose to weld 49 times — 22 upgrading its own board (Ichor Wellspring → Sundering Titan, Pyrite Spellbomb → Wurmcoil Engine), 27 against opponents (their Blightsteel Colossus → their Mind Stone). No exceptions.

That run found a bug in my own logic, now fixed in the first commit: the sub-ability's targets weren't cleared on entry and the chosen pair was re-derived after the scan, so 2 of 30 decisions committed to a different card than the one they were scored on. It now keeps the pair it scored — 0 divergences in 49.

Cost: the AtOppEOT gate keeps it off the per-decision path — 100 calls across 20 games, ~0.1s of AI time per game (per-call cost varied 8–28ms between runs).

Limit the run also exposed: the go/no-go is mana value, so a cheap-but-strong artifact is undervalued (it will trade a Sol Ring for a Mind Stone) and a token's 0 counts as 0. That's Forge's shared permanent valuation rather than anything specific to this card — happy to address it here if you'd prefer, but I didn't want to change the global model in this PR.