Let the AI use Thespian's Stage

[liamiak]

Logic is 30 lines; the rest is a test in a second commit you can drop.

Thespian's Stage ({2}, {T}: becomes a copy of target land, keeping this ability) is
AI:RemoveDeck:All, and CloneAi says why — its default target branch is

// Those can just use
// AI:RemoveDeck:All until this can do a reasonably good job of picking
// a good target
return false;

Picking a land

GameStateEvaluator.evaluateLand is what AiController.chooseBestLandToPlay already uses to
rank lands, so CloneBestLand ranks with it too. Scores on a board with four Swamps:

Land	ours	opponent's
Swamp	106	106
Thespian's Stage	156	156
Cabal Coffers	306	6
Ancient Tomb	209	209
Mishra's Factory	181	181
Wasteland	116	116
Gaea's Cradle	6	6

Two things follow from that table. Only lands we control are considered — a land is scored
under whoever controls it now, and Cabal Coffers swings between 6 and 306 depending on whose
Swamps get counted, so reading an opponent's board misjudges in both directions while the copy
arrives under ours either way. And our own legendary lands are skipped, since copying one
just means sacrificing one of the two.

The bar is the Stage's own score, which is deliberately conservative: the copy keeps the clone
ability, so a target really only has to beat a plain land (106) rather than the Stage's 156. It
passes on marginal upgrades like Wasteland (116) instead of risking a bad one.

getBestLandAI would have been the obvious helper, but it returns a random non-basic, which
is fine for "which land to fetch or destroy" and not for this.

When

checkPhaseRestrictions currently allows an activated clone only during the opponent's
declare-attackers step with attackers, which suits a temporary combat clone but not this. The
upgrade is permanent, so there is nothing to gain by taking it early, and the cost taps the
Stage — activating at the end of the turn before ours keeps the mana up as a possible threat
for as long as possible and still unlocks the Stage in our own untap step. That is exactly the
base class's AtOppEOT rule, so this delegates to it rather than restating the condition.

Existing cards are untouched: the override only diverges on CloneBestLand, and otherwise
calls super, which dispatches to the same method as before.

As an aside, that gate looks like it was meant for temporary clones — the block above it and
the one below it are both conditioned on Duration, and none of the 24 cards with an
activated or cast Clone has one, so the middle block is the only one that ever applies to
them. Happy to look at that separately if it's worth changing.

Measured

End of the opponent's turn, four Swamps and enough mana:

Board	Before	After
Our Ancient Tomb	CantPlayAi	copies Ancient Tomb
Our Cabal Coffers	CantPlayAi	copies Cabal Coffers
Our Mishra's Factory	CantPlayAi	copies Mishra's Factory
Our Wasteland	CantPlayAi	declines (116 < 156)
Our Gaea's Cradle	CantPlayAi	declines (legendary)
Opponent's Ancient Tomb	CantPlayAi	declines (not ours)
Basics only	CantPlayAi	declines
Any of the above, at any other point in the turn cycle	MissingPhaseRestrictions	unchanged

Testing

• mvn -pl forge-gui-desktop -am test: 289 tests, 0 failures.
• The test was verified to fail without the change — the AI never activates the ability, so the
  Stage keeps its own name — and one of the three cases fails if the timing is loosened back to
  our own main phase.
• No other card reaches either branch: CloneBestLand is new and only Thespian's Stage uses it.

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

[liamiak]

Force-pushed a timing change: the Stage is now taken at the end of the turn before ours, rather than in our own main phase.

Two reasons. The upgrade is permanent, so there is nothing to gain by taking it early, and holding the mana until then leaves a possible threat open. The second one I had missed: the cost taps the Stage, so activating in our own main phase costs us the copied land's mana for that whole turn, while at the opponent's end step it unlocks again in our untap step.

That is exactly the base class's AtOppEOT rule, so the override delegates to it rather than restating the condition. Added a third test asserting the AI declines during its own main phase.