LookaheadPolicy

The skeleton shared by every policy that chooses among available actions rather than constructing one directly: value-function approximations, and cost-function approximations that score candidates.

A modeler supplies the three model-specific pieces and nothing else:

class MyRule : LookaheadPolicy(ExhaustiveSearch) {
    override fun contribution(obs, action, ctx) = ...   // C(s, a), the immediate cost
    override fun postDecision(obs, action)      = ...   // S^x(s, a)
    override fun value(postDecision, ctx)       = ...   // V̄
}

The enumeration, the feasibility filter, the argmin and the empty-set fallback come from the library. Contrast §8.2.9's first VFA, which inlined all four inside one action method — the reason §8.2.11 argues the design was one abstraction where the problem has five.

Not every policy fits this shape, and that is deliberate. A rule that constructs an action — the greedy allocator of §8.2.8 sorts regions and fills them — implements plain PolicyIfc instead. Two shapes, because there are two ways to decide.