Let the AI use Goblin Welder

[liamiak]

Goblin Welder was AI:RemoveDeck:All. Its ability targets both the artifact that gets sacrificed and
the graveyard card that replaces it, so the two have to be read as one trade. PumpAi read neither —
with no AILogic it fell through to the generic pump targeting, which takes the most expensive
artifact on the battlefield regardless of who owns it.

Three misplays, measured by probing the decision directly (canPlaySa plus the resulting targets):

Board	Before
Welder, Chromatic Star, Sol Ring; Wurmcoil Engine in our own graveyard	sacrificed the Sol Ring
Welder, Sol Ring; only an Ornithopter in our graveyard	still welded — Sol Ring → Ornithopter
Opponent holds the biggest artifact, bomb in their graveyard	targeted theirs — sacrificed their Wurmcoil to hand them a Blightsteel Colossus

AILogic$ Weld picks the pair together, delegating each actual choice to something that already
exists:

• ours, going away — the existing SacCost preference (ComputerUtil.getCardPreference),
  otherwise the cheapest artifact we control. That is what keeps Sol Ring over Chromatic Star, which
  mana value alone cannot separate.
• theirs, going away — getBestRemovalTargetAI, since welding against an opponent is removal.
• coming back — pumpMandatoryTarget, already in this class: best from our own graveyard,
  cheapest from theirs.
• worth doing? — the card returning has to cost more than the one leaving, the test
  ZoneExchangeAi already uses for a swap across zones.
• timing — the shared AtOppEOT rule, rather than the mid-combat window the generic pump phase
  check happened to allow.

Both branches sit behind the new AILogic, which only this card carries, so nothing else that goes
through PumpAi changes.

Reasoned rather than measured: that welding at the opponent's end step beats acting earlier. The
ability is free and repeatable so holding it costs nothing, but I have not measured a win-rate
difference.

Known limit: it weighs one candidate pair per player rather than every combination, so a swap that a
different artifact of theirs would have justified can be missed.

Full desktop suite: 356 tests, 0 failures. The regression test is the second commit and can be
dropped on its own — two of its cases fail without the fix.

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