ChanceCollapse

Learned replacement for a root candidate's chance node: instead of expanding the opponent's 56 weighted dice outcomes and their replies, one batched call scores the position each candidate reaches and returns that expectation directly.

This is the search's most expensive layer by orders of magnitude — every candidate pays hundreds to thousands of leaf evaluations for one number — so a model that predicts that number turns a chance node into a single row of a batch. It is off by default and has to be enabled explicitly, because it trades an exact expectation for an estimate: what comes back is what the model thinks the expansion would have produced.

What it keeps: the immediate-king-capture shortcut and the forced pass above it, pre-ranking and its deterministic order, the random tie-break, the deadline's meaning (the batch is indivisible and is not started once the deadline has passed, so the anytime fallback to the pre-ranker's pick is unchanged), and root rescoring.

What it gives up: Star1/Star2 pruning and the transposition table have nothing to prune or store — a collapsed value is a different quantity from the exact expectation the table holds at that depth, so the table is left untouched rather than mixed — and ExpectimaxConfig.searchDepth has no effect, since no tree is built below the root.

It also gives up per-roll knowledge, and with it the one rule RootRescore treats as inviolable: a candidate that loses our king on some roll must never be rescued by a rescorer. lossGuard restores exactly that rule without expanding anything — KingCaptureProbability answers "can the opponent capture our king on their next roll" over the same 56 multisets, exactly (king-capture paths ignore the maximum micro-moves rule, so its depth-first search is not an estimate), at one 216-outcome search per root candidate. Leave it off for a pure collapse; turn it on when a rescorer is configured at a positive weight, or the rescorer can outvote a lost line.