PreRankModel

A model whose only job is to order the mover's legal turns, so the search expands the best ExpectimaxConfig.candidateLimit of them rather than the best the material proxy could find.

It gets its own session on purpose. Ordering candidates and valuing positions are different jobs: a ranker is trained on which turn is better, a value model on how good a position is, and the two have different targets, can use different feature schemas, and are not interchangeable just because both emit one number per row. Reusing the leaf model for ranking is the cheaper option this repository already offers — preRankWithModel — and the two are alternatives, not a pair.

What bounds its cost is the feature schema, not this type. Pre-ranking sees every legal turn, which in Dice Chess is routinely hundreds and can be thousands; at that width material-7-v1 is a thousand cheap rows while kcp-13 is a thousand 216-outcome capture-probability searches, which is not viable at this seam at all. The pre-rank pass is also paid in full before the deadline is consulted — it has to be, since its output is what the anytime fallback plays — so a schema too expensive for the position's candidate count overruns the move budget before the search proper begins.